1 01:00:00,000 --> 01:00:01,000 As an engineer, you want to build 2 01:00:01,000 --> 01:00:01,958 features, you want to 3 01:00:01,958 --> 01:00:03,791 build like real robust systems. 4 01:00:04,125 --> 01:00:05,083 You want to spend ideally 5 01:00:05,083 --> 01:00:06,708 all your time doing that. 6 01:00:06,958 --> 01:00:08,291 You don't want to spend your time 7 01:00:08,291 --> 01:00:09,291 fighting the language, 8 01:00:09,708 --> 01:00:10,958 building something that, you 9 01:00:10,958 --> 01:00:12,666 know, the language was not meant for. 10 01:00:12,666 --> 01:00:13,458 To give you an example, 11 01:00:13,458 --> 01:00:14,500 like concurrency in TypeScript. 12 01:00:14,833 --> 01:00:17,291 You want a way to build a concurrent 13 01:00:17,291 --> 01:00:18,375 system where you're 14 01:00:18,375 --> 01:00:19,375 thinking about what is 15 01:00:19,375 --> 01:00:20,750 the system actually doing rather than 16 01:00:20,750 --> 01:00:22,125 like setting up all the 17 01:00:22,125 --> 01:00:23,375 boilerplate in order to 18 01:00:23,541 --> 01:00:24,916 have that system in place. 19 01:00:25,458 --> 01:00:27,333 So I would say Effect really 20 01:00:27,333 --> 01:00:28,833 gives you the right premise. 21 01:00:28,875 --> 01:00:30,416 So now that we've completed this 22 01:00:30,416 --> 01:00:31,666 migration, kind of feel 23 01:00:31,666 --> 01:00:33,208 like this is the end game feels 24 01:00:33,208 --> 01:00:35,541 like more stable and at least if we 25 01:00:35,541 --> 01:00:36,666 continue building our 26 01:00:36,666 --> 01:00:38,000 code base on TypeScript with 27 01:00:38,041 --> 01:00:39,750 Effect, it doesn't feel like we're going 28 01:00:39,750 --> 01:00:41,416 to spend many more like 29 01:00:41,416 --> 01:00:43,041 cycles, like refactoring 30 01:00:43,041 --> 01:00:43,750 the code. 31 01:00:47,291 --> 01:00:49,375 Welcome to Cause & Effect, a podcast 32 01:00:49,375 --> 01:00:50,291 about the TypeScript 33 01:00:50,291 --> 01:00:51,958 library and ecosystem called 34 01:00:51,958 --> 01:00:54,083 Effect, helping engineers to build 35 01:00:54,083 --> 01:00:55,208 production-ready software. 36 01:00:56,041 --> 01:00:58,000 I'm your host, Johannes Schickling, and I've 37 01:00:58,000 --> 01:00:58,750 been building with 38 01:00:58,750 --> 01:01:00,000 Effect for over four years. 39 01:01:00,583 --> 01:01:02,541 With this podcast, I want to help others 40 01:01:02,541 --> 01:01:04,333 understand the powers and 41 01:01:04,333 --> 01:01:05,708 benefits of using Effect. 42 01:01:06,625 --> 01:01:08,583 In this episode, I'm talking to Michael 43 01:01:08,583 --> 01:01:10,375 Fester, co-founder of 44 01:01:10,375 --> 01:01:12,333 Markprompt, an AI-based customer 45 01:01:12,333 --> 01:01:13,208 support platform. 46 01:01:13,916 --> 01:01:16,083 In this conversation, we explore how 47 01:01:16,083 --> 01:01:17,333 Markprompt uses Effect 48 01:01:17,333 --> 01:01:19,291 across its entire architecture, 49 01:01:20,041 --> 01:01:21,958 including API and RPC endpoints, 50 01:01:22,500 --> 01:01:24,125 long-running workflows, and their 51 01:01:24,125 --> 01:01:25,166 customer-facing product. 52 01:01:25,666 --> 01:01:26,541 Let's get into it. 53 01:01:28,000 --> 01:01:28,958 Hey, welcome Michael. 54 01:01:29,250 --> 01:01:30,416 So nice to have you on the 55 01:01:30,416 --> 01:01:31,583 Cause & Effect Podcast. 56 01:01:31,583 --> 01:01:32,041 How are you doing? 57 01:01:32,708 --> 01:01:33,250 I'm good. 58 01:01:33,250 --> 01:01:33,666 Thank you. 59 01:01:33,666 --> 01:01:35,416 I'm really happy to be here. 60 01:01:35,416 --> 01:01:36,916 It's been, I guess it's been a while that 61 01:01:36,916 --> 01:01:37,666 we're trying to 62 01:01:37,666 --> 01:01:39,541 organize this, but I'm very 63 01:01:39,541 --> 01:01:40,708 happy that we're doing it now. 64 01:01:41,458 --> 01:01:41,666 Yeah. 65 01:01:41,833 --> 01:01:44,250 I mean, you're super, super busy. 66 01:01:44,458 --> 01:01:45,458 You're based in San 67 01:01:45,458 --> 01:01:47,791 Francisco building an AI company. 68 01:01:48,000 --> 01:01:50,083 So I'm really, really happy that you took 69 01:01:50,083 --> 01:01:51,625 the time to share 70 01:01:51,625 --> 01:01:53,041 your story about Effect. 71 01:01:53,875 --> 01:01:56,000 Maybe you can share a little bit about 72 01:01:56,000 --> 01:01:56,958 yourself, your 73 01:01:56,958 --> 01:01:58,625 background and what are you doing? 74 01:01:59,958 --> 01:02:00,625 Yeah, sure. 75 01:02:00,958 --> 01:02:02,208 So I have a background. 76 01:02:02,208 --> 01:02:03,208 I started in academia. 77 01:02:03,500 --> 01:02:04,625 Actually, I was doing 78 01:02:04,625 --> 01:02:05,500 research in mathematics. 79 01:02:05,708 --> 01:02:07,208 Then I started getting into AI and I 80 01:02:07,208 --> 01:02:08,083 built my first AI 81 01:02:08,083 --> 01:02:10,125 company in Europe, specialized 82 01:02:10,250 --> 01:02:11,750 on device voice recognition. 83 01:02:12,291 --> 01:02:13,916 And we got acquired by Sonos. 84 01:02:14,291 --> 01:02:17,333 And since then I moved here to San 85 01:02:17,333 --> 01:02:19,458 Francisco for my next 86 01:02:19,458 --> 01:02:21,750 venture, which is we're building 87 01:02:22,666 --> 01:02:24,666 AI infrastructure for customer support. 88 01:02:25,375 --> 01:02:26,916 So this is all like posts, 89 01:02:27,375 --> 01:02:29,375 chat GPT, GPT 4 coming out. 90 01:02:30,083 --> 01:02:31,583 And we saw this, this incredible 91 01:02:31,583 --> 01:02:32,958 opportunity to start 92 01:02:32,958 --> 01:02:35,541 working in this space. 93 01:02:35,625 --> 01:02:36,875 That is awesome. 94 01:02:36,875 --> 01:02:40,166 And I highly encourage everyone who is 95 01:02:40,166 --> 01:02:41,791 like interested in AI, 96 01:02:41,791 --> 01:02:43,333 particularly is looking 97 01:02:43,333 --> 01:02:46,500 for a better customer support solution to 98 01:02:46,500 --> 01:02:47,583 check out Markprompt. 99 01:02:48,041 --> 01:02:50,166 Can you share about a few companies who 100 01:02:50,166 --> 01:02:51,458 are already using Markprompt? 101 01:02:51,833 --> 01:02:52,291 Yeah, sure. 102 01:02:52,541 --> 01:02:55,250 So we've been working very closely with 103 01:02:55,250 --> 01:02:56,458 some companies very 104 01:02:56,458 --> 01:02:57,416 early on in the process. 105 01:02:57,708 --> 01:02:58,500 Actually, that's really interesting. 106 01:02:58,833 --> 01:03:00,958 Vercel started using us as one of our 107 01:03:00,958 --> 01:03:01,750 first customers, 108 01:03:01,750 --> 01:03:03,375 actually ElevenLabs, Replicate. 109 01:03:03,833 --> 01:03:05,666 Things are just moving so fast. 110 01:03:05,666 --> 01:03:08,541 So our approach here is not so much of a, 111 01:03:08,541 --> 01:03:09,375 you know, we have a 112 01:03:09,375 --> 01:03:10,583 vision of how things will 113 01:03:10,583 --> 01:03:11,583 end up, right? 114 01:03:11,583 --> 01:03:13,750 It's very much like a learning process, 115 01:03:13,750 --> 01:03:14,541 working very, very 116 01:03:14,541 --> 01:03:16,541 closely with customers, hand in 117 01:03:16,541 --> 01:03:17,916 hand getting their feedback, 118 01:03:17,916 --> 01:03:19,791 understanding together, how 119 01:03:19,791 --> 01:03:21,625 this thing is evolving, how 120 01:03:21,625 --> 01:03:23,625 we can improve, you know, the behavior of 121 01:03:23,625 --> 01:03:24,708 the LLMs or how we can 122 01:03:24,708 --> 01:03:25,958 extend the use cases. 123 01:03:26,625 --> 01:03:28,666 So this has been a very different way, I 124 01:03:28,666 --> 01:03:29,333 would say, of building 125 01:03:29,333 --> 01:03:30,708 a company from the past, 126 01:03:30,708 --> 01:03:32,333 because it's so new to 127 01:03:32,333 --> 01:03:33,250 everyone in the space. 128 01:03:33,916 --> 01:03:35,083 And we're just like scratching the 129 01:03:35,083 --> 01:03:35,875 surface of the 130 01:03:35,875 --> 01:03:38,333 opportunities of building these systems 131 01:03:38,333 --> 01:03:39,958 for you know, down the line. 132 01:03:40,833 --> 01:03:41,291 That's awesome. 133 01:03:41,541 --> 01:03:44,000 So before we go into the technical parts, 134 01:03:44,500 --> 01:03:45,416 maybe you can share 135 01:03:45,416 --> 01:03:47,208 through the lens of one 136 01:03:47,208 --> 01:03:49,125 of your typical customers, what 137 01:03:49,125 --> 01:03:51,000 Markpompt does and which 138 01:03:51,000 --> 01:03:53,083 problems it solves for some 139 01:03:53,083 --> 01:03:53,833 of those customers. 140 01:03:54,875 --> 01:03:55,000 Sure. 141 01:03:55,041 --> 01:03:57,333 So there's multiple facets of this 142 01:03:57,333 --> 01:03:58,916 running a support 143 01:03:58,916 --> 01:04:01,333 operation is it's not a one size 144 01:04:01,333 --> 01:04:02,125 fits all. 145 01:04:02,125 --> 01:04:04,250 Every company has their own ways of, you 146 01:04:04,250 --> 01:04:05,958 know, the nature of 147 01:04:05,958 --> 01:04:07,625 customer issues, the way that 148 01:04:07,625 --> 01:04:08,791 they engage with customers, 149 01:04:09,041 --> 01:04:10,958 is it like direct contact? 150 01:04:11,208 --> 01:04:14,583 Is it, you know, is it mostly self serve? 151 01:04:14,791 --> 01:04:16,375 Technical challenges are very different 152 01:04:16,375 --> 01:04:18,500 from one company to the other. 153 01:04:18,916 --> 01:04:20,625 So we offer really an 154 01:04:20,625 --> 01:04:21,500 end to end solution. 155 01:04:21,875 --> 01:04:23,208 So there's multiple 156 01:04:23,208 --> 01:04:24,791 components in our system. 157 01:04:24,791 --> 01:04:26,500 It all starts with understanding your 158 01:04:26,500 --> 01:04:28,458 system of record, which is 159 01:04:28,458 --> 01:04:30,250 things like your underlying 160 01:04:30,416 --> 01:04:33,291 customer knowledge, your, you know, your 161 01:04:33,291 --> 01:04:37,708 documentation, your CRM, the API is that 162 01:04:37,708 --> 01:04:39,250 you use for automating 163 01:04:39,250 --> 01:04:41,333 parts of your support. 164 01:04:41,750 --> 01:04:43,166 And we connect all that together. 165 01:04:43,166 --> 01:04:45,833 And then we build applications that are 166 01:04:45,833 --> 01:04:46,958 either customer facing 167 01:04:46,958 --> 01:04:49,458 or agent facing so and use 168 01:04:49,458 --> 01:04:50,958 internally for the company. 169 01:04:51,041 --> 01:04:52,666 So it can be, you know, everything from, 170 01:04:52,666 --> 01:04:55,458 you know, your contact form powered by AI 171 01:04:55,666 --> 01:04:58,041 to your chatbot to your 172 01:04:58,041 --> 01:05:00,625 Slack and Discord bots, right? 173 01:05:00,833 --> 01:05:01,333 This is customer 174 01:05:01,333 --> 01:05:04,041 facing email order replies. 175 01:05:04,375 --> 01:05:08,083 So you send an email and the AI agent 176 01:05:08,083 --> 01:05:09,458 will start, you know, 177 01:05:09,833 --> 01:05:11,500 trying to resolve the issue 178 01:05:11,500 --> 01:05:12,000 for you. 179 01:05:12,333 --> 01:05:13,250 And it might be possible 180 01:05:13,250 --> 01:05:15,291 that it resolves it fully. 181 01:05:16,250 --> 01:05:17,541 And if not, it's going to ask you for all 182 01:05:17,541 --> 01:05:19,791 the questions that the team then needs to 183 01:05:19,791 --> 01:05:21,791 use once you know, the ticket 184 01:05:21,791 --> 01:05:23,125 hits the inbox internally. 185 01:05:23,333 --> 01:05:25,125 And then there's the all the features 186 01:05:25,125 --> 01:05:26,708 which are internal 187 01:05:26,708 --> 01:05:28,583 facing for the support 188 01:05:28,583 --> 01:05:30,916 team to be able to streamline their work. 189 01:05:31,250 --> 01:05:32,583 Because the truth is that customer 190 01:05:32,583 --> 01:05:36,625 support before AI involved 191 01:05:36,625 --> 01:05:38,916 a lot of essentially like 192 01:05:38,958 --> 01:05:41,000 robotic work, just answering the same 193 01:05:41,000 --> 01:05:42,291 questions over and over again. 194 01:05:42,625 --> 01:05:44,250 This is not super engaging 195 01:05:44,250 --> 01:05:45,500 and exciting just to have 196 01:05:45,500 --> 01:05:46,416 to answer the same 197 01:05:46,416 --> 01:05:47,375 question over and over again. 198 01:05:47,625 --> 01:05:49,458 So if you can start like taking that 199 01:05:49,458 --> 01:05:50,333 burden off of your 200 01:05:50,333 --> 01:05:52,625 shoulders and help you, you know, 201 01:05:52,875 --> 01:05:56,291 work on more like high leverage strategic 202 01:05:56,291 --> 01:05:58,625 questions, then, you 203 01:05:58,625 --> 01:05:59,375 know, we can really help 204 01:05:59,375 --> 01:06:00,500 streamline the support operation. 205 01:06:00,708 --> 01:06:01,916 what happens when you've put 206 01:06:01,916 --> 01:06:03,625 an AI or both like customer 207 01:06:03,625 --> 01:06:06,375 facing and agent facing, then you can 208 01:06:06,375 --> 01:06:07,041 start, you know, 209 01:06:07,750 --> 01:06:09,208 extracting a lot of intelligence, 210 01:06:09,208 --> 01:06:10,666 a lot of knowledge, because you're really 211 01:06:10,666 --> 01:06:11,583 at the nerve center 212 01:06:11,583 --> 01:06:13,375 in a company, you are, 213 01:06:13,625 --> 01:06:15,333 you know, the interface between your 214 01:06:15,333 --> 01:06:17,708 customers and your company, right? 215 01:06:18,000 --> 01:06:20,583 So you have so much like knowledge that 216 01:06:20,583 --> 01:06:21,916 often just gets siloed 217 01:06:21,916 --> 01:06:23,166 within the support team, 218 01:06:23,166 --> 01:06:23,375 right? 219 01:06:23,375 --> 01:06:25,791 It's it's actually very rare that support 220 01:06:25,791 --> 01:06:28,458 team reports to product, right? 221 01:06:28,958 --> 01:06:29,500 And I think that's a 222 01:06:29,500 --> 01:06:30,375 that's really a problem. 223 01:06:30,583 --> 01:06:32,375 But if we can start rooting 224 01:06:32,375 --> 01:06:33,875 out all this insight and 225 01:06:33,875 --> 01:06:35,166 knowledge, we can start 226 01:06:35,166 --> 01:06:36,333 having a really, really good 227 01:06:36,333 --> 01:06:37,916 understanding on, you know, 228 01:06:37,916 --> 01:06:39,416 product shortcomings, where 229 01:06:39,416 --> 01:06:40,625 the engineering team might, 230 01:06:40,625 --> 01:06:42,708 you know, put some effort, right? 231 01:06:43,083 --> 01:06:44,625 Because this is where your customers have 232 01:06:44,625 --> 01:06:46,250 been experiencing a 233 01:06:46,250 --> 01:06:47,333 lot of friction, right? 234 01:06:47,666 --> 01:06:49,208 So we can start generating these 235 01:06:49,208 --> 01:06:51,375 reporting and all sorts of 236 01:06:51,375 --> 01:06:52,833 other insights that can help 237 01:06:53,208 --> 01:06:54,750 you as a company 238 01:06:54,750 --> 01:06:55,708 streamline your operations, 239 01:06:56,291 --> 01:06:57,583 things like spotting knowledge 240 01:06:57,583 --> 01:06:59,541 gaps, right, helping you write, you know, 241 01:06:59,541 --> 01:07:00,750 keeping your your 242 01:07:00,750 --> 01:07:02,000 documentation up to date, 243 01:07:02,000 --> 01:07:02,958 these are all things that 244 01:07:02,958 --> 01:07:04,625 we can start also working on. 245 01:07:05,041 --> 01:07:05,791 So this concept of 246 01:07:05,791 --> 01:07:07,875 agent is very versatile. 247 01:07:08,208 --> 01:07:10,041 The agents can solve issues, they can, 248 01:07:10,041 --> 01:07:11,500 you know, answer emails, 249 01:07:11,500 --> 01:07:12,541 but they can also start 250 01:07:12,541 --> 01:07:14,041 being tasked of gathering 251 01:07:14,041 --> 01:07:15,375 insights and producing new 252 01:07:15,375 --> 01:07:16,333 knowledge and keeping your 253 01:07:16,333 --> 01:07:17,625 system of record, 254 01:07:17,625 --> 01:07:18,458 always up to date. 255 01:07:18,458 --> 01:07:19,708 So there's really a lot of 256 01:07:19,708 --> 01:07:20,875 things that need to happen. 257 01:07:20,875 --> 01:07:22,041 You cannot approach the 258 01:07:22,041 --> 01:07:23,041 problem just from one 259 01:07:23,041 --> 01:07:24,750 angle, you really want to tie 260 01:07:24,750 --> 01:07:26,291 the whole thing together, because this is 261 01:07:26,291 --> 01:07:27,083 where you start creating 262 01:07:27,083 --> 01:07:28,250 these positive feedback 263 01:07:28,291 --> 01:07:30,958 loops, so called data flywheels, right, 264 01:07:31,041 --> 01:07:32,208 that enable you to 265 01:07:32,208 --> 01:07:33,416 start deploying the solution 266 01:07:33,416 --> 01:07:34,625 and trust it that it's 267 01:07:34,625 --> 01:07:35,750 operating really well. 268 01:07:36,458 --> 01:07:37,083 So yeah, there's, 269 01:07:37,083 --> 01:07:38,625 there's a lot of things to it. 270 01:07:38,666 --> 01:07:41,458 That sounds fascinating, and also quite 271 01:07:41,458 --> 01:07:42,958 complex in terms of the 272 01:07:42,958 --> 01:07:44,708 different surface areas, etc. 273 01:07:45,208 --> 01:07:46,083 So maybe we can go 274 01:07:46,083 --> 01:07:47,541 slightly more technical. 275 01:07:47,791 --> 01:07:49,791 And before we go into all of the Effect's 276 01:07:49,791 --> 01:07:52,416 specifics, can you more from like an 277 01:07:52,416 --> 01:07:53,458 engineering perspective 278 01:07:53,750 --> 01:07:56,000 provide an overview over your system? 279 01:07:56,875 --> 01:07:59,000 And which different high level components 280 01:07:59,000 --> 01:08:00,958 it's composed out of? 281 01:08:01,708 --> 01:08:01,916 Sure. 282 01:08:02,125 --> 01:08:04,875 So this is a, you know, apart from the 283 01:08:04,875 --> 01:08:06,666 machine learning things, 284 01:08:06,666 --> 01:08:08,041 the LLM things, this is a 285 01:08:08,041 --> 01:08:09,625 classical, you know, setup. 286 01:08:09,625 --> 01:08:11,041 there's some servers, both 287 01:08:11,083 --> 01:08:13,500 like short, like serverless instances for 288 01:08:13,500 --> 01:08:14,875 API endpoints, then 289 01:08:14,875 --> 01:08:15,750 there's some long running 290 01:08:16,166 --> 01:08:18,875 servers to do things like, you know, pre 291 01:08:18,875 --> 01:08:21,333 processing data, post 292 01:08:21,333 --> 01:08:23,791 processing, interactions. 293 01:08:23,916 --> 01:08:25,041 So yeah, there's a there's a bunch of 294 01:08:25,041 --> 01:08:25,875 servers here, everything 295 01:08:25,875 --> 01:08:27,041 is in node, everything is 296 01:08:27,291 --> 01:08:28,375 or TypeScript. 297 01:08:29,041 --> 01:08:32,125 The front end is a React application. 298 01:08:32,583 --> 01:08:35,958 So both we have a dashboard, which is a V 299 01:08:35,958 --> 01:08:37,583 duplication, there is 300 01:08:37,583 --> 01:08:38,708 some integrations 301 01:08:38,708 --> 01:08:39,916 inside of Salesforce, 302 01:08:39,916 --> 01:08:41,750 Zendesk, Slack, and so on. 303 01:08:41,750 --> 01:08:43,875 So it's a classical full 304 01:08:43,875 --> 01:08:46,625 stack, React TypeScript application. 305 01:08:47,541 --> 01:08:49,750 So you've built quite an expansive system 306 01:08:49,750 --> 01:08:51,375 here and building all 307 01:08:51,375 --> 01:08:52,250 of that in TypeScript 308 01:08:52,291 --> 01:08:54,291 I think TypeScript is sort of notorious 309 01:08:54,291 --> 01:08:55,958 for only scratching the 310 01:08:55,958 --> 01:08:57,250 surface and only covering 311 01:08:57,250 --> 01:08:58,083 the happy path. 312 01:08:58,750 --> 01:09:00,541 There's a lot of things that can go wrong 313 01:09:00,541 --> 01:09:03,208 in all of those systems, integrating with 314 01:09:03,208 --> 01:09:04,125 different systems. 315 01:09:04,750 --> 01:09:06,916 All of those things I think have led you 316 01:09:06,916 --> 01:09:08,666 to take a more serious 317 01:09:08,666 --> 01:09:10,166 look at Effect and ultimately 318 01:09:10,208 --> 01:09:10,958 adopting it. 319 01:09:11,375 --> 01:09:14,708 So can you share that journey of going 320 01:09:14,708 --> 01:09:16,041 from a non-Effect code 321 01:09:16,041 --> 01:09:18,583 base to converting step 322 01:09:18,583 --> 01:09:19,458 by step to Effect? 323 01:09:20,166 --> 01:09:22,875 Sure, so obviously there's the common 324 01:09:22,875 --> 01:09:24,791 complexity increase that 325 01:09:24,791 --> 01:09:25,833 happens when you're building 326 01:09:25,833 --> 01:09:27,791 a product and putting it into production. 327 01:09:27,791 --> 01:09:29,833 So as you add features and so on, then 328 01:09:29,833 --> 01:09:31,875 you hit sort of a ceiling 329 01:09:31,875 --> 01:09:33,333 and then you start assessing 330 01:09:33,333 --> 01:09:35,333 the situation and then you might start 331 01:09:35,333 --> 01:09:36,708 refactoring and so on. 332 01:09:36,708 --> 01:09:39,958 And then next phase, you hit some other 333 01:09:39,958 --> 01:09:41,250 limitations and then you 334 01:09:41,250 --> 01:09:43,291 reassess again the situation. 335 01:09:44,041 --> 01:09:45,916 So that's just the common thing of 336 01:09:45,916 --> 01:09:46,833 working on an 337 01:09:46,833 --> 01:09:48,458 increasingly growing code base. 338 01:09:49,500 --> 01:09:51,791 But then there's LLMs and these are 339 01:09:51,791 --> 01:09:54,958 non-deterministic systems and it poses 340 01:09:54,958 --> 01:09:56,166 some completely new kinds 341 01:09:56,166 --> 01:09:56,833 of challenges. 342 01:09:56,833 --> 01:09:58,916 How do you monitor the behavior of the 343 01:09:58,916 --> 01:09:59,666 system, for instance? 344 01:10:00,500 --> 01:10:02,708 How do you deal with readable streams 345 01:10:02,708 --> 01:10:04,958 instead of just sending over a payload? 346 01:10:05,416 --> 01:10:07,333 This has the potential to making your 347 01:10:07,333 --> 01:10:08,375 code base more complex 348 01:10:08,375 --> 01:10:10,666 if you need to stream some 349 01:10:10,666 --> 01:10:12,375 tokens down to the customer, but you 350 01:10:12,375 --> 01:10:13,375 actually also need to 351 01:10:13,375 --> 01:10:16,125 store these chunks on your own 352 01:10:16,125 --> 01:10:16,916 for post-processing. 353 01:10:17,291 --> 01:10:18,125 How do you deal with that? 354 01:10:18,625 --> 01:10:19,416 TypeScript doesn't give you 355 01:10:19,416 --> 01:10:22,041 some great answers to this. 356 01:10:22,458 --> 01:10:24,166 How do you deal with reliability? 357 01:10:24,750 --> 01:10:27,875 OpenAI and Anthropic and so on, they offer 358 01:10:27,875 --> 01:10:30,291 a service which is in high demand. 359 01:10:31,000 --> 01:10:33,125 We are in a period right 360 01:10:33,125 --> 01:10:34,416 now with scarce resources. 361 01:10:35,250 --> 01:10:37,458 So these APIs, they go down and they 362 01:10:37,458 --> 01:10:38,875 actually go down more 363 01:10:38,875 --> 01:10:39,875 than you would think. 364 01:10:40,250 --> 01:10:42,458 And so how do you build fallback 365 01:10:42,458 --> 01:10:44,833 mechanisms so that, say that 366 01:10:44,833 --> 01:10:46,125 OpenAI is down as your main 367 01:10:46,125 --> 01:10:48,833 provider, your customers cannot be down. 368 01:10:48,833 --> 01:10:49,750 If OpenAI is down, 369 01:10:49,750 --> 01:10:50,541 that's not their problem. 370 01:10:50,791 --> 01:10:51,416 That's our problem. 371 01:10:52,125 --> 01:10:53,541 So how do you build a 372 01:10:53,541 --> 01:10:55,458 fallback system to Anthropic? 373 01:10:55,458 --> 01:10:58,625 What happens a few weeks ago when OpenAI 374 01:10:58,625 --> 01:10:59,916 was down for, I think, 375 01:11:00,041 --> 01:11:01,208 five hours and then the 376 01:11:01,208 --> 01:11:04,250 entire industry shifts towards Anthropic 377 01:11:04,250 --> 01:11:05,166 that just has this 378 01:11:05,166 --> 01:11:06,958 spike that they're unable 379 01:11:07,000 --> 01:11:08,708 to deal with? 380 01:11:08,708 --> 01:11:11,583 So they are experiencing outage or just 381 01:11:11,583 --> 01:11:13,125 very degraded performance. 382 01:11:13,750 --> 01:11:15,541 There's outage and then there's actually 383 01:11:15,541 --> 01:11:16,500 performance, the 384 01:11:16,500 --> 01:11:19,083 speed of tokens and so on. 385 01:11:19,958 --> 01:11:22,916 All that, there's no good answer, no 386 01:11:22,916 --> 01:11:25,750 clear best practice in 387 01:11:25,750 --> 01:11:27,125 vanilla TypeScript to deal 388 01:11:27,125 --> 01:11:27,666 with these things. 389 01:11:28,208 --> 01:11:29,916 And then there's the fact that we are a 390 01:11:29,916 --> 01:11:31,541 startup working in a 391 01:11:31,541 --> 01:11:33,250 fast-paced environment where 392 01:11:33,250 --> 01:11:35,625 things change all the time. 393 01:11:36,166 --> 01:11:37,750 And we want to be efficient and we want 394 01:11:37,750 --> 01:11:39,291 to be able to not deal 395 01:11:39,291 --> 01:11:40,916 with all the complexity 396 01:11:41,125 --> 01:11:42,916 and all the problems and down times and 397 01:11:42,916 --> 01:11:44,833 reliability, but we need to 398 01:11:44,833 --> 01:11:46,458 be effective in shipping new 399 01:11:46,458 --> 01:11:47,708 features to constantly 400 01:11:47,708 --> 01:11:49,291 meet up with the demand. 401 01:11:49,875 --> 01:11:50,625 the zero to one of 402 01:11:50,625 --> 01:11:51,875 building a company here. 403 01:11:51,875 --> 01:11:54,125 So there's multiple factors that really 404 01:11:54,125 --> 01:11:55,833 made us think hard about 405 01:11:55,833 --> 01:11:57,916 what would be the foundation 406 01:11:58,083 --> 01:12:00,625 for us to be able to operate really fast 407 01:12:00,625 --> 01:12:02,125 with a small set of 408 01:12:02,125 --> 01:12:04,416 resources from a startup and 409 01:12:04,500 --> 01:12:07,458 be a highly reliable solution because we 410 01:12:07,458 --> 01:12:08,958 are catering to larger 411 01:12:08,958 --> 01:12:10,083 companies, enterprise 412 01:12:10,083 --> 01:12:10,541 companies. 413 01:12:10,875 --> 01:12:13,333 So reliability is just really key. 414 01:12:13,791 --> 01:12:15,500 Obviously, I can talk a lot about the 415 01:12:15,500 --> 01:12:18,750 whole telemetry part of 416 01:12:18,750 --> 01:12:20,041 things, making sure everything 417 01:12:20,041 --> 01:12:21,708 is properly monitored and that we can 418 01:12:21,708 --> 01:12:24,041 actually spot when things 419 01:12:24,041 --> 01:12:25,750 go wrong before our customers 420 01:12:25,750 --> 01:12:28,125 even see it and then we can take action. 421 01:12:28,125 --> 01:12:29,333 So there's a bunch of things here. 422 01:12:29,833 --> 01:12:31,708 And Effect just came at a very good 423 01:12:31,708 --> 01:12:33,916 moment where we still had 424 01:12:33,916 --> 01:12:36,166 a sufficiently small code 425 01:12:36,166 --> 01:12:38,291 base that we could actually start 426 01:12:38,291 --> 01:12:39,750 seriously thinking about 427 01:12:39,750 --> 01:12:40,958 completely converting it to 428 01:12:40,958 --> 01:12:41,125 Effect. 429 01:12:41,666 --> 01:12:42,916 So the timing was really good. 430 01:12:43,750 --> 01:12:45,291 So you've mentioned in regards to 431 01:12:45,291 --> 01:12:46,750 adopting Effect for the 432 01:12:46,750 --> 01:12:48,375 sake of improving reliability, 433 01:12:49,083 --> 01:12:52,250 one specific use case was that Markbrom 434 01:12:52,250 --> 01:12:54,625 is using LLMs such as 435 01:12:54,625 --> 01:12:56,125 the products provided by 436 01:12:56,125 --> 01:12:56,541 OpenAI. 437 01:12:56,958 --> 01:12:58,958 And if they go down, you need to fall 438 01:12:58,958 --> 01:13:00,041 back to another system. 439 01:13:00,250 --> 01:13:02,208 And in case that system also goes down, 440 01:13:02,541 --> 01:13:03,791 fall back to yet another one. 441 01:13:04,333 --> 01:13:07,000 All of that is helping to improve the 442 01:13:07,000 --> 01:13:08,916 reliability of Markprompt. 443 01:13:10,291 --> 01:13:12,125 Which other challenges did you face 444 01:13:12,125 --> 01:13:14,041 similar to that one in 445 01:13:14,041 --> 01:13:15,458 regards to reliability and 446 01:13:15,458 --> 01:13:16,791 how did you improve 447 01:13:16,791 --> 01:13:18,125 reliability through Effect? 448 01:13:19,166 --> 01:13:21,875 I mean, there's lots of examples where 449 01:13:21,875 --> 01:13:23,750 Effect has helped us 450 01:13:23,750 --> 01:13:24,875 consolidate our system. 451 01:13:25,125 --> 01:13:28,333 One example is one day we woke up and we 452 01:13:28,333 --> 01:13:29,250 realized that we were under 453 01:13:29,250 --> 01:13:32,208 attack by a malicious hacker 454 01:13:32,375 --> 01:13:35,750 who was interacting with a public facing 455 01:13:35,750 --> 01:13:37,916 deployment of Markprompt and 456 01:13:37,916 --> 01:13:40,208 sending, I think, 20 million 457 01:13:40,208 --> 01:13:43,166 requests per hour. 458 01:13:43,500 --> 01:13:46,166 And this was across the globe, so it was 459 01:13:46,166 --> 01:13:47,708 kind of hard to deal with. 460 01:13:47,708 --> 01:13:50,458 So Ethan Niser, he spent the summer with 461 01:13:50,458 --> 01:13:53,583 us and he's been working 462 01:13:53,583 --> 01:13:54,750 on a lot of Effect-related 463 01:13:54,833 --> 01:13:56,666 things and on his first day, 464 01:13:56,833 --> 01:13:58,708 this attack started happening. 465 01:13:58,708 --> 01:14:02,000 And so he built this great rate limiting 466 01:14:02,000 --> 01:14:03,291 system in the middleware 467 01:14:03,291 --> 01:14:04,708 that is capable of looking 468 01:14:04,708 --> 01:14:08,250 at multiple facets of a payload. 469 01:14:08,250 --> 01:14:09,541 And the interesting thing here is that 470 01:14:09,541 --> 01:14:11,000 you cannot just rely 471 01:14:11,000 --> 01:14:12,791 on IP and fingerprinting 472 01:14:12,791 --> 01:14:13,250 and so on. 473 01:14:13,250 --> 01:14:14,833 You actually also want to start looking 474 01:14:14,833 --> 01:14:16,375 at the LLM payload itself. 475 01:14:16,375 --> 01:14:18,125 What kinds of messages is there, pattern 476 01:14:18,125 --> 01:14:20,000 that we can spot here in 477 01:14:20,000 --> 01:14:22,250 the nature of the requests? 478 01:14:22,250 --> 01:14:24,208 And so it's a multi-tiered rate limiting 479 01:14:24,208 --> 01:14:25,750 system that is very easy 480 01:14:25,750 --> 01:14:27,208 to automatically trigger 481 01:14:27,208 --> 01:14:28,375 and also manually trigger. 482 01:14:28,375 --> 01:14:30,541 So this is an example also where building 483 01:14:30,541 --> 01:14:33,916 with composability has just allowed us to 484 01:14:34,250 --> 01:14:35,375 build a fairly sophisticated 485 01:14:35,375 --> 01:14:37,458 system here, which just works. 486 01:14:38,291 --> 01:14:39,583 We are now very 487 01:14:39,583 --> 01:14:40,750 confident that it just works. 488 01:14:40,750 --> 01:14:42,083 And by the way, one of the reasons that 489 01:14:42,083 --> 01:14:43,083 we're confident that it 490 01:14:43,083 --> 01:14:44,333 works is that it's very 491 01:14:44,333 --> 01:14:46,250 easy to test because the whole dependency 492 01:14:46,250 --> 01:14:47,791 injection story is 493 01:14:47,791 --> 01:14:48,958 just such a natural thing 494 01:14:48,958 --> 01:14:49,666 to do. 495 01:14:49,666 --> 01:14:51,291 You just operate a provider and you can 496 01:14:51,291 --> 01:14:53,458 very easily simulate a 497 01:14:53,458 --> 01:14:55,291 situation that is hard to 498 01:14:55,291 --> 01:14:56,541 reproduce with other means. 499 01:14:56,583 --> 01:14:59,250 So dependency injection, I would say, is 500 01:14:59,250 --> 01:15:00,458 a huge factor to the 501 01:15:00,458 --> 01:15:01,791 reliability because we 502 01:15:01,791 --> 01:15:05,458 can restart testing all sorts of 503 01:15:05,458 --> 01:15:06,583 different combinations of 504 01:15:06,583 --> 01:15:08,583 non-happy paths and building 505 01:15:08,583 --> 01:15:12,083 that with just change or service that you 506 01:15:12,083 --> 01:15:13,208 provide is just very, 507 01:15:13,208 --> 01:15:14,875 very easy with Effect. 508 01:15:14,875 --> 01:15:15,958 Telemetry is obviously a thing. 509 01:15:16,000 --> 01:15:19,750 So I think that the whole testing story 510 01:15:19,750 --> 01:15:22,000 is not just about one thing. 511 01:15:22,000 --> 01:15:23,208 You want to do unit tests. 512 01:15:23,208 --> 01:15:25,333 You want to do integration tests with 513 01:15:25,333 --> 01:15:26,500 dependency injection. 514 01:15:26,500 --> 01:15:28,166 You also really want to have a full 515 01:15:28,166 --> 01:15:29,083 understanding of what's going 516 01:15:29,083 --> 01:15:30,250 on when you are in production. 517 01:15:31,375 --> 01:15:33,083 And so being able to have good traces, 518 01:15:33,750 --> 01:15:35,000 good logging of every 519 01:15:35,000 --> 01:15:36,208 component of your system 520 01:15:36,666 --> 01:15:37,500 is non-trivial. 521 01:15:38,625 --> 01:15:41,708 And so we rely heavily on telemetry and 522 01:15:41,708 --> 01:15:43,875 Effect allows us to just 523 01:15:43,875 --> 01:15:45,750 add spans everywhere in 524 01:15:45,750 --> 01:15:46,916 a very, very simple way. 525 01:15:46,916 --> 01:15:48,333 And it just naturally fits 526 01:15:48,333 --> 01:15:49,791 into the rest of the system. 527 01:15:50,125 --> 01:15:51,791 And this really has allowed us to have a 528 01:15:51,791 --> 01:15:52,541 very, very good 529 01:15:52,541 --> 01:15:55,125 understanding of failure scenarios 530 01:15:55,166 --> 01:15:56,916 before our customers even see them. 531 01:15:57,333 --> 01:15:59,750 So these are just general patterns that 532 01:15:59,750 --> 01:16:01,958 are very natural to do 533 01:16:01,958 --> 01:16:02,875 in Effect that we don't 534 01:16:02,916 --> 01:16:05,583 even think about because they actually 535 01:16:05,583 --> 01:16:07,333 don't incur any sort of 536 01:16:07,333 --> 01:16:08,958 additional thinking on our 537 01:16:08,958 --> 01:16:09,333 part. 538 01:16:09,333 --> 01:16:11,083 It's just adding a line 539 01:16:11,083 --> 01:16:12,333 and then it just works. 540 01:16:12,875 --> 01:16:14,166 Yeah, this makes a lot of sense. 541 01:16:14,458 --> 01:16:17,875 I can totally see how reliability is 542 01:16:17,875 --> 01:16:20,250 almost sort of like a emerging factor, 543 01:16:20,416 --> 01:16:22,916 like a second order effect that comes 544 01:16:22,916 --> 01:16:25,541 when you have a composable system, 545 01:16:25,791 --> 01:16:28,791 and when you have observability, when you 546 01:16:28,791 --> 01:16:30,541 can look into what's going on, 547 01:16:30,875 --> 01:16:32,583 when you can eat your 548 01:16:32,583 --> 01:16:34,250 vegetables, then you're healthy. 549 01:16:34,750 --> 01:16:35,291 And so 550 01:16:35,291 --> 01:16:37,583 Building a reliable system is going the 551 01:16:37,583 --> 01:16:39,333 extra mile on all of those things. 552 01:16:39,333 --> 01:16:40,333 This makes a lot of sense. 553 01:16:41,375 --> 01:16:42,416 Given that you built 554 01:16:42,416 --> 01:16:44,541 quite an expansive system, 555 01:16:44,916 --> 01:16:46,875 maybe we can try to break it down a 556 01:16:46,875 --> 01:16:49,083 little bit more with a technical lens on. 557 01:16:49,708 --> 01:16:51,916 So I'm sure that you have 558 01:16:51,916 --> 01:16:53,166 a client side application. 559 01:16:53,291 --> 01:16:54,916 You've mentioned that it's like a single 560 01:16:54,916 --> 01:16:56,500 page app built with Vite, 561 01:16:57,208 --> 01:16:59,083 but more on the backend side. 562 01:16:59,583 --> 01:17:02,041 Can you describe to me what are the 563 01:17:02,041 --> 01:17:04,041 different subcomponents of your system? 564 01:17:04,541 --> 01:17:07,875 Yeah, so in a very simple way, the main 565 01:17:07,875 --> 01:17:09,500 subcomponents here, 566 01:17:09,500 --> 01:17:11,166 there's the application server, 567 01:17:11,166 --> 01:17:12,708 which is essentially an RPC 568 01:17:12,708 --> 01:17:14,541 server that the client 569 01:17:14,541 --> 01:17:15,833 application is speaking with. 570 01:17:15,833 --> 01:17:17,750 Then we have a public API. 571 01:17:18,458 --> 01:17:19,833 So Markprompt is built as 572 01:17:19,833 --> 01:17:21,166 Stripe for customer support, 573 01:17:21,791 --> 01:17:23,375 in the sense that every part of the 574 01:17:23,375 --> 01:17:24,208 system can be 575 01:17:24,208 --> 01:17:26,166 interacted with via our APIs 576 01:17:26,916 --> 01:17:28,708 so that you can actually build deeper 577 01:17:28,708 --> 01:17:29,666 product experiences 578 01:17:29,666 --> 01:17:31,500 inside of your core product. 579 01:17:32,166 --> 01:17:34,208 So we have a public facing API server. 580 01:17:34,208 --> 01:17:37,625 Then we have the ingestion engine, which 581 01:17:37,625 --> 01:17:40,250 takes in all your systems data. 582 01:17:40,250 --> 01:17:41,875 So this can be public facing like a 583 01:17:41,875 --> 01:17:42,833 website that we would 584 01:17:42,833 --> 01:17:44,375 scrape on a regular basis. 585 01:17:44,750 --> 01:17:47,541 It could be GitHub repository or 586 01:17:47,541 --> 01:17:49,250 historical tickets from 587 01:17:49,250 --> 01:17:51,500 Zendesk and Jira and Confluence 588 01:17:51,541 --> 01:17:53,125 and databases from 589 01:17:53,125 --> 01:17:54,750 Salesforce and whatnot. 590 01:17:55,125 --> 01:17:57,458 So we do a bunch of things here. 591 01:17:57,458 --> 01:17:58,291 We import data. 592 01:17:58,291 --> 01:17:58,875 We chunk it up. 593 01:17:58,916 --> 01:18:02,041 We monitor the whole structure of 594 01:18:02,041 --> 01:18:03,333 documents through the AST 595 01:18:03,333 --> 01:18:04,083 and then we build 596 01:18:04,083 --> 01:18:06,000 embeddings for multi-levels. 597 01:18:06,541 --> 01:18:07,291 And then we store them 598 01:18:07,291 --> 01:18:08,166 in a vector database. 599 01:18:08,708 --> 01:18:09,583 And then we have all the 600 01:18:09,583 --> 01:18:11,416 application related servers, 601 01:18:11,416 --> 01:18:13,416 so things like Discord or 602 01:18:13,416 --> 01:18:15,625 Slack or Zendesk and so on, 603 01:18:15,666 --> 01:18:18,541 which are more like simple servers here. 604 01:18:18,541 --> 01:18:18,916 Got it. 605 01:18:18,916 --> 01:18:20,541 Yeah, that makes a lot of sense to break 606 01:18:20,541 --> 01:18:22,000 it up into those different pieces. 607 01:18:22,500 --> 01:18:24,125 I'd actually love to learn a 608 01:18:24,125 --> 01:18:25,333 little bit more about each, 609 01:18:25,791 --> 01:18:26,833 maybe starting with 610 01:18:26,833 --> 01:18:28,083 the application server. 611 01:18:28,333 --> 01:18:29,416 So you mentioned that you're 612 01:18:29,416 --> 01:18:31,166 using Effect RPC in that regard. 613 01:18:31,625 --> 01:18:33,208 What did you use before 614 01:18:33,208 --> 01:18:34,208 you were using Effect? 615 01:18:34,791 --> 01:18:37,500 And what are the qualities of using 616 01:18:37,500 --> 01:18:39,125 Effect for an 617 01:18:39,125 --> 01:18:41,166 application for an RPC server here? 618 01:18:41,958 --> 01:18:42,416 Yeah, sure. 619 01:18:43,000 --> 01:18:44,791 I mean, this before was 620 01:18:44,791 --> 01:18:46,458 just like simple API handlers. 621 01:18:46,750 --> 01:18:48,166 So you would have two 622 01:18:48,166 --> 01:18:51,041 completely decoupled systems. 623 01:18:51,916 --> 01:18:53,333 And now pretty much all the 624 01:18:53,333 --> 01:18:56,041 endpoints are in full RPC. 625 01:18:56,500 --> 01:18:57,958 It starts with a schema, actually. 626 01:18:58,500 --> 01:19:01,375 We define the input schema, the output 627 01:19:01,375 --> 01:19:03,041 schema, and the error schema 628 01:19:03,375 --> 01:19:04,666 for each of these endpoints. 629 01:19:05,250 --> 01:19:07,125 And then it trickles down all the way to 630 01:19:07,125 --> 01:19:07,833 the client application. 631 01:19:08,125 --> 01:19:08,708 So we have this. 632 01:19:09,083 --> 01:19:10,500 Ethan actually wrote this and I hope he's 633 01:19:10,500 --> 01:19:11,208 going to publish it. 634 01:19:11,583 --> 01:19:13,750 Sometimes soon, essentially React query, 635 01:19:13,750 --> 01:19:15,375 but wrapped around with Effect 636 01:19:15,458 --> 01:19:17,541 so that we have the full schema from the 637 01:19:17,541 --> 01:19:18,416 backend all the way to 638 01:19:18,416 --> 01:19:19,375 the client side code. 639 01:19:19,791 --> 01:19:21,791 And we have nothing more to deal with. 640 01:19:21,791 --> 01:19:22,875 So the whole thing is completely 641 01:19:22,875 --> 01:19:24,583 streamlined in terms of the 642 01:19:24,583 --> 01:19:25,916 communication between the client 643 01:19:26,250 --> 01:19:27,375 and the server. 644 01:19:28,125 --> 01:19:28,958 That makes a lot of sense. 645 01:19:29,166 --> 01:19:30,458 And so before you 646 01:19:30,458 --> 01:19:32,208 didn't have any schema layer, 647 01:19:32,833 --> 01:19:35,291 anything to express the kind of data 648 01:19:35,291 --> 01:19:36,958 you're sending across? 649 01:19:37,791 --> 01:19:38,500 No, we didn't actually. 650 01:19:39,375 --> 01:19:39,708 Got it. 651 01:19:39,916 --> 01:19:42,291 Well, I suppose that through the process 652 01:19:42,291 --> 01:19:43,750 of just modeling the schema, 653 01:19:44,458 --> 01:19:46,750 I'm sure that there were a couple of code 654 01:19:46,750 --> 01:19:47,666 paths that you didn't 655 01:19:47,666 --> 01:19:48,833 even account for before. 656 01:19:49,833 --> 01:19:50,583 So the-- 657 01:19:50,583 --> 01:19:50,916 Absolutely. 658 01:19:51,750 --> 01:19:55,000 The public API server where you're 659 01:19:55,000 --> 01:19:58,791 exposing an open API spec, I suppose, 660 01:19:59,083 --> 01:20:01,583 how similar or different is that to 661 01:20:01,583 --> 01:20:02,708 your application server? 662 01:20:02,708 --> 01:20:03,958 Can you describe the public 663 01:20:03,958 --> 01:20:05,583 API server setup a bit more? 664 01:20:06,458 --> 01:20:06,875 Yeah, sure. 665 01:20:06,875 --> 01:20:08,333 So this is more of a 666 01:20:08,333 --> 01:20:10,583 standard HTTP server. 667 01:20:10,583 --> 01:20:11,958 This is actually using Effect HTTP. 668 01:20:12,333 --> 01:20:13,416 So again, we have the full-- 669 01:20:13,958 --> 01:20:17,000 Everything starts with the Effect runtime 670 01:20:17,000 --> 01:20:18,333 taking over directly 671 01:20:18,333 --> 01:20:19,291 at the request level. 672 01:20:19,916 --> 01:20:20,958 So again, we can model 673 01:20:20,958 --> 01:20:23,083 everything very nicely via schemas. 674 01:20:23,125 --> 01:20:26,291 And one really nice thing of having 675 01:20:26,291 --> 01:20:28,541 everything defined by schemas is 676 01:20:28,708 --> 01:20:30,250 that we can, for instance, automatically 677 01:20:30,250 --> 01:20:31,833 generate the open API spec. 678 01:20:32,250 --> 01:20:34,166 So this is something that we now have 679 01:20:34,166 --> 01:20:35,375 every single endpoint is 680 01:20:35,375 --> 01:20:36,458 automatically documented 681 01:20:37,166 --> 01:20:38,250 just via the schema. 682 01:20:39,208 --> 01:20:39,583 Got it. 683 01:20:39,958 --> 01:20:41,833 How much overlap does your 684 01:20:41,833 --> 01:20:43,458 internal application server, 685 01:20:43,916 --> 01:20:45,500 that RPC server with 686 01:20:45,500 --> 01:20:47,208 the public API server, 687 01:20:47,500 --> 01:20:49,000 does it have a lot of overlap? 688 01:20:49,000 --> 01:20:50,958 Do you reuse a lot of code internally? 689 01:20:51,083 --> 01:20:52,291 Or are those rather 690 01:20:52,291 --> 01:20:54,125 fulfilling different purposes? 691 01:20:54,958 --> 01:20:55,875 They're fulfilling 692 01:20:55,875 --> 01:20:57,458 completely different purposes, yes. 693 01:20:58,625 --> 01:21:00,000 So yeah, I mean, these things are 694 01:21:00,000 --> 01:21:01,541 completely separate. 695 01:21:02,375 --> 01:21:04,000 So to give you one very important 696 01:21:04,000 --> 01:21:07,791 differentiator is that on the RPC server, 697 01:21:07,958 --> 01:21:09,750 we provide permissions. 698 01:21:10,666 --> 01:21:12,875 So you want to make sure that a given 699 01:21:12,875 --> 01:21:14,875 user can perform an action 700 01:21:14,875 --> 01:21:16,000 that they are allowed to. 701 01:21:16,000 --> 01:21:19,791 And so we have a very nice system that 702 01:21:19,791 --> 01:21:21,583 provides the level of 703 01:21:21,583 --> 01:21:23,166 permissions of a user. 704 01:21:23,166 --> 01:21:25,250 This is completely transparent. 705 01:21:25,250 --> 01:21:28,583 We actually used RLS before, 706 01:21:28,583 --> 01:21:29,625 role level security to 707 01:21:29,625 --> 01:21:30,791 enforce these policies. 708 01:21:31,333 --> 01:21:33,333 But there was a real performance impact. 709 01:21:33,541 --> 01:21:34,708 It was just slowing down the crazy, 710 01:21:34,916 --> 01:21:36,958 especially when you're dealing with 711 01:21:36,958 --> 01:21:39,250 hundreds of thousands of data points. 712 01:21:39,500 --> 01:21:41,083 So this had a real impact. 713 01:21:41,333 --> 01:21:42,500 So we got rid of that. 714 01:21:42,791 --> 01:21:43,583 And actually, everything 715 01:21:43,583 --> 01:21:44,875 now is modeled in code. 716 01:21:44,916 --> 01:21:46,666 So the permission, which I guess was 717 01:21:46,666 --> 01:21:48,291 always how things have 718 01:21:48,291 --> 01:21:51,041 been done without RLS. 719 01:21:52,208 --> 01:21:54,333 And so this is, again, an example where 720 01:21:54,333 --> 01:21:56,416 Effect at the type level, 721 01:21:56,416 --> 01:21:58,500 you can immediately spot whether this is 722 01:21:58,500 --> 01:22:00,916 a user that is allowed to 723 01:22:00,916 --> 01:22:02,958 perform a single action, 724 01:22:02,958 --> 01:22:04,125 because that service 725 01:22:04,125 --> 01:22:05,875 requires these sets of permissions. 726 01:22:06,500 --> 01:22:07,333 And so this is very, 727 01:22:07,333 --> 01:22:08,458 very nice to model as well. 728 01:22:09,708 --> 01:22:10,583 That sounds very interesting. 729 01:22:10,583 --> 01:22:13,083 So you've built yourself an internal DSL 730 01:22:13,083 --> 01:22:16,500 to model those permission rules. 731 01:22:16,833 --> 01:22:18,791 Can you explain a little bit more how 732 01:22:18,791 --> 01:22:20,041 that's implemented and how 733 01:22:20,041 --> 01:22:21,833 that can be imagined from 734 01:22:22,500 --> 01:22:23,541 like a type level 735 01:22:23,541 --> 01:22:25,583 perspective or from an API perspective? 736 01:22:26,833 --> 01:22:27,000 Yeah. 737 01:22:27,000 --> 01:22:28,583 I mean, so every user comes 738 01:22:28,583 --> 01:22:30,125 with a set of permissions, right? 739 01:22:30,125 --> 01:22:30,750 And these permissions 740 01:22:30,750 --> 01:22:32,916 can be modeled via types. 741 01:22:33,666 --> 01:22:38,750 And services can declare what level of 742 01:22:38,750 --> 01:22:40,958 permission they required in 743 01:22:40,958 --> 01:22:42,166 order to perform an action. 744 01:22:42,791 --> 01:22:45,041 And so we immediately know if within an 745 01:22:45,041 --> 01:22:47,416 RPC endpoint, we're using a service, 746 01:22:47,666 --> 01:22:48,583 but that the provided 747 01:22:48,583 --> 01:22:51,166 permissions don't match. 748 01:22:52,000 --> 01:22:53,500 Well, then we can flag it. 749 01:22:53,500 --> 01:22:55,083 And so we can guarantee that there's no 750 01:22:55,083 --> 01:22:56,458 code being shipped where 751 01:22:56,458 --> 01:22:57,833 there's an inconsistency here. 752 01:22:59,000 --> 01:22:59,541 I love it. 753 01:22:59,541 --> 01:23:00,500 That sounds very elegant. 754 01:23:00,791 --> 01:23:02,833 I would love to see some code for that. 755 01:23:03,541 --> 01:23:05,708 Could you imagine that there could be a 756 01:23:05,708 --> 01:23:08,833 more commonly reusable package for that? 757 01:23:08,833 --> 01:23:10,750 Or is that only something that makes 758 01:23:10,750 --> 01:23:12,000 sense within your code base? 759 01:23:13,583 --> 01:23:14,250 Good question. 760 01:23:15,916 --> 01:23:18,958 Obviously, we have been just crunching 761 01:23:18,958 --> 01:23:19,666 through this code and 762 01:23:19,666 --> 01:23:20,458 getting into production. 763 01:23:20,666 --> 01:23:23,500 So we haven't thought too much about how 764 01:23:23,500 --> 01:23:24,625 this would extend to a 765 01:23:24,625 --> 01:23:27,166 more generic use case. 766 01:23:27,208 --> 01:23:29,541 But I guess this could be 767 01:23:29,541 --> 01:23:30,625 a very interesting topic. 768 01:23:31,541 --> 01:23:31,750 Right. 769 01:23:31,875 --> 01:23:34,208 Well, anyone in the audience is happily 770 01:23:34,208 --> 01:23:35,666 invited to try to 771 01:23:35,666 --> 01:23:37,083 replicate this by themselves. 772 01:23:37,708 --> 01:23:39,166 Maybe that's even something that we 773 01:23:39,166 --> 01:23:42,041 could, as part of the Effect org, also 774 01:23:42,041 --> 01:23:43,083 look into providing. 775 01:23:43,666 --> 01:23:44,625 To hear you describing 776 01:23:44,625 --> 01:23:46,291 this use case in this DSL, 777 01:23:46,750 --> 01:23:47,541 that makes a lot of 778 01:23:47,541 --> 01:23:48,958 sense for a lot of use cases. 779 01:23:49,958 --> 01:23:52,875 So moving then to the ingestion engine, 780 01:23:53,416 --> 01:23:56,083 can you describe a little bit more of how 781 01:23:56,083 --> 01:23:57,750 that looks like from a 782 01:23:57,750 --> 01:23:58,916 lifecycle perspective? 783 01:23:59,208 --> 01:24:01,000 So you've mentioned the ingestion engine 784 01:24:01,000 --> 01:24:03,958 is mostly meant to bring in data from 785 01:24:03,958 --> 01:24:05,083 various data sources, 786 01:24:05,333 --> 01:24:05,750 whether it's 787 01:24:05,750 --> 01:24:07,458 Salesforce, whether it's Notion. 788 01:24:07,875 --> 01:24:10,625 Walk me through how that works and how 789 01:24:10,625 --> 01:24:13,000 does that work in regards to Effect? 790 01:24:13,000 --> 01:24:14,750 Do you use Effect streams for that? 791 01:24:15,666 --> 01:24:15,875 Yeah. 792 01:24:15,875 --> 01:24:16,541 How does that work? 793 01:24:17,250 --> 01:24:19,208 So it's still pretty bare bones here. 794 01:24:19,208 --> 01:24:20,541 This is something that we have been 795 01:24:20,541 --> 01:24:21,625 discussing also with the 796 01:24:21,625 --> 01:24:23,291 Effect team and the Effect cluster. 797 01:24:23,583 --> 01:24:26,541 This is where we will probably be going. 798 01:24:26,541 --> 01:24:28,958 So now we have our own system, which is 799 01:24:28,958 --> 01:24:31,041 basically just some 800 01:24:31,041 --> 01:24:32,958 long running node servers 801 01:24:32,958 --> 01:24:36,416 that are able to go and fetch the data. 802 01:24:36,708 --> 01:24:39,958 So that could be make a large request to 803 01:24:39,958 --> 01:24:42,583 an SQL database or starting a scraper 804 01:24:42,583 --> 01:24:45,541 that visits a website and when it's done 805 01:24:45,541 --> 01:24:47,625 and purged all the pages, 806 01:24:48,083 --> 01:24:49,958 it is then indexed in a 807 01:24:49,958 --> 01:24:51,750 temporary cache on our side. 808 01:24:52,083 --> 01:24:53,500 And then it goes into the post processing 809 01:24:53,500 --> 01:24:56,458 pipeline, where again, we obviously we 810 01:24:56,458 --> 01:24:57,541 parallelize all this 811 01:24:58,375 --> 01:24:59,708 because every page 812 01:24:59,708 --> 01:25:01,166 needs some post processing. 813 01:25:01,500 --> 01:25:05,500 So we chunk it up and we try to get a 814 01:25:05,500 --> 01:25:06,500 sense of the structure 815 01:25:06,500 --> 01:25:07,583 of the document, the AST, 816 01:25:08,041 --> 01:25:10,166 and then we build these embeddings for 817 01:25:10,166 --> 01:25:12,958 each levels in the AST and 818 01:25:12,958 --> 01:25:14,875 then we store it in a database. 819 01:25:15,666 --> 01:25:17,166 And so the main challenge here is that 820 01:25:17,166 --> 01:25:18,791 these are long running 821 01:25:18,791 --> 01:25:20,500 tasks which can time out, 822 01:25:20,500 --> 01:25:23,708 they can fail and whatnot. 823 01:25:24,291 --> 01:25:26,583 And so now we rely on external systems to 824 01:25:26,583 --> 01:25:28,708 do the whole workflow orchestration, 825 01:25:28,708 --> 01:25:30,083 but this is something where we are going 826 01:25:30,083 --> 01:25:30,916 to be investing some 827 01:25:30,916 --> 01:25:32,250 effort as well because 828 01:25:32,583 --> 01:25:35,708 Effect is the perfect solution here 829 01:25:35,708 --> 01:25:37,250 because we have the full 830 01:25:37,250 --> 01:25:40,125 sets of tools for dealing with 831 01:25:40,250 --> 01:25:41,833 these kinds of situations. 832 01:25:42,958 --> 01:25:45,125 Yeah, I'm really looking forward for 833 01:25:45,125 --> 01:25:46,791 Effect Cluster and Effect 834 01:25:46,791 --> 01:25:48,291 Workflows to mature further. 835 01:25:49,208 --> 01:25:52,583 I think there's more and more use cases, 836 01:25:52,583 --> 01:25:53,500 particularly when it 837 01:25:53,500 --> 01:25:55,250 comes to AI applications 838 01:25:55,666 --> 01:25:56,916 where you want to express 839 01:25:56,916 --> 01:25:59,250 things as long running workflows. 840 01:25:59,916 --> 01:26:02,625 And I think this is where Effect is 841 01:26:02,625 --> 01:26:04,750 really like such a perfect foundation. 842 01:26:05,250 --> 01:26:08,208 So yeah, I think we're going to see some 843 01:26:08,208 --> 01:26:09,416 very interesting updates 844 01:26:09,416 --> 01:26:10,458 throughout the year here. 845 01:26:11,291 --> 01:26:13,083 So stay tuned for that. 846 01:26:13,083 --> 01:26:15,833 Maybe changing gears slightly. 847 01:26:15,833 --> 01:26:17,875 Initially, you've built all of the 848 01:26:17,875 --> 01:26:19,958 system, I think mostly by yourself. 849 01:26:19,958 --> 01:26:21,166 This is where you started 850 01:26:21,166 --> 01:26:22,708 adopting Effect by yourself. 851 01:26:23,333 --> 01:26:25,416 and then later you onboarded other 852 01:26:25,416 --> 01:26:27,250 engineers who didn't have 853 01:26:27,250 --> 01:26:28,791 prior Effect experience. 854 01:26:29,458 --> 01:26:31,416 Can you share a bit more about that 855 01:26:31,416 --> 01:26:33,041 story, how you onboarded 856 01:26:33,041 --> 01:26:34,291 other engineers who didn't 857 01:26:34,291 --> 01:26:36,583 have Effect experience to become 858 01:26:36,583 --> 01:26:38,625 successful working on your code base? 859 01:26:39,208 --> 01:26:39,791 I mean, this is such 860 01:26:39,791 --> 01:26:40,541 an interesting question. 861 01:26:40,625 --> 01:26:44,500 Because you would think that Effect is 862 01:26:44,500 --> 01:26:47,625 mainly for a TypeScript audience. 863 01:26:47,625 --> 01:26:50,750 Web developers who are proficient now in 864 01:26:50,750 --> 01:26:52,458 TypeScript and realize they 865 01:26:52,458 --> 01:26:53,583 need to go to the next level 866 01:26:54,000 --> 01:26:56,958 in terms of building complex systems. 867 01:26:56,958 --> 01:27:00,208 My take here is actually that having a 868 01:27:00,208 --> 01:27:01,916 code base which is already 869 01:27:01,916 --> 01:27:04,750 fully Effect will actually 870 01:27:04,958 --> 01:27:07,958 lower the barrier to entry for people who 871 01:27:07,958 --> 01:27:10,166 have less experience with TypeScript. 872 01:27:10,458 --> 01:27:13,000 So we've seen that with Elliot, our 873 01:27:13,000 --> 01:27:14,000 founding engineer who has 874 01:27:14,000 --> 01:27:15,166 joined a few months ago, 875 01:27:15,666 --> 01:27:17,750 with very little TypeScript experience, 876 01:27:17,750 --> 01:27:20,166 was able to very quickly take 877 01:27:20,166 --> 01:27:21,625 it up and then be productive. 878 01:27:22,375 --> 01:27:23,916 We've run some trials with some 879 01:27:23,916 --> 01:27:26,000 candidates in the last few weeks with, 880 01:27:26,541 --> 01:27:27,875 again, little experience in 881 01:27:27,875 --> 01:27:28,250 TypeScript. 882 01:27:28,250 --> 01:27:30,250 The thing is that we are hiring a lot of 883 01:27:30,250 --> 01:27:32,416 machine learning engineers. 884 01:27:32,541 --> 01:27:35,208 And so they have typically already a lot 885 01:27:35,208 --> 01:27:36,625 of experience with more 886 01:27:36,625 --> 01:27:37,416 of the machine learning 887 01:27:37,833 --> 01:27:39,000 sets of tools like Python 888 01:27:39,000 --> 01:27:41,541 and very little on TypeScript. 889 01:27:41,541 --> 01:27:44,250 But what we have seen is onboarding them 890 01:27:44,250 --> 01:27:45,583 on our code base, they 891 01:27:45,583 --> 01:27:46,541 were actually able to be 892 01:27:46,541 --> 01:27:47,875 very productive, very early on. 893 01:27:48,166 --> 01:27:50,583 And the great thing is that there's not 894 01:27:50,583 --> 01:27:51,916 much room for bad 895 01:27:51,916 --> 01:27:53,333 practice or the kinds of things 896 01:27:53,333 --> 01:27:55,250 that you need to learn when you get 897 01:27:55,250 --> 01:27:56,541 started with a new language. 898 01:27:56,583 --> 01:27:58,458 So what I think is that actually 899 01:27:58,458 --> 01:27:59,500 Effect will 900 01:27:59,500 --> 01:28:02,375 allow them to be proficient and 901 01:28:02,375 --> 01:28:03,416 build very reliable 902 01:28:03,416 --> 01:28:05,083 code within a few weeks. 903 01:28:05,625 --> 01:28:07,541 If it had not been an Effect and more 904 01:28:07,541 --> 01:28:08,625 like vanilla TypeScript, 905 01:28:08,833 --> 01:28:09,958 yes, sure, there's fewer 906 01:28:09,958 --> 01:28:10,708 things to learn. 907 01:28:10,708 --> 01:28:12,125 It's kind of easy and so on. 908 01:28:12,625 --> 01:28:15,041 But you would build up this understanding 909 01:28:15,041 --> 01:28:16,458 and intuition for the 910 01:28:16,458 --> 01:28:17,458 language over the course of 911 01:28:17,458 --> 01:28:17,875 years. 912 01:28:18,083 --> 01:28:18,333 Right. 913 01:28:18,333 --> 01:28:19,791 And then within years, you'll be able to 914 01:28:19,791 --> 01:28:21,791 build very complex and 915 01:28:21,791 --> 01:28:23,083 reliable systems because you 916 01:28:23,125 --> 01:28:26,375 have been proficient in the language. 917 01:28:26,875 --> 01:28:29,833 So Effect gives you these guardrails that 918 01:28:29,833 --> 01:28:31,250 with the whole thing is 919 01:28:31,250 --> 01:28:32,291 in Effect, sure, you need 920 01:28:32,291 --> 01:28:33,333 to understand the 921 01:28:33,333 --> 01:28:35,208 whole catalog of things. 922 01:28:35,208 --> 01:28:36,291 And this you just take up 923 01:28:36,291 --> 01:28:37,541 as you have new use cases. 924 01:28:37,916 --> 01:28:39,500 But if you have a really good foundation, 925 01:28:39,875 --> 01:28:41,125 like the full RPC 926 01:28:41,125 --> 01:28:43,041 server, you basically just 927 01:28:43,166 --> 01:28:45,166 you start writing schema, right? 928 01:28:45,166 --> 01:28:46,583 You write a schema and then 929 01:28:46,583 --> 01:28:48,041 you fit it into the system. 930 01:28:48,541 --> 01:28:50,625 And then once it's there, it's done. 931 01:28:51,125 --> 01:28:52,458 You don't need to change it. 932 01:28:52,500 --> 01:28:54,125 You haven't done anything bad 933 01:28:54,125 --> 01:28:56,000 or anything that will break. 934 01:28:56,250 --> 01:28:56,458 Right. 935 01:28:56,541 --> 01:28:59,041 So my take on this is that actually it 936 01:28:59,041 --> 01:29:00,041 might actually have the 937 01:29:00,041 --> 01:29:01,708 opposite effect of what 938 01:29:01,708 --> 01:29:03,958 we would expect, which is to be able to 939 01:29:03,958 --> 01:29:06,375 onboard non TypeScript 940 01:29:06,375 --> 01:29:08,958 engineers much, much faster. 941 01:29:09,375 --> 01:29:11,000 Yeah, that definitely sounds 942 01:29:11,000 --> 01:29:12,333 a little bit counterintuitive. 943 01:29:12,458 --> 01:29:15,375 And the first time I heard about this 944 01:29:15,375 --> 01:29:16,333 anecdote, and you shared 945 01:29:16,333 --> 01:29:17,791 this privately with me before, 946 01:29:18,416 --> 01:29:21,500 and that person you've just mentioned, 947 01:29:21,500 --> 01:29:23,833 Elliot, gave a brilliant talk at the last 948 01:29:23,916 --> 01:29:26,333 Effect Meetup in San Francisco where he 949 01:29:26,333 --> 01:29:28,333 shared his first hand 950 01:29:28,333 --> 01:29:29,750 experience learning Effect. 951 01:29:30,500 --> 01:29:32,125 And yeah, I think that is a 952 01:29:32,125 --> 01:29:33,166 little bit counterintuitive. 953 01:29:33,291 --> 01:29:35,166 But when you think about it more, it 954 01:29:35,166 --> 01:29:37,791 makes a lot of sense because most 955 01:29:37,791 --> 01:29:39,250 experienced TypeScript 956 01:29:39,291 --> 01:29:42,541 engineers over years, they had to first 957 01:29:42,541 --> 01:29:44,375 learn how to do anything 958 01:29:44,375 --> 01:29:46,458 with TypeScript at all and then 959 01:29:46,500 --> 01:29:49,958 later learn to ignore like those 80 paths 960 01:29:49,958 --> 01:29:52,125 out of 100 how you could do 961 01:29:52,125 --> 01:29:53,166 something with TypeScript, 962 01:29:53,416 --> 01:29:54,500 but that will lead you 963 01:29:54,500 --> 01:29:55,958 down a dangerous road. 964 01:29:56,500 --> 01:29:58,833 Whereas Effect constrains you a little 965 01:29:58,833 --> 01:30:01,208 bit more and puts you on a 966 01:30:01,208 --> 01:30:02,791 path to success in a way. 967 01:30:03,708 --> 01:30:04,875 That's also like it has been my 968 01:30:04,875 --> 01:30:06,666 experience for I think for me 969 01:30:06,666 --> 01:30:08,875 coming more from TypeScript, 970 01:30:09,916 --> 01:30:12,500 I had to unlearn a lot of things how I 971 01:30:12,500 --> 01:30:14,041 used to do things in TypeScript before. 972 01:30:14,708 --> 01:30:17,375 But this is where for me, a good 973 01:30:17,375 --> 01:30:19,416 foundation to pick up Effect 974 01:30:19,416 --> 01:30:21,791 was actually dabbling in other 975 01:30:21,791 --> 01:30:24,541 languages such as Swift and Rust, where I 976 01:30:24,541 --> 01:30:28,750 already primed my mind more on more saner 977 01:30:29,250 --> 01:30:31,708 engineering practices and learning the 978 01:30:31,708 --> 01:30:33,208 right habits and 979 01:30:33,208 --> 01:30:34,875 embracing the right abstractions. 980 01:30:35,583 --> 01:30:38,041 And so I can totally see how someone who 981 01:30:38,041 --> 01:30:39,708 doesn't have prior TypeScript experience 982 01:30:40,166 --> 01:30:42,375 maybe experience in other languages, 983 01:30:42,375 --> 01:30:45,666 which have more appropriate primitives 984 01:30:45,833 --> 01:30:48,083 can be up and running with Effect in no 985 01:30:48,083 --> 01:30:50,125 time. That is amazing to hear. 986 01:30:50,666 --> 01:30:53,125 So can you walk me through a little bit 987 01:30:53,125 --> 01:30:54,583 more through the process 988 01:30:54,583 --> 01:30:56,250 how you went tactically 989 01:30:56,416 --> 01:31:00,291 about migrating the code base as it was 990 01:31:00,291 --> 01:31:02,041 before Effect one by one 991 01:31:02,041 --> 01:31:03,500 step at a time to migrating 992 01:31:03,750 --> 01:31:06,833 it to Effect. Rome wasn't conquered in a 993 01:31:06,833 --> 01:31:08,250 day, so wasn't your code 994 01:31:08,250 --> 01:31:09,750 base. So how did you go about it? 995 01:31:10,041 --> 01:31:12,083 We had some great help. And this is where 996 01:31:12,083 --> 01:31:13,375 I think the Effect community 997 01:31:13,375 --> 01:31:15,458 is such a wonderful community 998 01:31:15,458 --> 01:31:19,875 of people who are nice and helpful. You 999 01:31:19,875 --> 01:31:22,541 kicked it off with us. So we've been 1000 01:31:22,541 --> 01:31:25,541 this for years. And immediately I saw the 1001 01:31:25,541 --> 01:31:27,291 importance of this, but 1002 01:31:27,291 --> 01:31:29,666 never found the time to 1003 01:31:30,250 --> 01:31:32,666 actually learn how to do it in practice. 1004 01:31:32,666 --> 01:31:33,458 I think this is also 1005 01:31:33,458 --> 01:31:34,625 maybe related to the lack of 1006 01:31:34,625 --> 01:31:36,375 documentation and concrete examples back 1007 01:31:36,375 --> 01:31:37,125 in the day. How do you 1008 01:31:37,125 --> 01:31:38,625 actually convert an API 1009 01:31:38,666 --> 01:31:41,750 endpoint to Effect? How do you just do 1010 01:31:41,750 --> 01:31:43,666 it? So I think this little 1011 01:31:43,666 --> 01:31:45,041 catalyzer was really important, 1012 01:31:45,666 --> 01:31:49,625 which allowed us, our team, to then get a 1013 01:31:49,625 --> 01:31:50,708 sense of where we could 1014 01:31:50,708 --> 01:31:52,125 start fitting it in. It was still 1015 01:31:52,125 --> 01:31:53,958 like very, very new. This was earlier 1016 01:31:53,958 --> 01:31:57,375 last year. And then we had the pleasure 1017 01:31:57,375 --> 01:31:59,000 of having Ethan with 1018 01:31:59,000 --> 01:32:01,625 us over the summer. And he wrote a lot of 1019 01:32:01,625 --> 01:32:03,041 very important code for both 1020 01:32:03,041 --> 01:32:04,333 the backend side of things, 1021 01:32:05,166 --> 01:32:08,250 schema, the database, the React query 1022 01:32:08,250 --> 01:32:10,125 part, which put us in a 1023 01:32:10,125 --> 01:32:11,000 really good foundation. 1024 01:32:11,500 --> 01:32:13,875 We had most of it basically sketched out. 1025 01:32:14,500 --> 01:32:17,791 And then in September, we decided to 1026 01:32:18,125 --> 01:32:20,083 do a big, big push and just get 1027 01:32:20,083 --> 01:32:23,166 everything migrated. So this was like a 1028 01:32:23,166 --> 01:32:23,833 bit slow in the beginning. 1029 01:32:24,500 --> 01:32:27,583 And then very, very quickly, within a few 1030 01:32:27,583 --> 01:32:28,583 weeks, we had the 1031 01:32:28,583 --> 01:32:31,375 whole thing migrated over. 1032 01:32:32,125 --> 01:32:34,291 Good. That makes a lot of sense. And 1033 01:32:34,291 --> 01:32:36,875 happy to have acted as 1034 01:32:36,875 --> 01:32:38,291 a little catalyst here, 1035 01:32:38,750 --> 01:32:40,833 but you've picked up everything so 1036 01:32:40,833 --> 01:32:43,791 quickly. So I think this was rather a 1037 01:32:43,791 --> 01:32:45,625 little bit of giving you 1038 01:32:45,875 --> 01:32:48,791 more courage to go this way. And I think 1039 01:32:48,791 --> 01:32:50,625 you figured out quickly 1040 01:32:50,625 --> 01:32:52,916 how to work with everything, 1041 01:32:53,291 --> 01:32:57,416 and you migrated everything in no time. I 1042 01:32:57,416 --> 01:32:58,875 would say, honestly, 1043 01:32:59,500 --> 01:33:00,583 the important thing is the 1044 01:33:00,583 --> 01:33:03,958 catalyzer here. Once you have the starter 1045 01:33:03,958 --> 01:33:05,875 blocks, then it's quite easy to 1046 01:33:05,875 --> 01:33:07,291 generalize. It's more 1047 01:33:07,291 --> 01:33:09,208 like where do you start? And the 1048 01:33:09,208 --> 01:33:10,666 documentation obviously is much better 1049 01:33:10,666 --> 01:33:11,625 now. So I think this is 1050 01:33:11,666 --> 01:33:14,041 a great thing. But I think this little 1051 01:33:14,041 --> 01:33:17,125 initial nudge is 1052 01:33:17,125 --> 01:33:18,458 what made the difference here. 1053 01:33:19,875 --> 01:33:22,250 If you think about your system as sort of 1054 01:33:22,250 --> 01:33:27,083 like a tree of different subsystems, 1055 01:33:27,416 --> 01:33:29,291 you can go about converting an existing 1056 01:33:29,291 --> 01:33:31,458 system to Effect in different ways. 1057 01:33:31,458 --> 01:33:33,875 Either you go first for 1058 01:33:33,875 --> 01:33:36,083 the leaves, and then you go up one step 1059 01:33:36,083 --> 01:33:37,583 at a time until your entire 1060 01:33:37,583 --> 01:33:39,375 program, your entire system is 1061 01:33:39,416 --> 01:33:41,875 refactored with Effect. But you can also 1062 01:33:41,875 --> 01:33:44,375 isolate specific sub nodes 1063 01:33:44,375 --> 01:33:47,791 and from the outside call the 1064 01:33:47,791 --> 01:33:50,208 effect and inside of it call other 1065 01:33:50,208 --> 01:33:52,250 things that are not yet Effect with 1066 01:33:52,250 --> 01:33:54,833 Effect.runPromise or 1067 01:33:54,833 --> 01:33:57,666 effect dot promise to run some effects, 1068 01:33:57,875 --> 01:33:59,708 some non-effect code yet. 1069 01:34:00,166 --> 01:34:02,208 Did you find any particular 1070 01:34:02,375 --> 01:34:06,250 topology adoption strategy for you to be 1071 01:34:06,250 --> 01:34:08,916 working well? So did you go first 1072 01:34:08,916 --> 01:34:11,500 all for the leaves? Or did you go 1073 01:34:11,500 --> 01:34:12,666 wherever you wanted to 1074 01:34:12,666 --> 01:34:14,958 refactor something anyway? How did you 1075 01:34:14,958 --> 01:34:17,791 go about the decision process of when to 1076 01:34:17,791 --> 01:34:19,208 refactor what with Effect? 1077 01:34:19,791 --> 01:34:20,666 Well, yeah, it was the leaves in 1078 01:34:20,708 --> 01:34:24,250 the beginning. And then it was the top 1079 01:34:24,250 --> 01:34:26,208 down afterwards, because 1080 01:34:26,208 --> 01:34:27,291 we just realized that this 1081 01:34:27,291 --> 01:34:29,750 was what we needed. And we just needed as 1082 01:34:29,750 --> 01:34:31,750 quickly as possible. So 1083 01:34:31,750 --> 01:34:33,500 then there was a combined push 1084 01:34:34,000 --> 01:34:36,458 from the whole team to just get the 1085 01:34:36,458 --> 01:34:37,791 migration done as quickly as 1086 01:34:37,791 --> 01:34:40,500 possible. So the RPC was written 1087 01:34:40,500 --> 01:34:43,166 really like top down, right? Because, 1088 01:34:43,166 --> 01:34:44,750 well top down in the 1089 01:34:44,750 --> 01:34:46,666 sense that one route would be 1090 01:34:47,000 --> 01:34:49,833 migrated to RPC, Effect RPC, then another 1091 01:34:49,833 --> 01:34:52,333 one and so on. But for 1092 01:34:52,333 --> 01:34:53,625 sure, this is a nice thing with 1093 01:34:53,625 --> 01:34:55,125 Effect is that you can do it on the 1094 01:34:55,125 --> 01:34:57,208 leave to start with and you're not 1095 01:34:57,208 --> 01:34:58,083 touching anything else 1096 01:34:58,333 --> 01:35:00,208 of your code base. But yeah, you 1097 01:35:00,208 --> 01:35:01,666 definitely feel the urge to 1098 01:35:01,666 --> 01:35:05,250 get the whole thing end to end. 1099 01:35:05,625 --> 01:35:08,250 The runtime should be initiated at the 1100 01:35:08,250 --> 01:35:10,416 very beginning of the 1101 01:35:10,416 --> 01:35:11,625 thing, right? So that you have 1102 01:35:11,666 --> 01:35:14,708 the full happy path, because you get to 1103 01:35:14,708 --> 01:35:16,541 love the happy path here, right? 1104 01:35:16,541 --> 01:35:18,625 Nice. So if I remember 1105 01:35:18,625 --> 01:35:20,833 correctly, I think it's now been pretty 1106 01:35:20,833 --> 01:35:22,750 much a year since you've been using 1107 01:35:22,750 --> 01:35:24,000 Effect in production. 1108 01:35:24,666 --> 01:35:26,125 looking back, what were some 1109 01:35:26,125 --> 01:35:27,666 of the things that were 1110 01:35:27,666 --> 01:35:30,000 like unexpectedly good with 1111 01:35:30,000 --> 01:35:32,041 Effect and also some things that were 1112 01:35:32,041 --> 01:35:33,791 more challenging that you thought or 1113 01:35:33,791 --> 01:35:35,333 unexpected in some ways? 1114 01:35:36,583 --> 01:35:39,083 I didn't expect the Schema package 1115 01:35:39,083 --> 01:35:42,000 to be, you know, I didn't have any 1116 01:35:42,000 --> 01:35:43,375 expectations that this 1117 01:35:43,375 --> 01:35:45,875 was really a core thing. In fact, because 1118 01:35:45,875 --> 01:35:49,375 this is really wonderful, how you can 1119 01:35:49,375 --> 01:35:50,708 control both like the 1120 01:35:50,875 --> 01:35:53,500 decoding part, the encoding part, how you 1121 01:35:53,500 --> 01:35:54,875 can pipe everything 1122 01:35:54,875 --> 01:35:59,333 together. This is one of the sort of, 1123 01:36:00,125 --> 01:36:01,916 it shifts away how you think about a 1124 01:36:01,916 --> 01:36:02,875 problem. Because usually, 1125 01:36:03,375 --> 01:36:05,375 now when I think of a problem, 1126 01:36:05,375 --> 01:36:06,458 I start thinking about what is the 1127 01:36:06,458 --> 01:36:07,750 schema, right? It is very 1128 01:36:07,750 --> 01:36:09,833 different from what is the actual 1129 01:36:09,875 --> 01:36:11,708 application code. You redesign your 1130 01:36:11,708 --> 01:36:13,916 schema, you think, well, how should it 1131 01:36:13,916 --> 01:36:16,000 transform the data here? 1132 01:36:16,416 --> 01:36:18,791 And then you work your way up to getting 1133 01:36:18,791 --> 01:36:21,583 it all the way in your 1134 01:36:21,583 --> 01:36:23,250 code. So I think that schema was 1135 01:36:23,250 --> 01:36:24,833 a very, very positive, 1136 01:36:24,833 --> 01:36:26,791 unexpected thing that came out of this. 1137 01:36:27,708 --> 01:36:28,291 To give you an example, 1138 01:36:29,083 --> 01:36:32,875 I always wanted to, again, in the quest 1139 01:36:32,875 --> 01:36:34,833 for extreme lining and being 1140 01:36:34,833 --> 01:36:36,291 as efficient as possible as a 1141 01:36:36,291 --> 01:36:39,375 startup, things like documentation, I 1142 01:36:39,375 --> 01:36:41,416 care a lot about documentation. And the 1143 01:36:41,416 --> 01:36:42,250 more we can put that 1144 01:36:42,250 --> 01:36:44,875 on like autopilot, the better, right? So 1145 01:36:44,875 --> 01:36:46,041 the ability to take 1146 01:36:46,041 --> 01:36:47,250 your public API endpoints, 1147 01:36:47,958 --> 01:36:49,833 and the other schemas generate an open 1148 01:36:49,833 --> 01:36:52,458 API spec is exactly the kind of things 1149 01:36:52,458 --> 01:36:54,708 that you want, right? 1150 01:36:54,708 --> 01:36:57,250 So this was a very, very nice discovery. 1151 01:36:57,791 --> 01:36:59,291 Right. That makes a lot of 1152 01:36:59,291 --> 01:37:00,583 sense. And I think that goes 1153 01:37:00,666 --> 01:37:04,000 hand in hand with what I've also found in 1154 01:37:04,000 --> 01:37:06,375 my development practices over the years, 1155 01:37:06,833 --> 01:37:10,083 is that I always started to think more 1156 01:37:10,083 --> 01:37:13,000 about the types of a system. 1157 01:37:13,000 --> 01:37:13,875 That's like typically what I 1158 01:37:13,916 --> 01:37:16,541 start out with. I just like described the 1159 01:37:16,541 --> 01:37:18,416 types. But then I've also 1160 01:37:18,416 --> 01:37:20,000 found myself thinking about, 1161 01:37:20,208 --> 01:37:22,708 actually, this data of this particular 1162 01:37:22,708 --> 01:37:25,291 type doesn't just live inside of this 1163 01:37:25,291 --> 01:37:26,291 boundary of the system, 1164 01:37:26,625 --> 01:37:29,416 but it needs to move across network 1165 01:37:29,416 --> 01:37:31,416 boundaries, across threat boundaries, 1166 01:37:31,416 --> 01:37:33,333 etc. And now you're, 1167 01:37:33,333 --> 01:37:35,541 you need to serialize, deserialize that 1168 01:37:35,541 --> 01:37:38,250 variable. And this is where 1169 01:37:38,250 --> 01:37:40,916 schemas come in. So I ended up 1170 01:37:40,916 --> 01:37:43,333 doing exactly the same as you. Like I 1171 01:37:43,333 --> 01:37:44,541 start thinking about like, 1172 01:37:44,541 --> 01:37:46,291 okay, what data am I dealing 1173 01:37:46,291 --> 01:37:49,125 with here? And I modeled it as schemas 1174 01:37:49,125 --> 01:37:50,541 and derived the types from it. 1175 01:37:51,500 --> 01:37:53,625 And well, we see this here as 1176 01:37:53,625 --> 01:37:56,416 like a big revelation turns out in other 1177 01:37:56,416 --> 01:37:58,250 programming languages, that's what you 1178 01:37:58,250 --> 01:37:59,833 always do. When you have 1179 01:37:59,833 --> 01:38:02,250 like in Rust, etc. When you have like a 1180 01:38:02,250 --> 01:38:03,500 struct definition, you can 1181 01:38:03,500 --> 01:38:06,250 derive schemas from it. But 1182 01:38:06,250 --> 01:38:08,416 I think for TypeScript, this is really 1183 01:38:08,416 --> 01:38:10,166 for most engineers, that's 1184 01:38:10,166 --> 01:38:12,458 something novel. And I think 1185 01:38:12,458 --> 01:38:14,708 this is a pattern that should be embraced 1186 01:38:14,708 --> 01:38:16,291 much more since I think 1187 01:38:16,291 --> 01:38:18,625 default is you just ignore 1188 01:38:18,875 --> 01:38:21,208 the existence of a schema and you JSON 1189 01:38:21,208 --> 01:38:23,416 stringify everything. And then you're 1190 01:38:23,416 --> 01:38:25,041 surprised what happens 1191 01:38:25,041 --> 01:38:27,791 to your classes or to your dates. So 1192 01:38:27,791 --> 01:38:29,916 yeah, that schema is such 1193 01:38:29,916 --> 01:38:31,708 an integral part of Effect, 1194 01:38:32,208 --> 01:38:34,916 I think is no coincidence. And I'm very 1195 01:38:34,916 --> 01:38:36,708 happy that it gets so 1196 01:38:36,708 --> 01:38:37,916 much attention for detail. 1197 01:38:38,666 --> 01:38:41,125 Yeah, I mean, I would say it's not only 1198 01:38:41,125 --> 01:38:42,125 the fact that you have like 1199 01:38:42,125 --> 01:38:44,666 decoding and encoding and so 1200 01:38:44,666 --> 01:38:48,083 on, it's the fact that it's so much in 1201 01:38:48,083 --> 01:38:49,750 harmony with the rest of 1202 01:38:49,750 --> 01:38:51,333 the whole like Effect way of 1203 01:38:51,333 --> 01:38:55,958 doing things, right? So it just feels so, 1204 01:38:55,958 --> 01:38:57,916 you know, an integral 1205 01:38:57,916 --> 01:38:59,750 part of effect itself. 1206 01:38:59,958 --> 01:39:04,166 Yeah, I fully agree. So in regards to the 1207 01:39:04,166 --> 01:39:05,666 more challenging aspects, 1208 01:39:06,166 --> 01:39:08,166 what were some of the challenges that you 1209 01:39:08,166 --> 01:39:10,416 found working in an Effect code base? 1210 01:39:11,333 --> 01:39:14,625 Sure. I would say it can give you a sense 1211 01:39:14,625 --> 01:39:17,166 of, you know 1212 01:39:17,166 --> 01:39:17,916 maybe a 1213 01:39:17,916 --> 01:39:18,958 false sense of security. 1214 01:39:19,333 --> 01:39:20,833 Because you think that everything 1215 01:39:20,833 --> 01:39:21,916 is dealt with at the 1216 01:39:21,916 --> 01:39:23,458 type level and you get no 1217 01:39:23,458 --> 01:39:26,208 flagged of compile time errors and so on. 1218 01:39:26,833 --> 01:39:28,208 And this is partly 1219 01:39:28,208 --> 01:39:30,416 true. But I think it's very, 1220 01:39:30,416 --> 01:39:33,041 very important. It's very important to 1221 01:39:33,041 --> 01:39:34,375 understand that this is not 1222 01:39:34,375 --> 01:39:36,666 the case. And it gives you 1223 01:39:36,833 --> 01:39:39,625 really good like foundation, but you're 1224 01:39:39,625 --> 01:39:44,291 not exempt of or die or catch all that 1225 01:39:44,291 --> 01:39:46,833 was put at the end of 1226 01:39:46,833 --> 01:39:48,958 your pipes. And so all the other errors 1227 01:39:48,958 --> 01:39:52,083 in between are, you know, just you don't 1228 01:39:52,083 --> 01:39:52,875 see them. And we had 1229 01:39:52,875 --> 01:39:55,625 an example where we had this 1230 01:39:55,625 --> 01:39:56,750 bug in our code base, and it 1231 01:39:56,750 --> 01:39:58,708 was quite hard to figure out 1232 01:39:59,291 --> 01:40:01,416 where it went wrong. So I think this is 1233 01:40:01,416 --> 01:40:02,500 something that, you know, 1234 01:40:02,666 --> 01:40:03,875 Effect is a very, very powerful 1235 01:40:03,875 --> 01:40:08,166 tool. And it brings out the best in us, 1236 01:40:08,166 --> 01:40:09,833 you know, engineers, I think. 1237 01:40:10,250 --> 01:40:12,041 But it also needs to be 1238 01:40:12,041 --> 01:40:15,708 dealt with knowing that, you know, 1239 01:40:15,708 --> 01:40:17,708 it's not like a safety guard 1240 01:40:17,708 --> 01:40:19,208 or anything, you can mess up 1241 01:40:19,250 --> 01:40:21,916 and you can do things that are very hard 1242 01:40:21,916 --> 01:40:23,666 to then investigate because 1243 01:40:23,666 --> 01:40:24,666 you've built up this, this 1244 01:40:24,666 --> 01:40:26,125 sense of everything is just working 1245 01:40:26,125 --> 01:40:28,458 smoothly. So this is one thing I think 1246 01:40:28,458 --> 01:40:30,416 that this is as important 1247 01:40:30,416 --> 01:40:32,250 to keep in mind when you're building, you 1248 01:40:32,250 --> 01:40:34,041 know, on top of Effect. I 1249 01:40:34,041 --> 01:40:35,375 think another thing that it gets, 1250 01:40:35,708 --> 01:40:38,166 a little bit hard to track is, 1251 01:40:38,166 --> 01:40:39,333 dependency 1252 01:40:39,333 --> 01:40:41,083 injection is just wonderful, 1253 01:40:42,166 --> 01:40:43,958 in Effect. And I think that's, that's, I 1254 01:40:43,958 --> 01:40:44,708 would mention this probably 1255 01:40:44,708 --> 01:40:46,166 as the second most important, 1256 01:40:46,625 --> 01:40:48,500 discovery or delight of 1257 01:40:48,500 --> 01:40:50,750 working with Effect. But, 1258 01:40:50,750 --> 01:40:53,291 you know, where do you put your 1259 01:40:53,291 --> 01:40:55,083 dependency? Where is it in your code? How 1260 01:40:55,083 --> 01:40:56,458 do you like chase the 1261 01:40:56,458 --> 01:40:58,166 dependency, you know, the service 1262 01:40:58,333 --> 01:41:01,333 that you provided, that is now affecting 1263 01:41:01,333 --> 01:41:02,666 a completely different 1264 01:41:02,666 --> 01:41:03,666 part of your code base, 1265 01:41:03,958 --> 01:41:05,208 that might be, you 1266 01:41:05,208 --> 01:41:06,166 know, some some tooling 1267 01:41:06,166 --> 01:41:07,291 to build inside of the 1268 01:41:07,791 --> 01:41:10,458 IDE itself, to be able to make it easy to 1269 01:41:10,458 --> 01:41:11,875 spot the connection, right. 1270 01:41:12,541 --> 01:41:13,958 But now we still like in this 1271 01:41:13,958 --> 01:41:15,833 very barebone situation where it's kind 1272 01:41:15,833 --> 01:41:17,041 of hard to understand 1273 01:41:17,041 --> 01:41:20,166 exactly, you know, the hierarchy of 1274 01:41:20,166 --> 01:41:23,125 your dependencies, right, and where they 1275 01:41:23,125 --> 01:41:24,375 are actually being used 1276 01:41:24,375 --> 01:41:26,208 in your code. So it can be 1277 01:41:26,208 --> 01:41:27,708 a little bit hard to navigate sometimes. 1278 01:41:28,791 --> 01:41:29,833 Right. Yeah, I found that 1279 01:41:29,833 --> 01:41:31,541 myself as well in the project 1280 01:41:31,541 --> 01:41:34,125 I'm working on. And it is actually 1281 01:41:34,125 --> 01:41:35,125 something that we're thinking 1282 01:41:35,125 --> 01:41:37,416 about, that we could possibly 1283 01:41:37,791 --> 01:41:41,958 enhance through a IDE integration, etc. 1284 01:41:41,958 --> 01:41:44,041 So stay, stay tuned for 1285 01:41:44,041 --> 01:41:46,125 for that. You said something 1286 01:41:46,666 --> 01:41:48,833 interesting just before, where you said 1287 01:41:48,833 --> 01:41:50,583 effect brings out the 1288 01:41:50,583 --> 01:41:53,208 best in us engineers. Can you 1289 01:41:53,208 --> 01:41:55,500 elaborate on that? What do you mean by that 1290 01:41:55,500 --> 01:41:58,125 Well, as an engineer, 1291 01:41:58,125 --> 01:41:58,958 you want to build features, 1292 01:41:59,125 --> 01:42:00,583 you want to build like real robust 1293 01:42:00,583 --> 01:42:03,041 systems, you want to spend 1294 01:42:03,041 --> 01:42:05,375 ideally all your time doing 1295 01:42:05,375 --> 01:42:08,083 that. You don't want to spend your time 1296 01:42:08,083 --> 01:42:10,125 fighting the language, fighting, you 1297 01:42:10,125 --> 01:42:11,375 know, building something 1298 01:42:11,458 --> 01:42:13,458 that you know, the language was not meant 1299 01:42:13,458 --> 01:42:14,208 for, to give you an 1300 01:42:14,208 --> 01:42:15,333 example, like concurrency and 1301 01:42:15,333 --> 01:42:17,000 TypeScript, right, these are things that 1302 01:42:17,000 --> 01:42:21,125 you know, you want a way to, 1303 01:42:21,125 --> 01:42:22,916 you know, build a concurrent 1304 01:42:22,916 --> 01:42:25,416 system, where you're thinking about what 1305 01:42:25,416 --> 01:42:26,458 is the system actually doing 1306 01:42:26,458 --> 01:42:28,000 rather than like setting up 1307 01:42:28,000 --> 01:42:30,958 all the boilerplate in order to have that 1308 01:42:30,958 --> 01:42:34,458 system in place. So I would 1309 01:42:34,458 --> 01:42:36,500 say Effect really gives you 1310 01:42:36,916 --> 01:42:40,041 the right premise. So one thing I'm 1311 01:42:40,041 --> 01:42:42,000 thinking about is, now that we've 1312 01:42:42,000 --> 01:42:43,125 completed this migration, 1313 01:42:44,666 --> 01:42:46,375 the full migration, it kind of feels like 1314 01:42:46,375 --> 01:42:48,500 this is the end game. I know 1315 01:42:48,500 --> 01:42:49,708 this might sound a bit like 1316 01:42:50,250 --> 01:42:54,125 naive or unrealistic, but it kind of 1317 01:42:54,125 --> 01:42:56,083 feels like more stable. 1318 01:42:56,333 --> 01:42:58,125 And at least if we continue 1319 01:42:58,541 --> 01:43:00,166 building our code base on TypeScript with 1320 01:43:00,166 --> 01:43:02,208 Effect, it doesn't feel 1321 01:43:02,208 --> 01:43:03,166 like we're going to spend 1322 01:43:04,250 --> 01:43:06,666 many more like cycles, like refactoring 1323 01:43:06,666 --> 01:43:09,000 the code, right, it feels 1324 01:43:09,000 --> 01:43:11,166 like this sort of timeless, 1325 01:43:12,166 --> 01:43:14,833 like code base now now now it's, it's, 1326 01:43:14,833 --> 01:43:16,125 it's there. Effect is 1327 01:43:16,125 --> 01:43:17,791 going to evolve, our code base is 1328 01:43:17,791 --> 01:43:19,416 going to evolve. It feels like this is 1329 01:43:19,416 --> 01:43:20,583 going to be like adjustments 1330 01:43:20,583 --> 01:43:22,000 and not like a full rewrite. 1331 01:43:22,250 --> 01:43:24,208 And this is what I mean by, you know, it 1332 01:43:24,208 --> 01:43:25,583 feels like kind of like 1333 01:43:25,583 --> 01:43:26,958 building a house with some really 1334 01:43:27,000 --> 01:43:29,583 good like walls and floor and ceiling and 1335 01:43:29,583 --> 01:43:31,416 so on. It's just there now. 1336 01:43:32,041 --> 01:43:34,083 And this is what we are all 1337 01:43:34,083 --> 01:43:36,500 I guess, striving for as engineers is to 1338 01:43:36,500 --> 01:43:38,166 have these, you know, the 1339 01:43:38,166 --> 01:43:39,958 right tools, right, the right 1340 01:43:39,958 --> 01:43:42,375 foundations, instead of having to fight, 1341 01:43:42,375 --> 01:43:44,416 you know, all these like small things 1342 01:43:44,416 --> 01:43:46,000 that degrade over time, 1343 01:43:46,083 --> 01:43:48,958 and then need to revisit them. And you 1344 01:43:48,958 --> 01:43:51,000 realize how silly you were 1345 01:43:51,000 --> 01:43:52,125 writing this code, you know, 1346 01:43:52,166 --> 01:43:54,708 last year kind of thing, it feels like 1347 01:43:54,708 --> 01:43:57,833 more harmonious and stable 1348 01:43:57,833 --> 01:43:59,708 Right. I love that. And 1349 01:43:59,708 --> 01:44:02,208 I think that is also something that has 1350 01:44:02,208 --> 01:44:04,041 really brought me to Effect. 1351 01:44:04,041 --> 01:44:06,625 And to flip it on its head, 1352 01:44:06,750 --> 01:44:10,250 something that I experienced so many 1353 01:44:10,250 --> 01:44:13,083 times and have observed so many other 1354 01:44:13,083 --> 01:44:15,083 teams as sort of like a 1355 01:44:15,541 --> 01:44:18,333 malfunction of engineering teams and of 1356 01:44:18,333 --> 01:44:20,541 projects that engineers are 1357 01:44:20,541 --> 01:44:22,916 afraid to make changes in a 1358 01:44:22,916 --> 01:44:25,791 project. And I've seen this time and time 1359 01:44:25,791 --> 01:44:27,333 and time again. And when 1360 01:44:27,333 --> 01:44:29,041 you're afraid to make changes to a 1361 01:44:29,041 --> 01:44:31,666 project, particularly if the code was 1362 01:44:31,666 --> 01:44:32,916 written by someone else, maybe 1363 01:44:32,916 --> 01:44:34,500 the person is no longer there. 1364 01:44:35,083 --> 01:44:36,958 That's terrible. Like if you're, if 1365 01:44:36,958 --> 01:44:38,250 you're afraid to change something, you 1366 01:44:38,250 --> 01:44:39,333 might not change it and 1367 01:44:39,333 --> 01:44:41,458 you might not improve it. Or if you 1368 01:44:41,458 --> 01:44:42,375 change it, you don't know 1369 01:44:42,375 --> 01:44:44,375 about the consequences, etc. 1370 01:44:45,208 --> 01:44:48,416 And effect allows you to, gives you the 1371 01:44:48,416 --> 01:44:51,250 confidence to change something and you 1372 01:44:51,250 --> 01:44:52,041 know what's going to 1373 01:44:52,041 --> 01:44:54,625 happen. It even makes it fun to do so. 1374 01:44:54,625 --> 01:44:55,375 And I think that's 1375 01:44:55,375 --> 01:44:58,750 sort of like a twin to the 1376 01:45:00,333 --> 01:45:03,416 aspect of composability, almost like the 1377 01:45:03,416 --> 01:45:04,791 other side of the same 1378 01:45:04,791 --> 01:45:07,958 coin, where if you can compose 1379 01:45:07,958 --> 01:45:10,125 things together, you can decompose things 1380 01:45:10,125 --> 01:45:11,833 that makes, that's, those 1381 01:45:11,833 --> 01:45:13,750 are just aspects of changing 1382 01:45:13,916 --> 01:45:17,166 something and changing not in a big bang 1383 01:45:17,166 --> 01:45:19,708 way, but just moving things 1384 01:45:19,708 --> 01:45:21,416 around almost feels like a, 1385 01:45:21,416 --> 01:45:23,458 like a dance. We're getting a little bit 1386 01:45:23,458 --> 01:45:26,458 too, too like pie in the sky 1387 01:45:26,458 --> 01:45:27,666 here, but that's really like 1388 01:45:27,666 --> 01:45:29,791 my reality as well, like working with 1389 01:45:29,791 --> 01:45:31,083 Effect and one of the things 1390 01:45:31,083 --> 01:45:33,000 that I like so much about it. 1391 01:45:33,625 --> 01:45:35,708 And another thing here may be looking a 1392 01:45:35,708 --> 01:45:36,625 little bit more into the 1393 01:45:36,625 --> 01:45:37,666 future, but we're increasingly 1394 01:45:37,875 --> 01:45:40,875 living in the future where we get, where 1395 01:45:40,875 --> 01:45:42,541 we do a lot of our code 1396 01:45:42,541 --> 01:45:44,625 development in an AI-assisted way. 1397 01:45:45,375 --> 01:45:48,166 Now it's maybe no longer just us, us as 1398 01:45:48,166 --> 01:45:50,250 engineers doing performing 1399 01:45:50,250 --> 01:45:52,625 changes on our code, but it's now 1400 01:45:52,625 --> 01:45:54,750 AI systems performing changes on our 1401 01:45:54,750 --> 01:45:56,541 code. And this is where the same 1402 01:45:56,541 --> 01:45:58,791 underlying dynamic is even more 1403 01:45:58,833 --> 01:46:02,958 severe. If there is a system, if there is 1404 01:46:02,958 --> 01:46:04,916 some actor here who's 1405 01:46:04,916 --> 01:46:07,333 changing our code, who is maybe 1406 01:46:07,375 --> 01:46:10,083 even less competent or less knowledgeable 1407 01:46:10,083 --> 01:46:12,333 about this, then maybe even 1408 01:46:12,333 --> 01:46:14,375 more scary to make changes 1409 01:46:14,375 --> 01:46:18,000 to a system. And if the material here, 1410 01:46:18,000 --> 01:46:18,833 the house that we're 1411 01:46:18,833 --> 01:46:19,750 building that you've mentioned, 1412 01:46:20,500 --> 01:46:24,708 is where it's inviting to have changes be 1413 01:46:24,708 --> 01:46:27,708 made, that sets an even better foundation 1414 01:46:27,708 --> 01:46:29,833 for those AI-assisted, 1415 01:46:30,625 --> 01:46:33,208 for this AI-assisted future. So that's 1416 01:46:33,208 --> 01:46:34,833 something I love thinking about. 1417 01:46:35,625 --> 01:46:41,041 Yeah, I agree. I think the guardrails 1418 01:46:41,041 --> 01:46:42,416 here are so important, 1419 01:46:42,416 --> 01:46:43,791 right? You want to make sure that, 1420 01:46:44,791 --> 01:46:49,541 whatever the LLM generates has as little 1421 01:46:49,541 --> 01:46:51,875 unintended side effects as possible, 1422 01:46:52,333 --> 01:46:55,083 because it could hallucinate some code. 1423 01:46:55,416 --> 01:46:57,541 So if you can make sure that 1424 01:46:57,541 --> 01:46:59,958 the blast radius is as tight, 1425 01:46:59,958 --> 01:47:01,875 small as possible, then I think you have 1426 01:47:01,875 --> 01:47:03,541 a really good foundation. Obviously, I 1427 01:47:03,541 --> 01:47:04,125 don't know where this 1428 01:47:04,125 --> 01:47:05,583 is heading. Maybe we're going to be 1429 01:47:05,583 --> 01:47:07,375 completely wrong and that we're not even 1430 01:47:07,375 --> 01:47:07,916 going to be thinking 1431 01:47:07,916 --> 01:47:12,166 about code in the same way in the future. 1432 01:47:12,166 --> 01:47:13,208 But it feels like it's a 1433 01:47:13,208 --> 01:47:16,583 really good way of ensuring 1434 01:47:16,708 --> 01:47:20,000 that whatever an LLM generates stays 1435 01:47:20,000 --> 01:47:24,000 within the intended radius. 1436 01:47:25,916 --> 01:47:28,208 Exactly. And I think just to round this 1437 01:47:28,208 --> 01:47:30,750 part out, Effect gives 1438 01:47:30,750 --> 01:47:33,958 you a middle ground between 1439 01:47:34,416 --> 01:47:37,916 a natural language like English that 1440 01:47:37,916 --> 01:47:40,125 might be more intuitive to 1441 01:47:40,125 --> 01:47:41,958 describe what a system should do, 1442 01:47:41,958 --> 01:47:45,791 but maybe less concise, as opposed to a 1443 01:47:45,791 --> 01:47:47,208 type system or code, 1444 01:47:47,666 --> 01:47:49,083 which can be more concise, 1445 01:47:49,166 --> 01:47:51,208 but maybe less intuitive. And I feel 1446 01:47:51,208 --> 01:47:52,625 Effect is the best of 1447 01:47:52,625 --> 01:47:54,583 both worlds, where you can, 1448 01:47:54,583 --> 01:47:56,791 through the help of a type system, and 1449 01:47:56,791 --> 01:47:57,666 here in this case, 1450 01:47:58,083 --> 01:48:00,083 TypeScript provides a marvelous type 1451 01:48:00,166 --> 01:48:02,750 system that allows you to give you the 1452 01:48:02,750 --> 01:48:04,541 correctness benefits that you wouldn't 1453 01:48:04,541 --> 01:48:06,416 get from English, yet 1454 01:48:06,916 --> 01:48:09,791 it gives you that more intuitive way to 1455 01:48:09,791 --> 01:48:13,125 express and compose things. I feel like 1456 01:48:13,125 --> 01:48:14,458 Effect is converging 1457 01:48:14,458 --> 01:48:16,333 towards something that's the best of both 1458 01:48:16,333 --> 01:48:18,250 worlds here, that is 1459 01:48:18,250 --> 01:48:19,833 actually a wonderful substrate 1460 01:48:20,250 --> 01:48:22,791 for AIs to also compose 1461 01:48:22,791 --> 01:48:24,125 systems out of. 1462 01:48:24,125 --> 01:48:26,083 Wish that the latest versions of LLMs 1463 01:48:26,083 --> 01:48:27,416 have even better 1464 01:48:27,416 --> 01:48:28,958 understanding of the latest developments 1465 01:48:28,958 --> 01:48:31,333 in Effect. But I think you mentioned that 1466 01:48:31,333 --> 01:48:32,875 you have a nice hack, which is to 1467 01:48:33,500 --> 01:48:36,625 clone the Effect repository in cursor and 1468 01:48:36,625 --> 01:48:38,375 then have it look at 1469 01:48:38,375 --> 01:48:39,416 that while you're coding. 1470 01:48:40,291 --> 01:48:42,750 Yeah, so that's actually a shout out to 1471 01:48:42,750 --> 01:48:44,416 our friend Tim 1472 01:48:44,416 --> 01:48:47,500 Suchanek. He taught me about 1473 01:48:48,250 --> 01:48:50,333 this little trick, and I've since applied 1474 01:48:50,333 --> 01:48:52,250 it many times also for other projects. 1475 01:48:52,750 --> 01:48:54,875 I think we're currently in this uncanny 1476 01:48:54,875 --> 01:48:57,666 valley, where LLMs always 1477 01:48:57,666 --> 01:48:59,875 lag behind on the latest and 1478 01:48:59,875 --> 01:49:02,833 greatest of actual information. The 1479 01:49:02,833 --> 01:49:04,541 Effect docs, for example, in 1480 01:49:04,541 --> 01:49:06,458 particular, they have really 1481 01:49:06,833 --> 01:49:09,291 advanced a lot, got a lot better over the 1482 01:49:09,291 --> 01:49:10,958 last couple of months. And 1483 01:49:10,958 --> 01:49:13,000 those LLMs that we're using 1484 01:49:13,000 --> 01:49:15,833 on a daily basis have not fully caught up 1485 01:49:15,833 --> 01:49:17,833 to that level yet, and 1486 01:49:17,833 --> 01:49:19,208 are also sometimes trained 1487 01:49:19,208 --> 01:49:21,791 on old code snippets, different syntax, 1488 01:49:22,041 --> 01:49:24,083 etc. And so when you ask 1489 01:49:24,083 --> 01:49:25,875 them, you might get an outdated 1490 01:49:25,875 --> 01:49:29,500 response. So what Tim is doing, and since 1491 01:49:29,500 --> 01:49:31,208 he's taught me about 1492 01:49:31,208 --> 01:49:32,375 this as well, is basically 1493 01:49:33,125 --> 01:49:36,500 check out the Effect code base on your 1494 01:49:36,500 --> 01:49:38,166 computer locally, open 1495 01:49:38,166 --> 01:49:39,541 Cursor in it, and then 1496 01:49:40,041 --> 01:49:43,291 Cursor uses the full Effect repo with 1497 01:49:43,291 --> 01:49:44,541 all the examples, tests, 1498 01:49:44,833 --> 01:49:46,833 etc. as the context. And then 1499 01:49:46,875 --> 01:49:49,625 inside of that, you use the Cursor chat 1500 01:49:49,625 --> 01:49:51,250 to ask your question, and 1501 01:49:51,250 --> 01:49:52,625 you're going to get a really 1502 01:49:52,625 --> 01:49:55,458 useful result. So I've been using this a 1503 01:49:55,458 --> 01:49:57,208 bunch myself, but we're 1504 01:49:57,208 --> 01:49:58,916 also working on some exciting 1505 01:49:59,166 --> 01:50:02,750 improvements to the Effect docs, etc., 1506 01:50:02,750 --> 01:50:04,791 where we're trying to 1507 01:50:04,791 --> 01:50:07,166 combine the latest and greatest 1508 01:50:07,375 --> 01:50:11,500 documentation with some of our own AI 1509 01:50:11,500 --> 01:50:13,750 developments as well. So stay tuned for 1510 01:50:13,750 --> 01:50:14,958 that. You won't need 1511 01:50:14,958 --> 01:50:17,791 those little tricks for too much longer. 1512 01:50:18,291 --> 01:50:19,041 So Michael, 1513 01:50:19,333 --> 01:50:21,041 thank you so much for taking 1514 01:50:21,125 --> 01:50:24,291 out so much of your busy day to share 1515 01:50:24,291 --> 01:50:26,500 about your experience and journey with 1516 01:50:26,500 --> 01:50:28,625 Effect. Really greatly 1517 01:50:28,833 --> 01:50:31,250 appreciate that you share all of that 1518 01:50:31,250 --> 01:50:32,666 with the entire audience 1519 01:50:32,666 --> 01:50:35,583 here. Before rounding this out, 1520 01:50:35,583 --> 01:50:37,500 is there anything else you would like to 1521 01:50:37,500 --> 01:50:38,958 share with the folks 1522 01:50:38,958 --> 01:50:42,666 listening? I'm sure that you're 1523 01:50:42,666 --> 01:50:44,875 very busy building a lot of things, but 1524 01:50:44,875 --> 01:50:46,958 you're also interested in 1525 01:50:46,958 --> 01:50:48,583 bringing on more engineers. 1526 01:50:49,541 --> 01:50:51,041 Anyone in particular who 1527 01:50:51,041 --> 01:50:52,541 should reach out to you? 1528 01:50:53,250 --> 01:50:57,208 Yeah, so we are constantly hiring and we 1529 01:50:57,208 --> 01:51:00,666 are obviously, it's a full TypeScript 1530 01:51:01,000 --> 01:51:04,875 Effect code base, but we are also very 1531 01:51:04,875 --> 01:51:05,708 much looking for 1532 01:51:05,708 --> 01:51:07,458 engineers with an ML background. 1533 01:51:07,916 --> 01:51:11,541 And we would warmly welcome them even if 1534 01:51:11,541 --> 01:51:12,250 they don't have 1535 01:51:12,250 --> 01:51:14,833 specific TypeScript experience. 1536 01:51:14,875 --> 01:51:18,333 So yeah, if you're interested in working 1537 01:51:18,333 --> 01:51:20,666 on such a code base, which I 1538 01:51:20,666 --> 01:51:21,833 think is really, really nice 1539 01:51:22,250 --> 01:51:24,708 and working on some really interesting 1540 01:51:24,708 --> 01:51:26,458 LLM related topics, 1541 01:51:27,208 --> 01:51:29,666 then absolutely reach out. 1542 01:51:30,791 --> 01:51:33,416 Yeah, I couldn't give you a warmer 1543 01:51:33,416 --> 01:51:35,333 recommendation to work at 1544 01:51:35,333 --> 01:51:37,958 Markprompt. Michael, Marie, 1545 01:51:38,583 --> 01:51:40,333 Elliot, all the folks working at 1546 01:51:40,333 --> 01:51:43,791 Markprompt are A plus humans. If I 1547 01:51:43,791 --> 01:51:45,500 wouldn't be doing all of the 1548 01:51:45,500 --> 01:51:47,583 things I'm doing, I'd strongly consider 1549 01:51:47,583 --> 01:51:49,166 working with you all as 1550 01:51:49,166 --> 01:51:50,875 well. So whoever gets a chance 1551 01:51:50,875 --> 01:51:53,291 to work with you, I think you're making a 1552 01:51:53,291 --> 01:51:55,375 very smart decision. So 1553 01:51:55,375 --> 01:51:57,916 now moving back my focus to 1554 01:51:57,916 --> 01:52:00,916 you, Michael, thank you so much again for 1555 01:52:00,916 --> 01:52:03,083 taking time, sharing all of your 1556 01:52:03,083 --> 01:52:04,666 experience and all of 1557 01:52:05,041 --> 01:52:06,958 those anecdotes with us. I greatly 1558 01:52:06,958 --> 01:52:08,708 appreciate it. Well, thank 1559 01:52:08,708 --> 01:52:09,750 you, Johannes. Always a pleasure 1560 01:52:09,791 --> 01:52:11,791 to speak with you. Awesome. 1561 01:52:12,250 --> 01:52:13,833 Thank you so much. Thank you.