SalesGPT: 100x your Sales Team with AI

Filip Michalsky joins the "Behind the Bots" podcast to discuss his open-source project, SalesGPT, an AI sales agent that can have natural conversations and handle common sales objections. Filip talks about how SalesGPT works using large language models and can be customized for different sales scenarios. He also discusses Crusty.ai, a startup using SalesGPT technology for scalable conversational sales over the phone. With over 900 stars on GitHub, SalesGPT shows the potential for AI to transform sales and marketing. Tune in to hear Filip's vision for the future of AI sales assistants.

SALESGPT
https://github.com/filip-michalsky/SalesGPT
https://www.crusty.ai/
https://twitter.com/FilipMichalsky
https://www.fry-ai.com/p/salesgpt-ai-sales-agents-warming-cold-calls


FRY-AI.COM
https://www.fry-ai.com/subscribe
https://twitter.com/lazukars
https://twitter.com/thefryai

Creators & Guests

Host

Ryan Lazuka

The lighthearted Artificial intelligence Journalist. Building the easiest to read AI Email Newsletter Daily Twitter Threads about AI

What is Behind The Bots?

Join us as we delve into the fascinating world of Artificial Intelligence (AI) by interviewing the brightest minds and exploring cutting-edge projects. From innovative ideas to groundbreaking individuals, we're here to uncover the latest developments and thought-provoking discussions in the AI space.

thank you let's say if you want like a certain output you wanted to say certain things
0:20
so that you can parse the string coming out that's actually not always as robust
0:25
now there is one way to uh to get around it and that's using open-air functions which actually like the main functions
0:33
like just in general and the actual difference there is that there's a guarantee you will always get a Json
0:39
back so that like even if your result is not what you want it will not break the flow of the code because like you know
0:46
you're gonna do something you're just relying on a you know a thought action
0:51
observation pattern right which is kind of what what the react to the react framework is what what people use is one
0:57
of the early early uh agents Frameworks if you don't get the thought react
1:04
observation framework out of you um so that's actually one of the uh foreign
1:19
it's pretty incredible like how good it's gotten like he has a he has a character like he has characters like
1:24
Elon Musk and other other celebrities um that you can talk with in real time and you can do it via a phone call or a
1:31
chat and the one thing that stood out that has gotten really good other than the voice that doesn't sound robotic is
1:36
he was demoing that the conversation like they would interrupt each other like mid-sentence it wasn't like someone
1:43
would ask the question and the AI bot would answer and then you'd wait for the AFI to answer someone would interject AI
1:49
bot and his technology was able to pick up the conversation like a human would like um you know if you had a question for me
1:56
right now and you want to interject you could do that whereas I don't think that was possible before with AI bot so it seems like that technology would really
2:02
really come along um in that aspect in more of a natural language type of uh that natural
2:08
communication type process you know kind of like Ryan said I think it frees people up to be able to allocate their
2:15
time and not necessarily more important because very important but things that
2:20
require that upskilling and creativity [Music]
2:30
foreign first of all thanks a lot Ryan and Hunter for having me on the show it's
2:36
always uh exciting to meet people who are interested in AI you know I've been into space since about 2015 and you know
2:43
at the time it was hardly possible to generate handwritten digits with these
2:48
generative models and you know fast forward eight years here we are like just having Flawless pictures being
2:53
generated we're in the cusp of engineering videos you know we have like extremely good language understanding
2:59
text to speech speech to text translation so there is just like a so much cool development but uh basically
3:06
about me um I am as you can hear from my accent I'm not from around here so I was born and raised in Prague the Czech
3:12
Republic and came to the US around 2012 to play Squash this small liberal arts
3:18
school called Bates College and I actually I was not a computer scientist major I studied mathematics as well as
3:24
chemistry I actually did my uh thesis in quantum chemistry when I was shooting lasers at gold nanoparticles and I
3:32
decided I did not want to do it a PhD in quantum chemistry so I actually ended up being in Consulting for a year you know
3:39
just like you're kind of out of the school uh really you know vanilla PowerPoint kind of um monkey job but uh
3:47
you know it was definitely a good learning experience so see how business works and I ended up going back to graduate school at Harvard where I got
3:54
my Master's in data science and that's kind of I really started getting into the deep space in Ai and been there ever
4:00
since um so I worked um so my my kind of uh educational background was focused on
4:07
Master's levels focused on AI bias and and fairness in machine learning so that's kind of what I did my thesis on
4:13
and then I joined Fidelity Investments where I worked on recommendation systems
4:19
um so personalization in general it's kind of the big bucket recommendation system is more like the machine learning
4:24
version of that problem and so the way I always explain this is you know if you
4:30
have a set of users instead of items you're trying to match up the items to those users so what does that mean so
4:36
let's say you go to Netflix and each person gets a personalized selection of the movies the cards they get like the
4:42
carousel uh you know I would be somebody who's creating that system which recommends you that selection of the
4:48
movies or you go to amazon.com and then you get a selection of products based on your past purchase history that's also
4:54
like a very similar problem uh now the algorithms below the surface are different based on the use case but
5:01
kind of in general like the idea is you know you're trying to recommend items to users who would like to like those items
5:07
Spotify is another example and so I've watched that Fidelity for about three years and after that I
5:12
started dabbling in startups um I actually was part of two uh proptech companies one uh sort of part-time while
5:20
I was still at Fidelity uh and then the second one really after I've left uh
5:25
last summer which was a a lending company where I served as a CTO and we
5:30
were basically creating a Marketplace for uh for Real Estate Investors to
5:35
connect them with sources of financing um ultimately the business was not growing as much as I wanted to I feel
5:41
like we couldn't really solve the cold start problem and so around March I
5:47
really kind of decided to pause the adventure and focus back to where my roots are which is you know artificial
5:53
intelligence and starting dabbling around uh things to do and you know upskilling myself and like kind of just
6:00
resuming the activities I did before which is like AI research and I you know
6:06
created this project called LGBT which you know which is why I also get invited to this show which is a project I can
6:13
talk about a bit more but that sort of what got me here plus I am working on a
6:19
early stage Venture called crusty Ai and at crusty we help businesses scale Revenue by allowing them to talk to
6:26
their customers to the scale so you think you can think of like a sales GPT connected to a low latency scale like
6:32
scalable distributed infrastructure so that you can call to all of your customers yeah wow what a background you
6:40
got a lot uh if Fidelity Fidelity you're meaning like the investment platform correct uh fiddly Investments yes so I
6:47
would say uh asset manager uh yeah it's one of the largest asset managers in the United States yeah yeah
6:54
um I use them a lot but uh just curious so what are your thoughts is Fidelity do you know Fidelity is making like I know
6:59
Bloomberg has a Bloomberg GPT you know Fidelity is making something similar uh I'm not aware of of that
7:06
um and you know I probably have stuff I'm covered under NDA still checking to talk about but uh you know we did have a
7:12
like a large data science um I would say Workforce so we had a lot of
7:18
data scientists working a lot of cool stuff stuff uh especially for me I was in the uh customer experience marketing
7:25
so we basically made the uh like the experience with like Fidelity digital uh
7:31
better by you know building like a scalable recommendation systems for for our for our customers
7:37
awesome okay so before we get too uh deep into the technical side of uh sales
7:44
GPT can you just give us a brief overview of what is sales GPT what's the vision of it what's it look like yeah so
7:51
Division behind cell gbt is really to build like an autonomous uh sales agent which is fully open source and can can
7:58
basically interact with you by any any channel a written channel so you can
8:05
think of interacting with a text message you can think about interacting by web Plugin or any other means can be
8:12
connected by email and the vision is really to be able to fulfill the tasks
8:18
of a sales uh representative but it is an AI agent it is actually not a real
8:23
human but it can act as one and you know obviously we have the ethical concern
8:28
has to disclose that is AI for most of the use cases but basically the experience of uh you
8:35
know interacting with it is extremely human-like because it can understand you
8:41
know what stage of the conversation especially in sales you're at it can react to your objections you know it can
8:47
present solution where necessary it can you know schedule an appointment as a follow-up or another call to action
8:54
whatever the actual user wants it to configure for so
8:59
um I guess in a nutshell sales GPD is a is open source AI agent which can uh
9:06
basically 100x your sales Workforce that's pretty cool so um so what I guess to get a little
9:14
bit behind the scenes a little bit what does your Tech stack look for something like that is it relatively simple those
9:19
are pretty complicated yeah so we are leveraging basically the you know latest advances in generative AI uh where we
9:27
have the access to foundational models right uh you can think of jigpt as like
9:34
one of them but there's other ones you notice anthropic there is Google here there is like there's a palm from Google
9:41
and basically um the way cell GPT is built is that you can swap any sort of brain if you will
9:49
uh to then to then power cells GPT so as far as the stack goes you know we're
9:55
using Lang chain uh to be able to do the orchestration of of the of the back end
10:01
llm as well as uh you know drum injections and then the agent actions
10:08
right so like basically we have two modes for sales GPT one is just a pure
10:14
conversation where you know one of the main problems currently with llms is that they hallucinate right so they
10:20
actually make stuff up and uh you can actually reduce that significantly by
10:25
constraining it to only talk about your products and that's kind of the example I have on our repo where you know the
10:32
agent is only going to be referencing the products you have you know for a catalog and that's just like that running example I use with mattress
10:39
selling agent right it's not going to sell you any mattress it's just going to only sell you the mattress it has in a catalog so that it can like look up
10:45
information and just reference it there so I would say as far as the repo itself
10:52
um you know it is not a uh you know a scalable distributed infrastructure just
10:57
kind of would be building at crusty but it's like a really I would say like an example where people can use to build
11:04
their own you know sales sales Bots and you know we got over 850 Stars people
11:10
are like reaching out opening issues you know creating it for their own use cases I get a ton of inbound interest through
11:17
these type forms I have put on the readme where people you know are reaching out to use this from anything
11:23
from like a Chinese uh you know like a materials like manufacture all the way
11:28
to you know somebody contacted me uh from a Dominican Republic call center they wanted to automate a call center in
11:34
the Dominican Republic so you know you get like a lot of people interested in this and like they come from all walks
11:40
of life so it's pretty cool that's awesome now how would the like say if someone wanted to implement this
11:45
at a call center how would they go about you you have the back end technology for the conversation but how does it how do
11:51
you hook it up to actually making the phone calls yeah so that's kind of what uh for the phone calls that's what we
11:56
work at crusty so we have a uh again the distributed low latency infrastructure uh which is kind of the the product
12:03
we're selling here and you know it is leveraging like obviously Telecom providers and then um it has a bunch of
12:11
like optimizations which I worked on while I was doing quantum chemistry to basically do it like almost real time
12:18
um as far as like hooking it up in your environment as a chat agent there is a
12:24
you know many open source solutions to do that uh you know you can use like streamlit to create like a quick app uh
12:30
which you can just power with sales GPT like there is like other I think kind of stack or sales sales stack or some of
12:36
these like they're they're just like a bunch of these OtterBox Solutions you can connect it as a front-end too
12:42
okay now can you can you talk about you alluded to it at the beginning and um about crusty AI it sounds like they're
12:48
two separate projects what's the uh as an overview what is uh across the AI and
12:53
how does it relate to sales GPT yeah so aggress the AI is a uh it is a
12:59
early stage of Entry which helps businesses uh with uh a few few key use
13:05
cases which were not able to be solved before like the generative AI Revolution one of them is talking to all of their
13:12
customers at scale so you can imagine right like if you're a business with a bunch of leads in your CRM you have a
13:18
certain budget of who your sales people can talk to right they're gonna qualify let's say the top 10 percent and talk to
13:24
them with uh with Rusty you can actually talk to all of your leads and like really
13:29
figuring out like who who is ready to be activated so let's say you have three
13:34
tier users let's say you are you know DocuSign or MailChimp or any other kind of company which is selling a product
13:41
which is a few hundred dollars a month uh or like a year even right and so those products are usually such a low
13:49
contract point or they have a low margin such that it doesn't make sense to employ A salesperson to call all of them
13:55
now that issue can actually be solved where you can literally talk to all of your customers at scale hyper
14:01
personalize the offering to them and then upsell them to a paid tier um so that's kind of what we focus on as
14:07
the first beachhead Market um is upselling you know B2B SAS businesses helping them upsell from free
14:13
tier to a paid version you know help it onboarding and stuff like that on the back end we are leveraging sales GPT as
14:20
the as kind of the the chat functionality but we do a bunch of to
14:26
obviously do that in a distributed fashion and like you know we host and we are doing that on low latency as well as
14:34
provide access to tools okay awesome now have you ever
14:41
like have you been on a call with uh when's one sales GPT GPT is at work and
14:47
sort of over like listen to what's going on or like how does that work I'm just curious on like how the actual
14:52
interaction happens if you do not disclose it it's an AI bot and the the AI bot calls somebody how does it
14:59
usually go Have you listened in on that and how does that work yeah so I mean from an ethical
15:05
standpoint when you do an outbound call you are supposed to disclose that it is
15:10
AI calling it's not a regulation which is signed yet but for example in the open AI in terms of service you actually
15:17
have to disclose to a human that they're they're interacting with the AI it's actually part of a TOS so if you're
15:23
doing outbound call uh if you're doing inbound the user is you know already opted in and they know that they're
15:30
calling an AI so you don't necessarily have to say then because you're like you know the person is kind of expecting that
15:36
um so those are the two uh you know inbound outbound and uh I would say
15:42
yeah like I mean like it's a visceral it's a visceral feeling where you know you're talking to AI but it doesn't
15:47
sound like an AI right like it doesn't sound like it sounds like an actual human you know you have sort of emotions
15:52
it can like really react to you what you're saying you know it is it is there it is like it is listening to you so
15:59
um yeah I would say you know there is gonna be a crossing the chasmine kind of thing where people who are you know
16:06
let's say in their 2000s were not not as comfortable paying with a credit card
16:11
online right they might probably not be as willing to talk to AI at this point and we're gonna have other people who
16:18
are more you know Advanced uh in technology adoption curve and so those
16:23
you know they might not have a problem with that now as we're as we are like you know raising a new generation you
16:29
know those who grew up on social media and now we're gonna have people who grew up with AI agents just being around them
16:35
right it's not going to be even even a problem for them talking to an AI so you know I think definitely there's gonna be
16:41
adoption curve uh I think you know you have so many companies or so many you have like a few other players uh in the
16:48
various parts of the kind of the AI voice ecosystem right like the notable ones you have character AI you have the
16:57
um stability Ai and all of these like kind of companion uh AI companion companies are creating these use cases
17:03
where you can actually talk to your AI and you know I think we're just in that start of that adoption curve and it's
17:10
just gonna get more and more prevalent as we go awesome and how how does so when that call is made how is it
17:15
disclosed is there some like right away is that the first sentence that comes out of the recording
17:21
um for lack of a better word yeah so it's not recording uh really because it generates the uh you know
17:28
everything what it says on the fly so it's it has the ability to basically react what you're saying because again
17:34
it's not recorded it's generated right so like it's it's generating a speech as I'm turning speech talking to you in a
17:40
way in a streaming in a streaming way um you know it can say things like hey you know this is this is John I'm an AI
17:46
assistant from XYZ company I'm talking to talk about talk about X you sign up for this product why you know and then
17:53
I'd love to talk to you about like how this uh additional features can help you or like what problems are you facing in your business
17:58
okay it seems like that's like like you said it's going to take time for people
18:03
to be to pick up the phone and hear that it's like oh this is an ai ai sales
18:08
person to be comfortable with that it's just going to take time because right away I would think a lot of people would hang up if they hang out with regular
18:15
sales calls they're probably gonna hang up right now but maybe that will take care of itself and there's probably workarounds from that as well to maybe
18:20
have some giveaways or something like that when you you know if you if I picked up an uh AI sales agent is on the
18:28
phone he's like I'm an AI sale sales agent but I'll give you 30 off of what I'm about to tell you maybe that keeps
18:34
them on for a long time or something like that yeah I remember like I'm not really focusing on the kind of aggressive sales use cases it's more
18:41
about like the person already knowing they're going to be talking to an AI so for example it's somebody who's on a
18:46
free tier and they would like to understand how the paid subscription is going to help them right you can either go through the FAQ Pages or you can wait
18:53
30 minutes in the phone call Center for somebody to talk to you or you can talk to an AI right away who can help solve your problem right so it's all about
19:00
aligning the incentives of the person and expectations right so you will not get tricked about talking to AI right
19:07
most of these use cases are actually inbound where you know like the person will call the AI in and will want to get
19:15
their problem solved so I mean they're still going to be adoption curve but it's almost like a help desk as well
19:21
from what sales GPT it can be yes like as far as the calling
19:27
solution 100 um I would say that like you know that's kind of the main use case the other one
19:32
is like you know you go to a e-commerce website and like within five minutes of placing your order like you can also
19:38
click a box I would like to talk about you know my order with like an AI agent who can help me do XYZ now that's gonna
19:44
take some time but you can imagine that like if you have a very specific question or like a very specific
19:50
business you would like to talk to somebody right about it like you might not find it in FAQs
19:55
um you know and like if you have a lower margin business you might either talk to somebody who is you know sitting
20:02
somewhere you know overseas uh and like you know they might not necessarily know
20:08
what's going on or they're just kind of like a more like a support here usually actually talk to something which is very knowledgeable about a product like it's
20:15
always there and it's ready to help you so as far as like the calling solution yeah and you know we can drop in the
20:21
link for the landing page people can check it out I have like a video but I mean maybe we can just uh kind of go
20:26
back to like the sales GPT uh repo which is you know a bit it's connected but
20:31
it's uh it's sort of like a different Focus right like on the sales GPT report you have to focus on the chat functionality and really just like
20:38
configuring the brain if you will so maybe we can just like Linger on that a little bit since I think that was the main point of the invitation today yeah
20:45
so I wanted I wanted to ask so you said you can do it through what
20:50
are all the channels that it's able to do it from you said text uh call you can do email
20:56
what was there was there other one or yeah so I I've seen so I've seen people uh wanting to integrated it to a
21:03
multiple uh channels as far as sales GPD goes right like and we have everything under the
21:10
sun we have text we have uh email we have like uh HubSpot action tasks we
21:18
have like Salesforce Integrations there's WhatsApp it's a big one right especially out of the US Facebook
21:25
Messenger or like Instagram Messenger as well as uh uh mostly Chinese one
21:32
um messaging app
21:38
WeChat I can't yeah WeChat WeChat yes yeah yeah yeah so I
21:44
mean you know we have everything under the sun right like and these are not coming from me these are from the users of the repo who are contacted me and
21:51
they're like hey I'm doing this for that right like I'm interested in this in this use case yeah
21:57
so I'm assuming like you said it's very uh human-like interaction so I'm
22:02
assuming there's some sort of like Persona that's giving off given off by this this uh sales GPT uh do they get to
22:10
does the user get to customize that at all or you know it with a voice what
22:15
kind of voice tools do you use something like that how does that work yeah so
22:20
um the customization that comes like directly into sales GPT prompting so you
22:26
know it's kind of similar if you are like talking to Chad GPT in like the
22:31
actual like UI of chai GPT you know most people just have like an interaction with it when they ask it a question you
22:37
know they want to solve a problem uh what is kind of the next level is you
22:42
can actually prepare what's called a system prompt and that's come directly like you can look it up in like language documentation and it is system prompt
22:50
basically you like constrain the llm to do what you want do the best of your
22:56
ability so like it's usually like a longer prompt and you can see that in the configuration file on sales GPT
23:01
right like you have some custom-made prompt and it's telling it hey your name is you know Rachel and you work for
23:06
Sleep Haven and then this is your business and then this is what like you're selling this is the reason why a
23:13
person is calling you or you're calling them and then this is like all of the stages of the conversation right so you
23:19
basically prompt it all out and that's something a user can um can like actually change you can tell
23:26
it hey like I want to react in this way so like you know you can you can act like as a pirate if you will like you
23:31
say hey say Arc after each like sentence you can you can set it up in that way right like the imagination is really the
23:38
only like limit here because it's just such a general like you know like general purpose uh foundational models
23:44
which were built so that you know train on the authority of the internet so you can like really use them in any way you
23:51
you would like the the follow-up was about um that's a really good answer I like that that was a really good explanation
23:56
uh the follow-up was what kind of uh voice generation tools oh do both tones yeah so you know there is a number of
24:04
providers in the uh voice generation space you have like play HD you have 11
24:09
laps you have like um a bunch of bunch of other ones like you know meta has released the audio
24:16
craft uh recently which is a uh just kind of like a music and like just voice like a noise generation one they have
24:23
voice box which is actually not open source um you know and there's like a bunch of other ones this bark you know there is
24:30
like a number of these you can use and and customize for your own liking okay and then you said you get to pick the
24:35
llms too right so yes absolutely and there's are there some that perform
24:40
better than others as far as sales and overcoming objections things like that that you found yes I mean that would be
24:47
great if they'll be great if somebody could like Benchmark that and let me know I have not actually done like the the benchmarking of like which
24:54
Foundation model is better performant in general uh you know you would like to
25:00
get you would like the foundational model to have access to data related to sales so I could imagine that Google
25:09
Palm model and that's just like my you know conjecture like I have not proven
25:15
an icing but like it might actually work better because it has access to all of this like you know like e-commerce data
25:22
which which open-air might not have right uh like everything which is from their ecosystem from like a closed
25:28
databases which they can use to train their Foundation model which somebody like opinion might not have access to
25:33
right um so I think like Google has like a huge advantage in that sense
25:38
um in the same way like you know if you have Elon Musk trying to build something very conversational like Twitter style
25:45
bot right with them with with Elon basically constraining the um
25:51
closing out their API and not really allowing like third parties to to train you know on that they might get that
25:58
Advantage so basically it really comes down to what data sets you have available to do to building your
26:03
foundation model uh the second one is budget obviously we're still at a point where we do have like open source models
26:10
uh you know llama version 2 is now the state-of-the-art in the open source but it's still lacking behind the closed
26:17
Source component uh just because the size as well as the the sheer volume of
26:22
data it can be trained on okay so does it cost to use right now or does it just cost uh due to the llm you're using also
26:29
cell GPT is is is uh is open source GitHub project anybody that can clone it
26:35
install it and then just use like you know it is a free to use
26:40
um under MIT license however for like connecting it to llm you either have to
26:46
train your own or like just spin up your own host your own which you can do now right like with things like quantization
26:52
and like running on a local machine you can actually spin up like a smaller model but if you're using like an API
27:00
key from open AI for example yes you have to pay for that gotcha so what's what's been the biggest complication to
27:06
the development of sales GPT what would you say is the biggest complication right now I mean I would say it's time
27:13
it's a time right component like uh I would love to get people uh like other
27:20
developers who would like to contribute to the repo I have like a bunch of like I have a ton of stars and ton of issues
27:25
but uh you know people using it for their own apps uh not that many people actually contribute back to the open
27:31
source so you know I I do have like a running to-do list of things I would like to improve but um
27:37
thing on our hands like other developers who would like to help making this better for everybody um and obviously they can like use that
27:44
project as on the resume or like if they want to build their own apps right like they can they can do that
27:51
um this is sort of like the backbone for building these things and on it it's built on Lang chain right so you can probably build a similar thing
27:58
online chain except this is more like I think sale specific so if somebody's looking for like a sales specific use
28:04
case um this is like more suited towards that it can give you like a better examples but understanding what's needed so like
28:10
what's what's one big Improvement you'd like to make or that you would need help with right now you you wish you could
28:16
just like get that done yes I think right now
28:21
the uh as I talked about usage of tools um you know it's not always
28:28
100 bulletproof and the reason for that is you're actually relying on the formatting of the llm out so like let's
28:34
say if you want like a certain output you wanted to say certain things so that you can parse the string coming out uh
28:40
it's actually not always as uh robust now there is one way to to uh to get
28:46
around it and that's using open air functions which actually like the main functions like just in general and the
28:53
the actual difference there is that there's a guarantee you will always get a Json
28:58
back so that like even if your result is not um
29:04
it's not what you want it will not break the flow of the code because like you know you'll get a Json viruses if you're
29:09
just relying on a you know a thought action observation pattern right which
29:15
is kind of what what the react uh the react framework is what what people use is one of the early early uh agent
29:22
um Frameworks if you don't get the thought react observation framework out
29:28
of the LM you know you can break your flow so that's actually one of the the pieces we just wanted to do is to make
29:34
it more robust it can be open AI functions could be something else but yeah improving the the the the
29:40
functionality of the agent in that sense have you seen any examples of uh people using the project uh in real life right
29:47
now that actually have help their business or created Created any kind of basic sales for them
29:53
absolutely I mean I have over 100 people who reach out to me on the typeform channels uh and they're all like using
30:01
it for their business right like anything from a Gmail integration to like a WhatsApp agent I think those are
30:08
the two most popular use cases um and yeah like they're using it to improve improve their their sales
30:13
process sales Pipeline and you know uh where I'm plugging in a bit crusty is kind of like white labeling that that
30:19
calling solution so we are building our own front-end and ability to you know um sort of help with your like
30:26
customization uh but you can also do it yourself a sales GPT right and and then like if you want calling that's kind of
30:33
where we plug in with across the um but yeah like definitely I mean people are using it actively you know I
30:40
mean it's it's about 140 Forks on that project so like you know I'm sure that like there's there's people who from
30:46
there you you are who are using it for their own business this and uh you know um obviously like it's a sales thing
30:53
people want to build products out of it it would be great sound they want to also bring a little PR back to
30:59
contribute back uh to the open source so this is a little Notch there no you're
31:04
definitely a hot project I think that's how I found you guys is on the GitHub trending projects uh because you're like
31:09
it sounds like you have a lot of stars and forks so that's awesome congrats on getting this far
31:14
um thank you yeah what uh what so you you mentioned a little bit well let me for I've got a couple of
31:21
more a couple questions um but for the Crusty AI is that sort of just to clarify for the for the everyday
31:26
user Krusty is going to be sort of the like you said the white label the interface if you want to use sales GPT
31:33
in the cloud um for example and if you want to use sales GPT locally you can just clone the GitHub repo and use it locally however
31:40
you wish is that sort of the 10 basic 10 000 foot overview do I have that right sort of I would say Krusty is the sales
31:48
GPD powered voice sales agent which can talk to you over the phone so it's more
31:54
suited towards the voice like conversation uh than the chat and then the basically the value out there is
32:01
that you don't have to build your own scalable low latency infrastructure and it's all managed for you and you can
32:08
prepare your agents with LGBT like the prompting the way the brain interacts right like the response is going to be
32:13
the same but you can do that all at scale with like additional optimizations and like the ability to basically call
32:20
and have that all manage for you so yeah so it is the uh it is the calling calling solution calling complement uh
32:27
to the uh to the open source and are you the I think you mentioned it earlier at the very beginning of the podcast but
32:33
are you the co-founder of crusty AI or what's your role at Krusty yeah I'm a
32:38
co-founder at crusty that's right all right awesome and what was your
32:45
like why did you feel the need to do sales is there a reason like it sounds like your background's in marketing a
32:51
little bit as well from the companies you worked at so is that sort of how it translates yeah computational computational marketing
32:58
right was one of the one of like the uh previous I guess roles like at Fidelity
33:03
I would say it's mostly computational marketing right like personalization you know at scale like you know we were
33:10
serving recommendations for like 30 million plus users at scale right so you have like these massive massive systems
33:16
uh basically right so um yeah like I'd say like the problem
33:21
I'm solving you know with sales GPT is just um the ability to like okay so like the
33:29
ability to qualify at least like in my previous business when it was a landing business um you know the way we were sourcing our
33:35
customers were through Facebook uh groups for Real Estate so you can imagine you have like Oregon Real Estate
33:41
Investors right and you have a bunch of people there and a bunch of people looking for loans so we would contact them
33:46
and because our value proposition I was like hey get a selection of loans without paying the broker fee now
33:52
there's gonna be a bunch of like issues with that like people don't trust us because we reach out to them for Facebook they don't know our brand right
33:57
but the main thing was that I got on all these phone calls with people who were not qualified to purchase so like I
34:05
would spend all this time talking to these leads who are not actually leads because they were not actually ready to take to like buy real estate do a deal
34:13
right like a flip or like buy a home and so I spent all this time talking to the people who are actually not customers
34:20
right we're not qualified buyers so you know I think a part of like creating sales gpe was for me to address this
34:26
problem is like can we create something which you can address like the big problem of like spending all this time
34:33
and money of qualifying leads right because I was sort of like a salesperson two of my previous startups right like you wear all the hats if you're early
34:39
right you build a product you do the marketing you do hiring but you also do the sales and if you have people wasting
34:46
your time in the sales process that's really not no early stage founder wants to be in that position right so I'd say
34:53
like subconsciously that was the reason why I chose this this uh this thing but it's also like a very exciting because
35:00
now you like kind of in a way of human computer interaction where
35:05
um you know this thing can actually like have a real conversation right like previous systems for robotic you know
35:10
they're kind of like really heavy to use you know like clumsy and clunky but now
35:16
you're like actually with the current technology and that by the way is just not just llms it's the very very good
35:23
speech recognition models which have been built in the last two or three years as well as the speech generation
35:29
models which were built in last few years and those combined with the llms allow you know Krusty to exist
35:36
um and for cell GPT you know uh the llm itself kind of a core brain that's kind of the reason why it connects us okay
35:42
yeah we re we uh interviewed a real Char uh last last week uh Sean wee is the
35:49
founder um yeah he did a little demo for it for us and it was pretty it's pretty incredible like how good it's gotten
35:55
like he has a he has a care like he has characters like Elon Musk and uh other other celebrities
36:01
um that you can talk with in real time and you can do it via a phone call or a chat and the one thing that stood out
36:06
that has gotten really good other than the voice that doesn't sound robotic is he was demoing that the conversation
36:12
like they would interrupt each other like mid-sentence it wasn't like someone would ask a question and the AI bot
36:18
would answer and then you'd wait for the bot to answer someone would interject AI bot and his technology was able to pick
36:25
up the conversation like a human would like um you know if you had a question for me right now and you want to interject you
36:31
could do that whereas I don't think that was possible before with AI bot so it seems like that technology has really really come along
36:37
um in that aspect in more of a natural language type of um natural communication type process
36:43
absolutely yeah no absolutely that's definitely a big part of you know solving for the interruptions and then
36:50
uh just understanding like end of speech right the recognition that Senator big one to reduce latency so all yes these
36:57
are definitely very important points so um so one thing that people might be
37:03
concerned about when they hear sales GPT and AI sales people is inevitably well
37:08
what about human sales people or are they going to take all of our jobs um how do you address that concern yeah
37:14
so I mean there's definitely you know a question which comes up and uh like you can think of like any sort of like
37:20
technology uh driven transition right like people didn't want cars they wanted
37:26
to Faster Horses right like and you had Henry Ford um I think here we're gonna see
37:32
basically an elevation in productivity right that's ultimately what this is about this is not about like you know
37:38
what AI can do this is about okay what can we do as like Society right like we want to grow GDP When I Grow you know
37:44
the overall output and like I think that this ultimately allows us to do that now there is like short-term concerns about
37:50
like job replacement um but I think what you're gonna see is that like you know you're gonna have massive or Skilling upskilling and
37:57
ultimately like a greater greater abundance right so I think you know when you are positioning this
38:03
as like helping the existing salespeople of like a one sales person can have a job of 10. now you're like really you
38:10
know putting a downward pressure on the cost of of like doing a certain job which will you know increase the
38:16
productivity of people who are there but ultimately you will always need a human in the loop right like I'm not talking
38:21
about like replacing a salesperson I'm talking about like hey like I can help you set an appointment so you can talk
38:27
to a real person kind of really reduce the sinks which are which sucks about
38:33
like doing sales like for example in cold calling if it makes sense or just doing delete qualification or additional
38:39
kind of things which which sucks as a salesperson so like are we like that's kind of like that that you know that
38:45
that doomsday scenario oh we're gonna like be all our place but yeah I think we're not there yet we're not gonna be there yet even with the current systems
38:51
right like you're gonna have like productivity improvements and I think the pace of change will be gradual like
38:57
yes like we'll have to reskill up skill certain certain you know certain job
39:02
types will not be here anymore we'll have to have we'll create new job types right like we might not need to type
39:08
writers anymore because we can like record you know uh like automatically you don't need typewriters anymore right
39:13
so like that's kind of like a transition I see uh where it it's like hey like
39:18
you're gonna have self-driving cars you know Uber drivers will have will have no chops anymore like that kind of thinking
39:25
like you know it's it's something which is happening on a micro level but I think it's happening uh gradually enough
39:31
so that people can can adjust and kind of you know get used to it yeah it seems like like when we go back to the sale
39:38
like uh real estate sales you know if a realtor wouldn't have to reach out or
39:43
qualify leads for example they could build their business or make something but like they don't have to deal with
39:49
that low-lying fruit so they can maybe expand their business or scale their business in better ways so I don't know
39:54
that argument about it like replacing their job maybe it just makes their job better and they can it makes their business better ultimately because
40:00
they're they're more freed up with things it's like you know absolutely yeah yeah I mean like ultimately in
40:05
every business you're selling to people right so like there needs to be a valuable position for somebody to to buy your product and you know I think that's
40:12
kind of like what what it is here is that you're like improving if like productivity of the existing people
40:18
working for that business yeah and to be clear I wasn't saying that that is the case I'm saying that's enough no
40:24
absolutely no there's definitely like a valid yeah that's a valid comment right like that's something you know you get a lot of
40:31
Automation in general yeah yeah and I think it you know kind of like Ryan said I think it frees
40:37
people up to be able to uh allocate their time and and not
40:42
necessarily more important that those things are very important uh but things that require that upskilling and
40:48
creativity that they might be able to contribute to more uh to advance project
40:53
uh solely um you mentioned like the sky is the limit you know it's kind of like the
40:59
Innovation you know the imagination kind of determines what's going to be done with this project do you see any other
41:04
limits to the project right now um or or is it basically just kind of up to the human imagination and what the
41:10
developers are willing to to do I mean like look it is an open source project right so like it doesn't
41:16
actually have like the product uh reliability or like you know any any sort of like you can't just like drop
41:22
this and like uses it as a product right like you just have to do like the
41:28
additional work to put that in a product like ultimately cell GPT is showing you how to do certain things how to create a
41:34
sales agent uh autonomous sales agent it doesn't doesn't give you all of the stuff you need around building like
41:40
actual product right which is like oh you need authentication like user management you know all of this stuff
41:45
which is kind of creating an actual product is not there so you know
41:50
um it's more meant to you know in a company which might have something building something and you just drop it in and part of their workflow right I
41:56
don't know if you kind of familiar with the traditional machine learning cycle which is probably changing too but like
42:02
let's say the last 10 15 years then you're creating a ai ai product the AI
42:07
is like a little piece there and then you have like the entire like infrastructure around it right like everything from like a data injection
42:12
you know data management like training serving inference security you know
42:18
quality of the output like all of that stuff around it needs to exist so you know I think that I've heard this
42:25
interesting uh comment I was listening to this podcast called No priors and like did Switzer Google and she was
42:31
talking about how like um not that my companies actually have
42:37
llms in production yet because like um that's what I'm talking about like you get to a demo stage but to go to that
42:42
like production level system you actually have a ton of stuff on top of that so you know that's kind of where we
42:50
are like I think I'll continuously like to improve the Core Concepts there like allow you to talk to it by terminal and
42:56
doing all of that or maybe like have like a little chat there but yeah I think like you know the time spent on
43:02
this project and then uh and have having contributions from others would be super helpful do you have a team of developers
43:09
right now if so how many like how many people are working on this project every day I mean cell GPT that's just my effort
43:16
right it's under my personal name on GitHub and I GitHub handle so I was the one who was developing all of that I have a few advisors on who helped me
43:24
with sort of like the uh the infrastructure as well as like um
43:29
you know like concept thing um as far as Krusty goes I have a small team of of uh folks who are helping me
43:36
out with uh with uh with a startup there but uh yeah it's it's a very small team
43:41
you know we are definitely like an early stage startup yeah and are you guys uh if you don't feel
43:48
like answering this question that's totally fine are you guys VC funded or are you looking for funding or how um
43:53
how are you funding crusty at least uh so yeah it's it's currently uh it's
44:00
currently not uh not VC funded yet um we are definitely open to that
44:05
um you know but we already have like paying customers so like you know we sort of are not in a in a huge like I
44:13
guess time crunch like obviously would be nice to grow faster and you know I'm definitely open to taking some
44:19
conversations uh but we're not actively you know raising uh right now um I think
44:25
that might become more relevant as time goes on in the next you know month or two uh we'll see how things go but uh
44:31
yeah like it's you know the sales GPT plus like like the open source I plus
44:37
crusty is the closed Source sort of like callable calling infrastructure cluster
44:42
uh where that's where we get like our traction and it's been it's been about right you know I I've been very very
44:49
busy my bad I mean how long you how long has it been since you launched it hasn't been long at all right
44:55
uh no so the the the the beta has been out for about the month uh like cell GPT
45:01
open source repo has been out for about you know three or four months right now I think like end of end of April so it's
45:08
like what like three and a half months or so uh since the first commit and then for the Crusty you know I literally just
45:15
put up the landing page I think like last week or something um yeah okay awesome well you're on the
45:21
right track I mean it looks great it's an incredible project makes me want to like go find out what to sell after this
45:28
podcast you know and test it out yeah yeah yeah but we can start mattresses
45:34
there we go mattresses hey everyone needs one right so yeah I
45:40
know it's the but the product Market fit right yeah so what is your uh what potential risks
45:47
do you see for this project do you see any potential risks that need to be addressed yeah I mean like I think one
45:53
thing is I was gonna say for sales GPT in general but you could also talk about crusty if you want uh yeah so I think
45:59
for cell GPT is just the um ability to get like other developers to
46:05
help out on features and just make it like an open source versus just like one-man show right like or like my team
46:10
kind of show um the other thing is like to continue
46:16
being useful I think as like an educational tool uh like continuing to push out like latest like Concepts and
46:24
you know like paradigms because that thing moves very very fast right so like the initial stuff which was there since
46:30
May might be a little bit outdated already right like there's new better ways to do prompting and do like you
46:36
know how to set up these agents and stuff like that so like keeping it I think current it's a challenge
46:42
um especially when you are you know volunteering it's open source repo right so like I'm not monetizing sales GPT so
46:48
like you know it's just like there for anybody free to use and I'm