Subscribe
Share
Share
Embed
Watts in Your Data, hosted by Denis Gontcharov, explores how enterprises in energy & utilities leverage Databricks to improve operations. Listeners can expect in-depth technical discussions and interview that break down complex topics automated data quality testing, and advanced analytics into understandable segments, actionable insights, and real-world applications.
More About Me: https://gontcharov.eu
Denis: Hello and welcome to
the Watts In Your Data podcast.
This is a new episode about solving
data problems and having solutions
for the energy and utilities industry.
And today I'm joined by Serena.
Hi, Serena, welcome to the show.
Serena: Hi.
Thanks for having me.
Denis: Well, it's a pleasure.
For people who may not be familiar with
your presentation on Databricks, could
you tell the listener a bit about your
current work and what you did in the past?
Serena: Yes, of course.
Um, so I am the lead data
architect at Blue Digit.
Bludigit is, um, the IT company of
EGAs and um, EGAs is, uh, now the first
guest distributor in, um, in Europe.
Uh, we don't.
Just work with gas, gas distribution.
But we also have, um, IT as like advisory.
We sell our, uh, proprietary solution, but
also we, we work in water, um, services
and also energy management services.
it's kind of a broad energy company.
Denis: By the way, for the listener,
the reason, um, how Serena actually got
on this podcast was that we met on her
presentation, uh, in, on the Databricks
Data and AI Summit in July, 2025, and
she had a presentation there about
Italgas and their use of Databricks.
And I thought it was quite
interesting to go a bit deeper on
certain topics she mentioned in what
they wanted to do in the future.
So now we're gonna focus on that.
And with that I can introduce the topic
of this episode, namely using AI agents
to automate, in this case IT operations.
But as we will see it, gas also uses
agents in other areas of their business.
So I think it's gonna be a
very interesting episode.
Can you tell us a bit more about
the ambitions of Italgas, when
it comes to AI agents and what
you have running at the moment?
Serena: Yes, of course.
Um, so the ambition is quite big.
We started, um, uh.
In 2017, a big data
transformation journey.
Um, last year, I would say around this
time, we made a strong move in our, um,
vision of the future of the company.
And so we, um, we decided
to invest a lot in ai.
we, we don't.
Believe that AI is going to replace
people, but we think it's going
to make process easier and it can
help people doing more interesting
jobs, more interesting task and
reduce the workload that at the
moment is really, really heavy.
So, um, what we did, we, it was to.
Like merged together, the IT
department and uh, the human resources.
And we created the ai AI factory,
which is, uh, uh, a business unit
inside blue digit, but it also has
HR people, not just technical people.
Um, starting from basically one year
ago we developed around 10 agents
that now we have in production.
And I think one of the most
interesting, uh, is, uh, the one
about ticket ticketing resolution.
Denis: Yeah, that sounds great.
Uh.
And one important caveat I also learned
from your presentation earlier was
that it, gas in fact is quite far with
having their entire data infrastructure
on Databricks, meaning that you have
your, let's say, ticketing program
known as ServiceNow, a very popular
ticketing system in the industry.
You have it fully integrated
with its Databricks.
Serena: Yes, exactly.
Basically we, uh, we are
using, um, lake Flow Connect
Connector.
So, uh,
something that is, uh, already
there in your Databricks platform.
quite new.
I think we started using it that it
was in public review, I think, but I
think now it's, uh, general available.
it's really, really easy
to, to manage and to set up.
So, you can get design in real time with.
Basically no effort.
Um, so that's, that
was our starting point.
So we started using the lake flow
connector and we thought, okay, cool.
Uh, this could actually help us
solving the ticketing problems.
'cause of course we have a lot of
tickets because, uh, I mean, we are in.
In a very particular moment of our, um,
our, uh, company history because we just
merged and double size of employees.
So a lot of
people dunno how to use our tools.
So we like tripled our tickets
and we were expecting that.
So we knew that was going to happen.
So what we decided to do was
to, uh, build vertical agents.
platform that could help our, um,
our new, employee, colleagues, uh,
to, to work in those, uh, platform.
So they were vertical and in embedded
into the front end of the application.
So they could ask question
as a, basically a chatbot.
We had those agents, we had the
ticketing, and we thought, , these
two things could be combined.
that's what we did.
Basically, we, an agent, uh,
with an MCP server so we can
talk with other agent and this.
This agent in ServiceNow, the front end
of ServiceNow, but it works on Databricks.
It's everything in Databricks.
Just the front end is serviceNow
Denis: let's focus a bit more on
the actual problem because indeed
it sounds very familiar to me.
I also currently work for a company
that has grown a lot in terms of,
let's say, connecting more and more
sites, in this case, wind mill parks.
So you have a lot of extra
systems, and whenever systems join,
they're usually very different.
So you have the problem of
onboarding people or systems.
So if I understand correctly, what.
Your agent solution allows is to help the
current IT personnel and new IT personnel.
Um, do use the new systems
or how would you describe it?
Serena: Yeah, we started with manual,
uh, building the, knowledge base with,
uh, manual and, uh, previous tickets.
So like whatever it was, it wasn't
working or it wasn't clear for.
Our, our user.
Um, so we developed these 2, 2, 3 agents,
to help in the most, um, crucial tools.
So, because I mean, we didn't have
time for, a wider audience, but we,
we will like to, to have something
like that on our own, platform.
So every tools that we have, uh,
but at the moment is only three.
Uh, so
basically we just build these agents
that call be your like buddy, so your
assistant
that is there.
Denis: Mm.
Mm-hmm.
Serena: how to do something, you can
ask, I dunno, uh, can you help me design
a new pipeline for, um, uh, these, uh.
This territory or whatever, and
it helps you and it tells you the
step that you have to do, where
the buttons are and these things.
Denis: Okay, that's interesting.
So comparing to the past where user
would have to open a new ServiceNow
ticket and stare at a lot of empty blank
fields and be a bit intimidated by not
knowing what to write, you now have and
let's say, an agent in a separate window
that helps the user with questions, or
how does user interact with the agent?
Serena: So the, the, vertical ones are.
It, it depends.
So basically, uh, you can have, uh,
the vertical one in the front end
of the application, and that's all.
It's very easy and very basic.
But the cool things that we did was to
like, um, connect these vertical agents
to the one that helps solving tickets.
So basically what happens is that the
user goes into ServiceNow front end.
Uh, start asking question and
say, oh, I have this problem.
Can you help me?
Um, so the.
General agent for ticketing, the
resolution to investigate a problem and
if it realize that it's something related
to the vertical agent, calls through
MCP server, the specialistic agent,
to solve the an to solve the problem.
So to give an answer.
So instead of, opening a ticket.
Most of the time, uh, the user
can solve its problem by themself.
So
Denis: Mm.
Mm-hmm.
Serena: the number of ticket.
If the agents are not able to solve
the problem, what happen is that the
agent helps you, right, uh, accurate
So the IT department has all the
information, all the ingredient,
to actually solve the problem.
Denis: I think the workflow make
sense where user asked help to
an agent, and then that agent.
Contacts the vertical agents.
Could you explain a bit more the actual
architecture and what you mean with like
the vertical agents and how they interact?
Serena: So what happens is that you.
Um,
in ServiceNow
through API, uh, it's everything
again done in, uh, Databricks.
So the the PI gateway
is the Y Mosaic gateway.
Uh, that?
Because we have, uh, already the
governance from Unity catalog.
we, so you have for free all
the security, uh, all the, um,
visibility reduction, uh, policy.
So if you can't have an information
on those data, you won't have it.
And it's everything embedded in
Unity Catalog and in AI Gateway.
So, uh.
You don't have to worry about anything.
It's already there.
It's already done.
other cool things is that
you can have, um, fallback.
So if it is, um, so for example,
when we have a high demand of,
uh, request, so API calls, um,
and maybe we can saturate our, um.
LLM uh, instance.
So if we are using, I dunno, open AI
or llama, whatever, um, it can get like
saturated so it can scale to another
LLM, which is quite cold because we
are not giving any, service to our
Uh, so.
Basically through, uh, the gateway,
through the Mosaic gateway, we get a
request in Databricks and the general
agents start to, uh, analyze the answer.
So it has, uh, all the knowledge
base, it's basically the ticketing
world, and start a conversation
with, um, with the user.
At the moment, the history of
the conversation is in, uh,
MongoDB, uh, database atlas.
Uh, but we are actually, uh, experimenting
with Lake Base, so maybe we will
move even that inside Databricks.
Denis: It's true.
Yeah.
They released it, uh, this year.
Right?
for listener Lake Base is the managed
Postgres SQL database on Databricks, which
allows you to store relational data or
like transactional data really easily.
So you want to replace mongoDB with
that, I suppose, in the future.
Serena: yes, exactly.
So everything is, uh, is.
In Databricks so it's easier to manage.
'cause now we, we kinda struggle,
especially on the network side, we have
to go through, um, an Azure function
for some configuration in the network.
so it'll be easier for us to manage
to ev everything in Databricks.
And basically, yes, it goes through
this conversation with the user.
to solve the problem at this point.
If it realize that the information,
the issue is one of, is in one of the
vertical agents, the specialized agent
that we already have, and since, since
the ER all developed in in house.
So we know the code, we, we can
interact with them using an MCP server.
Um, basically through the MCP service,
general agent contact, contact, uh,
the specialized agent he sends, um,
summarized, history of the conversation
Denis: Mm-hmm.
Serena: context and the problem,
what's the problem, and then the
specialized agents answer the question.
And so the.
The business user or whoever it
the the
one
Denis: Mm-hmm.
Serena: um, can give you a feedback and
say, okay, this is solved, or it's not.
So if it's not, and they can go
on, uh, asking more questions to
the specialized agent or otherwise,
it'll help you, compile the tasks.
So the body of your, um, request of
your ticket, and it will give a sum, a
summary of whatever you discussed it.
So you will already know, um,
what are the steps that you tried.
So, it will be quicker for the
IT department and, uh, it'll give
you all the information that.
You actually need.
what we realize is that, uh, most
of the time when we, uh, receive
some tickets, some incident, it's
that, um, the user doesn't actually
write well what it's written.
And you miss a lot of information
of
crucial information
Denis: Yeah, that's very typical.
Serena: Yeah.
Yeah.
But I think it's.
normal because as a user you
don't know what you actually
need to solve that problem.
So it's kind of normal that, uh,
you will need some information.
You won't communicate
because you didn't know.
Denis: Yeah, absolutely.
In fact, I've been listening to your
explanation and I had this metaphor pop
into my head that I think is, uh, perhaps
a good comparison to the workflow here.
It sounds a bit like if you are sick,
you may go to like a general physician,
let's say your own family doctor.
And then that person, if he cannot
solve the problem himself, he will
forward you to the say a specialist,
let's say a lung disease expert doctor.
Is that a bit similar?
Serena: Exactly.
That's exactly how it works.
It with a perfect metaphor.
I will use that actually.
Denis: Yeah.
Okay.
Great.
Can have that one.
That's super, that's super.
Um, well, let's focus a bit then once
you have, in case even the specialist
couldn't solve the ticket, well,
at least that then you have, um, a
ticket with nice information in it,
with all the steps that were taken.
And I imagine this would help
an actual human troubleshoot
the problem much faster.
Right.
Serena: Exactly.
But um.
On the other side, when we receive
the ticket, we also have an agent
that will help us solve the problem.
So,
yes,
we will have a
Denis: Wait, that's a
third agent, you, you mean?
Right.
Okay.
I'm still following.
Mm-hmm.
Serena: Yes.
Basically when we are as a ECT
department, we receive the ticket.
We also have, uh, an agent that
suggests us the right answer.
So in real time, whenever we received
an, uh, an incident or a request,
whatever from ServiceNow is ingested into
Databricks and it starts the agent, the.
Solution agent, and basically what
it does is just compile an answer
and whenever you log in into your
ServiceNow to solve the ticket, you
already have a suggested answer.
on one side, we are producing
the number of tickets because.
Many tickets are already sold by the
agent, and on the other time, we are
reducing the time that we need to answer
to an agent because we have this other
agent, uh, helping us with the resolution.
Denis: That's fascinating.
In terms of reducing, like do
you see or better set, how do
you benchmark any improvements?
Have you actually measured the time
to ticket resolution or your ticket
backlog before those agents and after?
Or how do you measure your improvements?
Serena: So we, we don't really have
at the moment the results because
we are still in the tryout period.
So, uh, we dunno how much
better is doing compared to.
To the previous one, uh, to,
to the method without agents.
But what are we measuring?
Is the number of tickets open?
So we are trying to reduce
that, that's for sure.
A KPI.
And what we did was to collect, all
the main metadata from the ticketing.
So from the, um, uh, when the
ICT department received the
ticket, how many steps it took.
So most of the time there were a lot
of, um, as I said before, a lot of
asking the user for more information.
we are, uh, checking if that
number of steps is reduced.
That's one another KPI that we are
tracking and the time that it took
to solve and close the, um, to solve
the, the ticket, also, um, how many
time it was closed and not reopened.
'cause sometimes it happened that it was.
Denis: Hmm.
Mm-hmm.
Serena: The ICT department thought it
was done, everything was done correctly.
So they closed the, the ticket,
then the business opened it again
because they didn't understand each
other.
So that's
Denis: Mm-hmm.
Serena: that we are tracking,
Denis: we also encounter a
lot that tickets are reopened.
We also have the issue that
oftentimes tickets are forgotten.
Very often you need an action
from a different department.
So you assign that ticket
to, let's say infrastructure,
request them to do an action.
But then that doesn't happen for a few
weeks and then people forget about it.
Do do your agents, for example,
um, are able to send you reminders
or follow up on all tickets.
Serena: No, uh, that we didn't, uh,
implement it, but it actually could be
a nice thing where I can take it into.
Account for the next MVP,
um, but what it does, it helps you
with, um, sending to other department
because, uh, sometimes what happen
is that the business is, uh, an issue
with the front end application, but
it actually is not a problem of that
application.
So
Denis: Hmm.
Serena: agent, we are tracking.
Where the problems is for
real and not where the user
is experiencing the problem.
So that is helping, actually, it's only.
One month that is in production.
And for now we are seeing
improvement that side.
But yes, we usually take three
months to see, to have results,
Denis: Mm-hmm.
Serena: to see the aggregated
results so that we have a better,
idea of how it is really going.
we, we will see in a couple of months.
Denis: That's an interesting, I'll
definitely reach out to, uh, to
see, um, what the results are.
But I think already it's very powerful
to, let's say now you liberate an
experienced IT professional who had to.
Go and troubleshoot him or herself
by an agent, allowing that person
to do more interesting work, as you
said in the beginning of the video.
Um, one, I think one important, let's say
holy grail of agent workflows is that we
see nowadays, agents can write codes, can
interact with systems, can SSH to servers.
Do you plan to explore the option of
training agents to actually carry out
IT operations themselves in the future?
Serena: Yes.
Uh, not only for the ticket
resolution, what we would like to
do in that department is that, in.
we are having this trial period where
the human gives also feedback, uh,
uh, whenever it close the ticket.
So it tells you what, uh, the
agent did well and what it didn't.
And we made it mandatory.
So
we actually have
Denis: Hmm.
Serena: 'cause sometimes
you, you don't make it
mandatory and
people will just
Denis: I don't have feedback.
Yeah.
Serena: Yes.
And yeah, I mean, you can't
improve without feedback.
So, uh, so yeah, we made it that
mandatory and what we would like
to do is, uh, on a very first
trial, make, uh, some small, action
automated and we are working on, um.
On the data quality aspect.
So whenever we have tickets on the
quality of data, we will like to make
it automated, that is the agent actually
do modify the data in the database.
And not only in the one that you
saw that was a problem, but in
everyone that is, uh, implied.
So for example, we have.
Smart meters.
That is the tool that, um,
count how much gas you use.
It's our main object.
So it's our, our business
is all based on that object.
So you have, uh, this information on SAP.
it's not just there.
It is also in Salesforce.
So if in a business user see a problem
in, uh, smart meter in SAP, it open a
ticket saying, okay, you have to, um,
correct this data, this information.
It's not like this.
It's like that The agent will.
Change the data on the
system that the business.
So, but also in all the other,
so for example, it sos in SAP,
it will also change the agent in
Salesforce so that everything is
coherent and everything is uh, align.
Denis: Mm-hmm.
Can you focus a bit more about, uh,
how you use Databricks because you
mentioned the agent will live there or
interface through Mosaic AI gets it.
Data from Unity Catalog that
is connected via Lake Flow.
Um, with SAP and ServiceNow, but
then you do want to update the
data in the source systems like
ServiceNow and SAP, is that correct?
Serena: yes, yes, exactly.
What happened is that we have,
all the data in Databricks, so we
know where that entity is in the
source,
uh,
Denis: Mm.
Serena: So what the agent will do is
look through the catalogs and the mis
data and look for those information and
make the correction.
Wherever is needed.
So this is the idea, or still
design this agent, but that's
what we would like to do.
Denis: Well, you're already in
a strong position by having all
your data on Databricks, right?
I mean, the client I currently work
for, we are not there yet at all.
So, um, and that then you
have the problem, you have
a lot of different systems.
One check that we can share that worked
quite well in our case was, for example,
it's not an agent yet, but we often, for
example, have to test for connectivity.
We have some remote sites with a
remote server that pumps data into
our central system like Azure.
But then sometimes the connection
breaks, and when you have, let's say
hundreds of sites, it's very hard to
diagnose this manually with the person.
So we wrote a couple of very simple
PowerShell scripts that just test, Hey,
can I connect to the source SCADA system?
Okay, can I connect to Azure?
So on, can I establish
an RDP or SH connection?
And now we basically pump the results of
this script into a tool for monitoring.
But I can imagine something
like that could be very well
done by an agent, right?
Serena: Yes.
So obviously we actually would like
to do something similar for, um,
the cloud infrastructure to see.
If the need to like scale up or
scale down any resource, that's
definitely something an agent could do.
we're in our next phase,
uh, of a AI for a cT.
That's what we would like to focus on.
So, but also, uh, like a
threat to detection for
cybersecurity, we would like to.
To do something about that.
So it's really a lot of, things
that you can do with AI nowadays.
Denis: Yeah, I mean, one of the risks
I, I see nevertheless is that, I mean,
recently you probably saw this article
on LinkedIn bouncing by of an agent,
deleting a production, a table in
production database for the company.
So how would you shield
yourself against these risks?
Right?
Because a human can always think
twice, whereas an agent may
just take the shortest route.
Serena: Uh, yeah, that's true.
Uh, so what we are doing,
well, they usually don't have
the lead, uh, permission.
So this is something,
Denis: Mm-hmm.
Serena: at the same times.
Um, so what you do is, what we
are trying to do is to have, um.
Trial period where you have
humans and agents that coexist.
you try to minimize the risk, uh,
like the super human supervise
the agents until you breach a
certain, um, security performances.
And so then you are, you feel a
bit more confident now, you know.
But at the end of the day, I
mean, these things will have, they
can happen even without agents.
So we, I remember like three years
ago that the person that worked
with us dropped a dp, so happen.
Denis: Yeah, we've all had these things.
I've also, in the beginning of my career,
have written a SQL query without a wear
statement and made a Cartesian join
on the entire database, essentially
freezing it for a very long time.
Um, so those things are
also done by humans.
Serena: yes, in Italian, we say that if
you, never do, you never make mistakes.
So.
Denis: It's true.
Well, speaking of mistakes in human,
have you had any, um, feedback from,
let's say it people or business
people about this new agent support?
What do they think about it?
Serena: Uh, so the ticketing
support is, um, is getting, uh.
It Different supports, sorry, feedbacks.
It depends.
Uh, the IT department are
actually quite happy to have it.
Uh, business users are not so happy
to have, uh, to talk to, uh, an agent
before, uh, compiling the ticket.
So we made it mandatory
to go through the agent.
Before actually open the ticket.
Maybe in the future we will, uh, skip.
We will not make it mandatory, at
the moment we are trying to, to
see, to have a bigger adoption.
So trying to see how it
actually works and, um, to have
as many feedbacks as we can.
Uh, other people are really, really
happy because it helps them and it.
Really, uh, cut the, the
time that they need to.
So to have support.
So sometimes it happen that, they didn't
get, uh, supports for weeks and now it's
like of minutes of talking to an agent.
Denis: Yeah, like we're still exploring
and it's, you're, you're like still
looking for your client base, so to
say, for the people who like it, for
when it's useful, I don't think the goal
is to force it on the entire company.
Right.
You will probably use
it where it makes sense.
Serena: Yes.
Um, even because now it works on
everything, but I think like creating
a new, uh, user or a new password,
it's not really helpful to, to have a.
Conversational agents that
help you write
the
Denis: Yeah, exactly.
Serena: mean it.
Yeah, so it really depends.
It really see, in some cases,
really helps in some other,
doesn't make any difference with.
Denis: Well, speaking of future use cases,
um, and to wrap up this episode, you
mentioned you had several vertical agents
that do all kinds of specialized things.
Can you tell us a bit more about one agent
that particularly excites you or something
that you want to develop in the future?
Serena: Okay.
Uh, so the, the one, uh.
The one that we like, well, I like the
most because it was my first agent.
Uh, it's the one for, um, checking
the document whenever we receive a,
a project for opening a new pipeline
or, uh, to install a smart meter
in your home, have to submit, uh.
Some paperwork.
So you have a project, you have,
uh,
to send
Denis: Sorry, that's a gas
pipeline, right?
Serena: things.
Yes.
Gas pipelines or, uh, smart meters or
if you have to like install the, the
tools to measure the gas in your house.
Those are all, um.
Something that was done by hand.
So there were people receiving all
these documents and checking by
hand if everything was correct.
What we did, and it was
really our first agent, uh, we
automated this pro, this process.
So basically the customer sent all the
documentation and we scan it, we read it.
Through the agent and, uh, we
check what is correct and what is
missing, and if it is, if something
is missing or is not correct.
We send an email through the agent
saying, these are the things that
Denis: hmm.
Serena: or you have to do this and that.
And that.
And then when they submit
the new paperwork, we receive
it, we check it again.
If everything is uh correct, we send
it through, um, the next step of the
process and we like give the approval.
So even here had like six months
of, uh, joint work with, uh, with
the agent and, um, and the humans.
Denis: Mm-hmm.
Serena: But it was really a
tedious work to do because it was
really no, no value in the job.
People were doing these things as
a, uh, they were looking at this
project process as a, a wasting of
their time because, I mean, they
didn't do anything really exciting.
This was one of the, it, it was a really
cool, uh, project, a really cool agent
because it actually helped people and the
company and, uh, and it was really useful.
Uh, and it was my first one there.
We, we learned that feedbacks are
really important and if more, and
most of the people don't give you
feedback, so you have to ask them.
Denis: Yeah, it seems that cooperation
between humans and machines is not yet,
um, will always be a difficult point.
So.
You mentioned that uh, learning the
importance of getting human feedback
is important given on your experience
with now developing several agents and
having 10 of them in production, is
it something that you would like to
share with the listeners who may also
be on Databricks or try to deploy some
agents, some things that you learned
and what you can give us advice.
Serena: Yes, of course.
Um, yeah, so again, uh,
having a feedback is like.
The base of your, of your work.
the cool things that you can do now
is to have even, uh, personalized
metrics that you can, um, design in
Databricks with, uh, AI Mosa and LLM
as a judge with, uh, all the MLA ops.
So I think, they are doing a
great work in the observability,
and I think observability is, uh.
It's a really important aspect that
sometimes people feel it's boring,
but it really gives you great insight.
Not only in the use of the agent,
so how on the adoption on the
performances, but also on the RESO
resources that you, you are using.
sometimes maybe you
underestimated or overestimated.
So observability is really important.
it will give you better idea of.
What your agent is doing, who is using
your agent, uh, how is performing?
And also if you it correctly,
it will give a better, um,
user experience to your users.
So I think it's a underestimated topic,
observability, but I think it's really.
Really important and really cool.
Uh, the other thing is, to not focus only
on your current problem and do something
that is not reusable, but, uh, to look
at the bigger picture and trying to
build reusable pipeline, reusable code.
agent.
I always like to, to say that for
me, agents are like microservices.
So basically agent adjunct architecture
is like, microservices, architecture.
have to really focus when you are,
uh, coding, when you're designing
an agent on what makes sense to
be an agent and what makes sense.
To be just a step on your agent,
because otherwise you end up having
a lot of small agents that are never,
um, call again, accepting that chain.
And, um, it'll licensees or will add,
uh, resources, consumer of resources.
So it'll not make a good experience
neither for the business,
neither for the IT department.
So.
I think it, I think more and more
you have to focus on really the
architecture and the big picture and,
uh, what are the idea behind this agent.
It's not just coding, it's not going
Uh, through the, through the problem
and trying to solve it, but you really
have to take a step back and see what
are the best action, what is the best,
um, strategy to solve this problem.
Denis: I think these are great points
and I can add one point of my own.
If you want to implement the point.
Serena just said, I think it's
very important to take your digital
transformation seriously and by
that really making Databricks
or whatever platform you have.
Um, I personally prefer Databricks.
Really the center of your data
infrastructure, have all data there.
Uh, with Mosaic ai because if you
still have your old systems and custom
solutions and then want to implement
agents, you will lose too much effort
with data integration and observability.
And I believe Databricks just makes
these things like, um, monitoring
connection, but also the metrics from
your agents all, all neatly in one place.
Right.
Serena: Yes, exactly.
Uh, when we moved everything on
Databricks, our, uh, ability and
our, readiness to move, uh, POC to
production ready solution increased
exponentially, we, really am.
Almost double, um, the activities, the
project that we deliver in one year.
And be fair, we actually
cut in half the costs,
Denis: That's fascinating
because I keep hearing that
all data picture was expensive.
Your cost will explode, but you
had the opposite experience then.
Serena: Yes, we, we had
the opposite experience.
To be honest, I think it's, uh,
maybe not so, uh, intuitive, like to
optimize and, uh, to, to use it well.
So that's the problem.
So it's really easy to spend
a lot of money, if you.
Again, take your time to really
understand what you have, what
you want to do, and how to do it.
I think it is really incredible.
It's super,
Super fast.
You get all your data in a
blink of an eye, uh, it process
data in a very, uh, low time.
You, you have all these
tools already embedded.
Everything is there.
You don't need, so many
tools, so many licenses.
It's everything.
Open sources.
So you own your data, you own
your code, your process, and.
Actually you don't spend that much money,
but of course everything in the cloud,
it's really easy to spend a lot of money
if you forget to like stop a cluster or
you, uh, didn't really set it properly.
Denis: Oh yeah.
I've been there with the surprise emails
At the end of the month when you have
a large bill, maybe you can develop
an agent one day, Serena, that will
like watch over your Databricks costs.
Who knows?
Serena: Yeah, I think I, I mean, with
all the observability, uh, a system table
now with the observability data that you
have, you can actually do it quite easily
and using agent Bricks, that's the name
of the new, the new tool from Databricks.
It's really, it could be really easy.
The only problem I think with
databricks is that they released many
new feature in so little time that
it's really hard to keep up with them.
Denis: Yeah, it's true and often people
who the clients are not always aware
of what Databricks can do, right?
Serena: Yes,
Denis: seems to be a lack of
expertise in this industry.
Serena: Yeah, even the A IBI, uh,
work from Databricks, it changed so
much in so little time that most of
the people don't really know that
they can, can do so many things and
Databricks even about data visualization.
Denis: Yeah, well hopefully after
this podcast episode, they will have
some new ideas of all the things
that are possible with agents.
Um, in that sense, I think we had a very
good discussion of all the things you,
uh, plan to do and what you've already
done, is there anything you would like
to add at the end of this podcast?
Uh, Serena.
Serena: Um, not really.
I mean, we really have a lot
of, uh, things in mind that
we would like to do and try.
So the next, uh, months will be.
Uh, interesting.
And, uh, we will be able to
deploy more agents to production.
Um, yeah, it was really
nice to talk with you.
Denis: Yeah.
Thank you for coming.
I had a great conversation and who
knows, maybe like in half a year
we'll talk again about maybe now
you have like 100 agents running
in production, your entire company.
I dunno.
Thank you all Serena for coming.
And thank you also for the
listener for listening to a new
episode of What's In Your Data.
And we'll see you in the next one.
Bye-bye.