Latest / AI for Founders with Ryan Estes / The AI Analyst That Never Sleeps: Burak Karakan of Bruin

Transcript

The AI Analyst That Never Sleeps: Burak Karakan of Bruin · AI for Founders with Ryan Estes

0:00
Your data is suffering and your company is the last to know.
0:03
Most founders are out there hiring analysts, building
0:05
dashboards and wiring together 8 different tools and the whole
0:08
time 80% of the answers they need are already sitting inside
0:11
the data they already own. They just can't get to it.
0:14
Every team in your company is spitting up their own little AI
0:17
agent, pulling from different sources, getting different
0:20
answers. Your sales numbers don't match
0:22
marketing. Marketing doesn't match OPS.
0:25
This is a data chaos problem, and it's getting worse by the
0:27
week. It's bleeding you dry.
0:29
And the scariest part isn't what you can see.
0:31
It's what your data already knows that you don't.
0:34
Your data knows your capacity is about to collapse.
0:37
It knows where you're bleeding before you feel the cut.
0:39
So if your company's data could talk to you every morning before
0:42
you made a single decision, what's the first thing I would
0:45
tell you that you've been too afraid to ask?
0:47
Welcome to AI for Founders. I'm Ryan Estes, and today's
0:50
guest is Barack Karakan, Co founder and CEO of Bruin.
0:54
If you're a founder who's tired of your data, living in a dozen
0:56
different places and never telling you anything useful,
0:59
this one's going to hit different.
1:00
And the AI for foundersnewsletter@aiforfounders.co
1:03
is where we break down exactly how founders like you can turn
1:06
tools like this into unfair advantages.
1:10
It lands in your inbox, not in your spam folder, because we
1:12
actually write it and it's not spam.
1:14
Head to AI for founders.co and get in before everyone else
1:17
figures out what you're about to know.
1:20
And if this episode made you think the kindest thing you can
1:23
do is leave a five star review because great guests read those
1:26
before they say yes. And right now I want to tell you
1:28
about Kick Caster. If you're a founder, an
1:30
executive, or an expert who wants to get booked on top
1:33
podcast, Kick Caster is the agency that makes it happen.
1:36
They have the relationships, the strategy, and the track record
1:39
to get your story in front of the audiences that matter most.
1:42
Stop waiting to be discovered and go to kickcaster.com to
1:45
start getting booked. That's Kick Caster by motorist.
1:52
Welcome to the show. Thank you so much for joining
1:53
me. Hi, Ryan.
1:55
Thanks so much for having me. You betcha.
1:56
Where you calling from today? I am currently in Istanbul,
1:59
Turkey. Istanbul.
2:01
OK, cool. And are you HQ D out of Berlin?
2:05
Well, it's, it's a, it's a mixed situation at this point.
2:08
We are like a pretty much a remote company.
2:10
We have a lot of folks in Istanbul, some people in Berlin,
2:14
some people in UK, and then let's spread across the globe a
2:17
couple of remote folks as well. Cool.
2:19
Tell me. Let's start with Istanbul.
2:21
Tell me about the tech community in in Istanbul.
2:24
Well, very interesting place Istanbul is, you know, it's,
2:26
it's quite large. It's it's very international.
2:29
There's a lot of, there's a lot of builders, there's a, there's
2:33
a huge mobile gaming culture here.
2:37
There's like a couple of Unicorn companies, gaming companies like
2:41
that came out of Istanbul. So we can, we can find a lot of
2:44
gaming folks here and, and quite a lot of like strong builders.
2:50
I think Turkey's in an interesting place where it's,
2:55
it's, it's very close to Europe in terms of time zones.
2:57
So it makes it very easy to, to work across the, across the
3:00
whole European market and the, and the American market as well
3:03
and has some pretty good schools and stuff.
3:05
So, I mean, I, I was away for, for quite a few years and now
3:09
coming back, I feel nice to be back.
3:12
You know, it's nice to have some like, excited, driven folks
3:16
around me that just like, you know, burning with desire to
3:18
make it so. Love it.
3:21
And it seems like that's the theme, you know, it's, you talk
3:23
about historical cities. That's got to be #1 right?
3:27
It's up there with Rome. Absolutely.
3:30
Yeah. It's crazy.
3:30
Like, it's almost crazy. Like you're just going to walk
3:32
around and you're going to see like wolves that are like
3:34
hundreds of years old and people are just like having some tea
3:36
next to it. You know, It's, it's quite big
3:39
as well. Like the, the core Istanbul is
3:42
quite central and, and relatively small, but he can
3:45
smell the, the history in the, in the city, which is, which is
3:49
very nice. You know, it's, it's definitely
3:51
like a well recommended visit. You're gonna have like some,
3:54
there's some great food around here, very multicultural.
3:57
So, you know, if you're ever like making a stop by here, you
4:00
should spend a couple days. I would love to.
4:03
Yeah, it would. It definitely seems like that.
4:05
It's that kind of place. Historically, it's a a place for
4:08
ideas to coalesce, for cultures to kind of merge together.
4:12
It's always been that, right? Absolutely, yeah.
4:14
There's like, like it's, it's, it's a very interesting mixture
4:18
of like Eastern culture and the Western culture sort of like
4:21
putting them in an hole in a single bowl and and mixing them
4:23
up. And there's like a lot of
4:25
synthesis that can't, that comes out of that.
4:28
It's, it's quite unique in that sentence.
4:31
You know, it's not like I, I lived in Europe for like, but
4:34
more than six years. It's a bit more, a lot more
4:39
heterogeneous than than that because it mixes a lot of
4:42
different, different cultures. And that's, that's what makes it
4:45
beautiful. So because of where you are,
4:47
you're probably going to find like a nice corner of the city
4:50
that would, that would feel like home, you know, or that you will
4:52
make it home for yourself. So I really like that.
4:57
Had a really random question for you.
4:59
So a lifetime ago, I was the music director at a famous jazz
5:02
club here in Denver and one of the the kind of regulars was a
5:08
guy in the witness protection program.
5:09
He was an Israeli gangster and turned state's evidence and
5:13
ended up in Denver and would hang out one of the coolest guys
5:16
I've ever met in my life. He was awesome.
5:17
But he would, he always would say he's like Ryan, you look
5:20
Turkish, you look Turkish. Do I look Turkish?
5:24
You kind of do, actually. You, you, you definitely do.
5:27
I mean, look, the glasses put you in a different mood, I'll
5:32
give you that. So.
5:33
Like here we go. Absolutely, man, absolutely.
5:36
I look Turkish. You you look very much Turkish.
5:39
Yeah, you can easily pass this as Turkish Easily.
5:42
Easily. I didn't notice that actually
5:45
before. It's it's very true.
5:47
I wonder what that's about. Interesting.
5:49
He was also convinced I was a boxer.
5:50
He was. He's like, you have the puffy
5:52
eyes of a boxer. You're a Turkish boxer.
5:54
I'm like, all right, man. Nah, you could be, Yeah, Yeah,
5:56
you could be. You could.
5:57
You seem like a dangerous guy, man.
6:01
I'm glad we got that out of the way.
6:03
Thank you for clearing up this 20 year old mystery.
6:05
And I, I guess it's true. I'm to go back to 23andMe and
6:08
see what's going on. I'm glad to help.
6:11
Yeah, glad to help. But let me know the results.
6:14
Yeah, I'm quite curious now. Yeah.
6:18
Amazing. OK, let's dive in.
6:21
I want to learn all about your company, Brewin, because it's
6:25
it's very interesting and it's I, I love the design of the
6:28
website since I've seen it last. It looks very clear, very clean.
6:32
It says say hello to your AI data analyst.
6:35
So give us an idea of Brewin and the website isgetbrewin.com kind
6:39
of what you do and who your ideal customer is.
6:43
We are building in an AI data analyst.
6:46
Basically it is it is an AI agent that you can connect it to
6:49
your data warehouse, you can connect it to your different
6:53
data sources. You can ask any question in
6:56
Slack, in teams in our web UI and get an answer within 90
7:00
seconds. The the core idea behind that.
7:04
So I'll, I'll take a step back. We worked both my Co founder and
7:08
I, we worked with a lot of data before we worked in big
7:11
companies, helped a lot of small companies to, to build their
7:14
infrastructure as well. And essentially we started off
7:18
our journey trying to build a, an easy to use and manage data
7:22
platform for humans. And that ended up evolving into
7:28
into, but, but Bruin is today. And with AI agents becoming a
7:32
thing, we eventually said, OK, well, if we made humans
7:35
productive, there's a very good chance that we, we would be able
7:38
to make AI agents productive as well.
7:40
That's how our AI analyst was born.
7:44
The core sort of benefit is what normally would take hours, if
7:48
not days, now takes minutes and you can ask any question, you
7:52
can get reliable answers out of that.
7:55
It will actually use the data that you have inside your
7:59
database and all of all the actions your agent takes are
8:03
verifiable or they're like they traced.
8:06
So you can, you can go and validate every step and suddenly
8:10
the time to insight starts becoming a non issue.
8:17
Basically, we work with, we work with companies that are smaller
8:24
than 500 people mostly. We have like some, some bigger
8:27
and multinational customers as well.
8:29
But we tend to work well with companies that are obsessed
8:31
with, with impact, you know, and the companies that, that have
8:35
some questions really. So, I mean, there's a lot of
8:37
organizations out there that that do not really have any
8:39
questions to ask. We, we, we prefer to work with
8:43
those that actually have questions that they can ask and
8:45
if they get the answers reliably, they can act on that.
8:48
That tends to be sort of like younger more agile organizations
8:52
and and also forward thinking teams that are not afraid of AI
8:57
well. You know, and kind of where
8:59
we're at right now is like, you know, kind of like, you know,
9:03
companies that want to run, they're really encouraging their
9:05
teams to to use AI tools. And one of the problems I see is
9:10
you have everyone who is close coding agents with clawed code
9:14
and put in their particular silo.
9:16
So they what data they're pulling from is different from
9:19
the agent that their other team is pulling from.
9:22
So is one way to think about this is this is kind of like a
9:25
giant MCP where like all of the company's data can be stored in
9:30
one place, kind of have one centralized agent to rule them
9:33
all kind of thing? That's a very nice way to put
9:35
it. The only like little addition I
9:38
would make is it's not us that stores the data.
9:41
The data still lives in your data warehouse, your database,
9:44
and we make this data accessible to all, to the rest of your
9:48
organization. Essentially, if you have a 24/7
9:52
data agent that's reliable, that can get you accurate answers in,
9:57
in a minute living in your Slack account, that allows a lot of
10:02
people in the company to start now being able to ask questions
10:06
and get answers and then use those answers to, to, to change
10:09
their decisions, to iterate on the business, right.
10:12
And, and the fact that this is centralized means that now you
10:16
have full visibility into how people in the company are
10:20
interacting with the data. If there are any common pain
10:23
points, if there are common pitfalls, they might be, they
10:27
might be making mistakes when they're looking into the data
10:30
And, and this in addition makes data governance much easier.
10:34
So you know who's accessing what and, and how, etcetera.
10:39
If you compare this with basically everyone trying to
10:41
build their own versions internally inside the
10:43
organization, now you have suddenly like 15 different
10:45
products competing internally, everyone doing it their own way.
10:49
Some of them have strengths over the other ones, but you
10:51
basically lose the plot because everyone is, you know, pulling
10:54
in a different direction, you know, so, so Bruin allows you to
10:58
eliminate that. Plus it's like you get the on
11:01
board intakes about like 2 minutes.
11:03
You can ask your first question in about in about two minutes
11:06
and and if that excites you then Bruin is Bruin Might be a a a
11:13
good idea to try it out. Yeah, it sounds amazing because
11:16
otherwise if we're like, you know, the, the guy in the office
11:20
next to me is like, no, we can do this.
11:21
I just need, you know, access to absolutely everything.
11:24
And now it's got my inbox credentials and my Google Docs
11:29
and they've got my Slack history and all this stuff where it's
11:32
like, no, no, I need to make sure this doesn't just leak into
11:36
God knows where, you know, the open Internet, you know, so, and
11:41
people are definitely doing that.
11:42
Like right now is they're, they're trying to find a
11:44
centralized place. So, so this becomes kind of like
11:47
the, the central protected home. And I'm assuming there's access
11:51
levels so that you can like see some people have visibility into
11:55
these things. Some people don't have
11:56
visibility to these things. These things always stay on our
12:00
machine. They're never connected.
12:03
Am I safe to assume that? Absolutely.
12:06
So we allow it. We allow teams to control
12:09
basically who can access what, who can ask what types of
12:13
questions that can be answered with what data that they have
12:15
available. Have very granular control over
12:20
all of these different agents and their permissions across
12:23
different even Slack channels. So you can do things like, hey,
12:26
I connect this AI agent to a marketing channel.
12:28
Marketing people should only have access to this and this and
12:31
that data, right? Whereas I have another executive
12:35
channel that will need to have a lot of access to, to the
12:39
sensitive internal data as well, right?
12:41
And these are living in completely different places.
12:44
So the organization admins have full control and visibility over
12:48
who can access what. And they can even define
12:51
completely separate credentials, completely separate access
12:54
rights. Well, we're building, we're
12:57
building something that we wish we would use for our own company
13:00
as well. And that's what we're doing as
13:03
well. So yes, you can control a lot of
13:06
these different aspects of access control.
13:09
We also offer different deployment models.
13:11
So you can, you can use our cloud offering as like a managed
13:15
product. We have like a hybrid model
13:17
where we can deploy part of our data plane, part of our product
13:20
into your environment. So it does not have any access
13:23
to sensitive data. We also allow deploying in full
13:28
on premise environments where if the organization has their own
13:31
AI models, if they host their own large language models
13:35
themselves, we're able to make our agent work with that too.
13:38
So depending on your appetite, we can meet you there.
13:44
This episode is brought to you by a dangerous idea.
13:47
It's called reading, specifically the AI for Founders
13:50
newsletter. 47,000 Founders strong now.
13:53
And it's the good stuff. No top ten AI tools that will
13:57
change your life written by someone who doesn't ship
13:59
product. It's just real breakdowns of
14:01
what's actually working in AI, who's raised, who failed, who
14:05
pivoted, and which founders quietly built $1,000,000 in ARR
14:09
while everyone else was arguing on X.
14:12
It's written for people building with AI every week.
14:14
You get tactical frameworks, unfair advantages, breakdowns of
14:17
real companies, and the occasional existential question
14:20
about whether your startup is intelligent or just, well,
14:23
branded. It's free, It's sharp.
14:25
It might mildly offend your current workflow.
14:27
Go to aiforfounders.co and subscribe before your competitor
14:30
does, because the only thing worse than being replaced by AI
14:34
is being replaced by a founder who reads.
14:37
That's the newsletter at AI for founders.co.
14:39
Now back to the show. It's incredible.
14:41
You know, I had a conversation with three agency owners last
14:45
week that they were talking about like, oh, we're going to
14:48
make giant MCPS and integrations into all of our clients into
14:52
specific things. And I was like, bro, Are you
14:54
sure you want that liability? Like you talk about data
14:58
governance. If there's some kind of like
14:59
misfire or slip a whoops dude, that's, that is litigation
15:05
waiting to happen, you know? Yeah, absolutely.
15:08
Yeah. So, I mean, maybe you should
15:10
introduce them to me, yeah. No, but I know from now on I'm
15:13
gonna be like, hey, Bruin is your first stop.
15:16
And speaking about that too. You know, we're we're starting
15:18
to introduce like our AI Co workers and kind of thinking of
15:21
them as an org chart, largely because it makes sense to our
15:25
brains. So, you know, if we're looking
15:27
at the org chart like Bruin would be, which this is like, is
15:31
this like data compliance officer?
15:34
And how would I kind of use it within?
15:36
Like if I'm spinning up other agents, if I'm spinning up a
15:39
research and development kind of agent or I'm spinning up a
15:42
marketing reporting agent, Is there kind of channels that you
15:45
could say like, you know, look at our projected pipeline and
15:50
give me a outlook of where most prospects are coming from this
15:55
kind of thing? And it would be able to kind of
15:57
pull from the tools we're using, but also make projections based
16:00
on kind of like where, where it sees we're going.
16:04
Absolutely. So the way I imagine Bruin and,
16:08
and sort of like these sorts of products evolving in the future
16:10
is we're going to have a lot of different AI agents working in
16:13
different domains. And those agents will need to
16:15
collaborate just as humans do as well.
16:17
They're just going to use different means of
16:19
collaborating. And, and in that world, Bruin
16:22
essentially serves as your virtual data team, right?
16:25
So any agent or a human being in your organization that needs to
16:29
have access to data that has any questions about data that wants
16:32
to get some forecasts, predictions, past reports, all
16:35
or simply asking like what data do we have available?
16:38
Do we already pull this data? Do we already have this data
16:41
accessible? They can now suddenly ask Bruin
16:44
about all of those and get answers in seconds.
16:48
This is why we connect to Slack for humans, we connect to teams
16:53
for humans. We have a web UI, but we also
16:55
expose our full capabilities both as like a raw API as well
17:00
as an MCP server. So if you have other agents that
17:04
now want to programmatically interact with our agent, they
17:07
can do so, right? So we're already seeing like
17:11
early signs of this where teams are starting to spin up, spin up
17:14
their own marketing agents where before they can make informed
17:20
decisions, they need access to past data.
17:22
And now they're suddenly connecting to our AP is to, to
17:25
ask AI generated questions and then get the answers back and
17:29
feed them back into AI. Another very interesting
17:32
scenario that that that I haven't thought of before is one
17:36
of our customers is, is dealing with a lot of inbound support
17:38
tickets, right? And, and many of their support
17:42
tickets requires the, the support agents to actually go
17:45
and validate if the customer has purchased a certain thing, if
17:49
they have actually have seen what they claim to have seen, if
17:52
they've done what they claim to have done.
17:54
This is a lot of data work. So even though they had their
17:57
support agent, they had to build a lot of like data capabilities
18:01
into their agent, which obviously it's, it's, it's like
18:04
it's, it's a profession of its own.
18:06
You know, So what they do now is they basically connected their
18:08
support agent to our AI analyst. And then every incoming support
18:14
ticket automatically gets routed to our agent, which then pulls
18:18
data, feeds it back into the support agent, which now can
18:21
generate over 95% of the support answers.
18:26
And it knows if the, if they, what customers claims are
18:28
accurate, it knows what the data is, it knows how much the
18:31
customer paid for a certain product, what they have done,
18:33
what they have seen, etcetera. All validated, which would
18:37
normally take like 2 to 3 hours per user, now takes around like
18:41
40 seconds. You know, and suddenly you start
18:45
seeing different constraints. Your Your world has changed.
18:49
Right. You know, and like, because
18:51
pulling reports is laborious and there's always constraints on
18:55
what you can get within what dashboard, and you got to pull
18:58
it out of here and look for that.
18:59
And it never really gives you exactly what you want, you know.
19:03
So for example, what I want is to look at results from my sales
19:06
team and say, go through the Google, Google calendars of
19:09
every single discovery call, refer that to, you know, who's
19:13
closed and analyze the, let's say all of the sales call
19:19
information to see what discrepancies exist for a
19:22
closing sales call and a non closing sales call.
19:26
Things like that, like deep curiosity into like the the
19:30
processes that are getting kind of maybe lost in the shuffle.
19:34
Yeah, I mean, I think like the the reports is like the first
19:38
obvious example of of this. It's very, very slow and and and
19:42
painful. There's a lot of like you can
19:44
either build like your custom agents that that he can do that
19:47
you try to fit into your scenarios, which sometimes might
19:50
be a good fit or there's product like ours.
19:53
But I think that's only the beginning.
19:55
The more interesting part for me is OK, you take care of like
19:58
doing reporting for retrospectively.
20:00
So you look at what happened in the past now.
20:04
Can we make predictions about what's gonna happen in the
20:07
future based on the data that we already have available?
20:10
And can we take actions based on those predictions to either make
20:14
them better or to prevent them from getting worse?
20:17
Right. So for instance, if you notice
20:18
that you have some team members that have some upcoming
20:24
vacations for a couple of weeks and suddenly like 2 weeks down
20:28
the line, your capacity is going to get worse.
20:30
And then you also predict that your incoming calls that are
20:33
scheduled per salesperson is going to increase over the next
20:36
couple of weeks. Now you're suddenly looking at
20:37
the capacity problem where you're going to lose a lot of
20:39
deals in the process. Can agents proactively inform
20:43
you about this upcoming problem? They notice and then and then
20:47
help you take action to begin with and then at some point
20:50
eventually being able to take action by themselves, such as, I
20:54
don't know, maybe hiring on demand sales people, maybe
20:57
hiring AI sales agents for you down the line, you know, or
21:02
maybe helping you manage your spend.
21:03
Even if you're doing like, let's say, I don't know, paid
21:05
marketing. And then you can say like, hey,
21:07
I mean, doesn't matter if you can pull these people in.
21:09
We don't have the capacity to serve them.
21:11
You know, we might as well not just not burn the money.
21:14
So yes, the reports is, is very it's saving a lot of time today.
21:20
But what excites me more is, is the future perspective, like a
21:24
lot of what we do is going to change, you know, not, not
21:27
because of us. There's like this, like the ways
21:30
of doing business is completely changing, right?
21:32
And, and, and that excites me a lot more.
21:36
And data sits at the core of that because if you have clean
21:40
data, if you have data that is accessible, that is accurate,
21:44
that is easy to understand, now suddenly you have a map from the
21:49
past that tells your future, right?
21:52
And based on the future, you can take a lot of actions that will,
21:55
that will result in a lot of improvements in, in how we do
21:58
business. And I'm very excited to be
22:00
sitting in the core if that's that domain.
22:04
It's amazing, it really is. And it's the, the kind of shift
22:07
that I'm trying to make mentally is that like, OK, it's smarter
22:11
than me now and now IQ is free. So now what is my, my job?
22:16
It's kind of discernment and taste, you know, So what I, I
22:20
hear you describing the product immediately, I'm thinking like,
22:22
OK, I want to say, you know, everybody that churned in 2025,
22:27
enlist them into a re engagement sequence on instantly and use
22:31
the sequence that's re engagement instantly too.
22:35
And just like start creating, having ideas.
22:38
But what you're talking about is probably even more valuable,
22:40
which is like having them analyze and see the holes that
22:44
I'm blind to, not outside of the scope of my ideas.
22:47
It's like I'm thinking sale, sale, sale, sale, sell, sell,
22:50
sell. But they're like, no, you're
22:51
bleeding over here though. Like let's patch up where you're
22:54
bleeding and then we can go get into your ideas.
22:57
So how, how do you build this kind of into a product where
23:01
everything is kind of reactive at this point?
23:03
Still, it's a chat bot or we're in Slack.
23:06
So it's it's waiting for me to engage.
23:08
So is that something you encourage just to say like, Hey
23:10
Bruin, what are my blind spots? 100% So I, that's one of my,
23:16
that's, that's one of my favorite questions that I asked
23:18
during onboarding. So I'm like, hey, what can I ask
23:21
you? And, and it, it will come out
23:23
with like a bunch of answers based on the data that, that it
23:27
seems irrelevant. And then I'll, I'll start
23:30
digging deeper into some of those vague ones just to be able
23:34
to have that mental exercise of, OK, there's like a, a thinking
23:37
being on the other side of the table here that allows us to,
23:41
you know, play together and, and, and dig deeper.
23:44
I think we are beyond the point of just treating these as, as
23:50
parroting sort of predictive machines.
23:54
It's, it's a lot more, it's a lot closer to intelligence.
23:57
It's, it's a lot smarter than me.
23:58
And and this, this, if we acknowledge and accept that fact
24:03
and make it a part of, if we embrace that into our workflows,
24:06
suddenly we start thinking about how do we enable them to make
24:11
the most of what we can do with our data.
24:14
In the example that you gave, you can still orchestrate these
24:19
different, different agents to take care of some of these, some
24:23
of the ideas that you might have, you give some hypothesis,
24:25
etcetera. But what's what, what I find
24:27
fascinating is now we have these machines that can do the same
24:31
exercise that you can, but thousand times more efficiently
24:36
than than a human being can in parallel can test out a ton of
24:40
ideas in parallel, observe the results, see what works, see
24:44
what doesn't, and then keep it to rating from there.
24:47
Right. And this is obviously happening
24:49
in, in software engineering already.
24:51
You know, there's a lot of like Carpathy recently has published
24:54
like a very interesting project called auto research, where
24:56
basically you give an AI agent a, a, a criteria to optimize for
25:01
a, a ways to validate it's output.
25:04
And then you let it go crazy, you know, and it constantly
25:06
keeps iterating and improving. Imagine being able to do this
25:10
with every part of your company. You know, you do have all of
25:14
that data as you do as, as you take actions, you're generating
25:18
even more data, right? So suddenly what a a marketing
25:22
person has to do in terms of optimizing their campaigns, you
25:25
might now have hundreds of agents, agents doing this in
25:28
parallel. You know, all of them are
25:30
generating data that that flows into your organization.
25:33
Now suddenly that data can be used by your operations team, by
25:37
your logistics team, It can be used by your product team, it
25:40
can be used by your finance, finance team.
25:43
You know, all of these suddenly start becoming possible where
25:47
we're multiplying our efforts in a much, much larger scale,
25:52
right? The problem with that is if you
25:55
let all of these go wild without any way to validate their
25:59
output, without any way to sort of gathering all that data, now
26:03
suddenly you're looking at like a, a Wild West, right?
26:07
And, and that needs to be that if, if you make data an
26:11
afterthought when you're designing systems like this,
26:14
that wireless comes very quickly.
26:16
You know, so that is what sort of that's, that was one of the
26:22
things that that that kept us building in this space.
26:24
You know, it's like it's an integral part of getting more
26:27
efficient. There's no way that we're going
26:29
to be able to make these agents work without giving them access
26:33
to data. Can we give them access to data
26:36
in a safe and reliable and an accessible way?
26:38
Yes. Then suddenly we unlock a lot of
26:41
different use cases. We love it.
26:43
And let's kind of hover here for a minute.
26:45
You know, there used to be this commercial in the 90s that would
26:48
play and it would say, do you know where your children are?
26:53
And it was supposed to like, I don't know, awake something and
26:55
some parents that didn't know where their kids were or
26:57
something. But you could kind of do the
26:59
same thing now, which is like, do you know where your agents
27:02
are? Because because if you, you
27:05
know, you got 2 open claw instances and they're just
27:09
running on loops, it's like you might not, you know, you
27:12
probably don't. So with Bruin, how do you make
27:15
sure you know where where Bruin is and what or what not they're
27:21
he's doing? Yeah.
27:22
So basically it all starts from like a solid data foundations as
27:28
part of the onboarding, you get a bunch of quality checks that
27:31
allow you to say, OK, is this data good enough for Bruin to to
27:35
operate on and and does this allow Bruin to to, to generate
27:41
meaningful analysis based on the data and metadata that you have
27:44
given us, right. So this already allows you to
27:49
have tight control over what the agent can and cannot do.
27:52
Now this has to be combined with strict access control
27:58
principles. In the end, if you give an, an
28:04
AI agent today, like a full admin capabilities, full access
28:09
to every piece of data, full access to do whatever you want.
28:13
And within your organization, there's probably like some
28:16
problems waiting for you to, to happen, regardless if it's
28:18
brewing or not. You know, like, let's be, let's
28:20
be real. But if you if you combine our
28:25
granular agent capabilities with strict access control on
28:30
different platform side so that for every agent you give exact
28:34
access to to exactly what it needs.
28:37
This already allows you to let it go wild because there's not
28:41
much it can do to go wild. You know, like if it only has
28:45
read only access to two of your marketing tables inside your
28:49
snowflake data warehouse, you know, how much harm could it do?
28:53
The worst it can do obviously is that it can generate inaccurate
28:57
answers and then now your team suddenly can use that answers to
29:00
make the make inaccurate decisions.
29:03
But in order to prevent that, Bruno already includes a bunch
29:06
of additional checks that an output of an agent is being
29:09
controlled by another agent. Your team gets notifications on
29:13
a regular basis in terms of what's happening, where for
29:16
every analysis, every answer you get, you get a link that takes
29:20
you to of individual agent sessions that allows you to like
29:25
exactly see what the agent do and how did it get to that
29:28
answer. Right.
29:29
So suddenly you have eliminated a ton of different types of
29:34
problems already and now you can let your agents move a lot
29:38
freer. Yeah, I guess like that would be
29:42
the that would be the core of, of, of how you know where Bruin
29:45
is and, and, and it doesn't it doesn't go wild.
29:48
We do offer additional controls such as like can Burn have
29:52
access to to these external network sources?
29:55
Is it allowed to access the Internet at all?
29:58
Is it allowed to do this and this and that or can it throw
30:01
that kind of queries? Can it run a query that will
30:03
cost me more than $2.00 at a time?
30:05
You know, a lot like very granular control we allow you to
30:08
do, but depends on your risk appetite basically.
30:13
Yeah, 100%. I mean, it's almost worth the
30:15
price of admission just so you don't have to log in with two
30:19
factor authentication to 17 platforms when you're asking a
30:22
couple of questions. You know, I mean, it's, it's,
30:26
I'm really excited about the, you know, the future of, of
30:29
agentic AI because it's just becomes APIs.
30:31
And so you find, OK, this tool specialized in what it's doing
30:34
really well, But like, I'm not trying to look all over your
30:37
platform to solve my problems. You know, I, I'm really excited
30:41
about agents like this, largely because I know there's features
30:44
and stuff I'm not using in my software that I just don't know
30:46
about it, you know, so if you could just ask it straight up,
30:49
like, Hey, what, what should I be doing?
30:51
It's like, oh, well, you actually know you're in the pro
30:54
plan of all this stuff. You could try this SAB like oh
30:57
OK maybe I will do that. Yeah.
31:00
I mean one of the one of the fascinating part of, of of this
31:04
business has been to, you know, suddenly shift our thinking from
31:08
not just building for humans, but also building for for
31:11
agents. Because precisely like, in order
31:14
to allow agents to do that kind of investigation or that kind of
31:19
like reasoning or what, what you have available, the tools need
31:22
to fit how they use. We have internally built some
31:25
tooling that allows us to get feedback from different types of
31:28
agents based on how our APIs work.
31:31
You know, like if we build an internal like a first version of
31:35
like part of our command line tool, we suddenly run a bunch of
31:38
agents on them as if they're human beings and ask them to
31:41
give us feedback. You know, so and and we run some
31:45
benchmarks to see which one of those were they able to use more
31:48
efficiently and and certain types of tasks right now.
31:52
What was? Feedback from the agents like
31:54
how do they want you to present data and information to them?
31:57
They really dislike when it's multiple steps.
31:59
So they usually try to get things into like a, a single
32:03
command, single space, more efficient interaction model,
32:07
which is not trivial for humans because if you try to build like
32:11
a single command that encompasses like a lot of
32:13
different functionality. Now you suddenly build this
32:15
giant command line tool that has like, I don't know, 200 options,
32:20
which would be incomprehensible for humans.
32:23
But you're not building for humans anymore.
32:26
You're building for agents. Agents are great at generating
32:29
those commands, which means they can surgically generate commands
32:33
that would get only the information they want and they
32:35
need. And then what would normally
32:38
take, I don't know, 200 seconds to finish now starts finishing
32:41
in 30 seconds, you know, because like agent was able to
32:44
construct, create a like a lot more efficient layer to
32:49
communicate with your, with your back end APIs.
32:52
They really dislike when we present them human like
32:57
interfaces when it comes to like command line tools and stuff,
33:00
you know? So it's what what my team likes
33:03
is not always what's what my agents like.
33:07
Interesting. What are what are some of the
33:09
human things it doesn't like? I mean, for instance, for humans
33:14
we use naming a lot. For instance, we offer it.
33:17
We have like a command line tool that allows you to interact with
33:19
your data pipelines in our platform.
33:22
And humans tend to use the name of a pipeline, but there can be
33:27
different different pipelines across different projects and
33:31
teams with the same name. You know, for humans, they
33:34
primarily work in the context of their single team.
33:37
So they use names, right. For agents, when they start
33:40
using names, it start getting confusing because there's
33:42
multiple pipelines with the same name and and agents work across
33:45
different teams. So they tend to prefer I DS a
33:48
lot more than than having human readable names, you know.
33:53
Yeah, and and this also allows them to the the the other
33:57
interesting thing was humans like seeing the information and
34:01
like a much easier to understand way and like tables and stuff,
34:05
you know, like we, we like our our visuals.
34:07
Agents don't care give them like role programmatically accessible
34:11
data. They love consuming that that,
34:13
that type of data. And suddenly you can stuck a lot
34:17
more different depth of information into the same
34:20
response and the agent will make use of that, you know, whereas
34:23
for humans, like if you design A multi dimensional table, it
34:27
already gets complicated. You know, agents can handle that
34:30
complexity. So it's, it's really a different
34:33
audience. It's it's, it's a quite a
34:35
surprising it's it's a very interesting thing.
34:37
Like, I don't know, I wouldn't have imagined this couple of
34:40
years ago. You know, it's pretty.
34:41
It's really interesting. They've gotten smart pretty
34:44
quick. I mean, it's really the last six
34:46
months has been like where I'm like, OK, they're smarter than
34:50
me now for sure. Absolutely, yeah.
34:54
Didn't take much, let me tell you.
34:58
Yeah, I mean, I would like after November was, I think November
35:03
sort of early December was sort of for me, it was like, wow, you
35:07
know, wow. Now like this wasn't possible
35:09
before and, and look what we can do now, right.
35:13
And, and it made, I mean, many of the ideas that are, that are
35:18
not possible today are not necessarily novel.
35:20
You know, it's just that they were not possible a year ago.
35:22
They were not possible two years ago.
35:24
They were possible, but they were not accurate.
35:26
They're not working well. They were more frustrating than
35:28
than actually bringing value. Whereas now, I mean, we do have
35:33
customers that have stopped hiring for data analysts.
35:38
You know, they have, they're instead like hiring for
35:42
different roles and they are suddenly given access to a lot
35:45
of different members that they're all different team
35:47
members to use the AI agent instead of making them talk to
35:52
the data team directly. Right.
35:55
We do see a real shift in, in how teams think about their
36:00
their budgets now. How do they think about the
36:02
access to data? You know, how do they think
36:04
about making that data useful within the organization?
36:07
And, and that was not possible a year ago.
36:12
Yeah, yeah. I mean, it's the, the
36:15
capabilities, but like also these predictive qualities that
36:18
the real, it's the, the analyst, you know, that they, and again,
36:25
this, these are kind of vague concepts because like, how do
36:27
you really measure intelligence? Well, there's probably testing.
36:31
And when they, you know, release a new model, they're like, oh,
36:33
it passed this benchmark for this, but it kind of gets all
36:36
washed over, you know where it ends up.
36:39
Is it it you kind of wake up to it, You know, it's like when you
36:43
you're, you're talking to, but buddy, you know, but I'm talking
36:46
to you like I can tell in the first couple of seconds of like,
36:48
oh, oh, I better be mind my PS and QS Barack's really smart.
36:52
So, but it with the, with the, with the agents, it takes maybe
36:57
a little bit more time. We like, Oh, this is different.
36:59
Like the, the response I'm getting is is smarter, it's
37:02
faster. It's like it's taking into
37:04
consideration additional context.
37:07
I don't need to prompt it as much.
37:08
It's kind of starting to intuit like what I want.
37:10
It's memory is getting longer and these aren't things that
37:13
necessarily show up on spreadsheets.
37:15
So I guess that's just the interesting thing is as it
37:17
begins to get intelligence is there's almost like a a
37:20
recognition that just kind of like emerges.
37:24
I mean, this is precisely the same reason why, you know,
37:27
benchmarks are good. Like we have quite a lot of
37:30
benchmarks internally. We do like a lot of a lot of
37:32
them. But like, in the end, it boils
37:34
down to like vibes. You know, that's the best kind
37:37
of like check that I do with like, you know, if there's like
37:39
a new model that we think that might work well, you know, we
37:42
connected to, to, to some of our data sets and we start testing
37:45
it out. And the core thing I'm looking
37:47
for is does it feel intelligent And does it feel more, does it
37:53
feel smarter than the, than the other alternative here?
37:56
And that smarter can be in different ways.
37:58
It can be that it analyzes the data better, but also it
38:01
interprets the human requests a lot, a lot better.
38:04
It, it comes up with proactive suggestions in a, in a lot, lot
38:09
smart way, much more smarter. And, and this essentially what I
38:14
call that, that vibe check is sort of how we in the end decide
38:18
what goes into, into the end product.
38:20
You know, and, and I do agree with you that I think like
38:25
things like, I don't know, like weird prompting tricks or like
38:29
different tooling that we build around, around the AI, etcetera.
38:33
I mean, they matter today. They're likely not going to
38:35
matter in three months down the line, 6 months down the line,
38:37
you know, and, and suddenly if you're building a product now
38:41
you have a product problem where you're not building for today's
38:43
problems, you need to try to think ahead to build what's
38:48
going to be the problem six months down the line with the
38:50
model capabilities that are available then so that we can we
38:54
can try to be ahead of the game. Yeah.
38:57
I mean, a year ago it was like prompt engineering.
38:59
Here's the right way to ask. Act like a data analyst from a
39:03
senior, you know, Fortune 500 company and look at this one
39:07
spreadsheet. You didn't say it just right
39:10
that the output wouldn't. Now it's just like you don't
39:12
need that at all. Yeah, just like doesn't matter,
39:16
you know, doesn't matter. Like one of the interesting
39:18
things was you can see the same with, with RAG, you know, like
39:21
we, we've built a lot of systems that allowed agents to fetch
39:26
information using like our like embeddings and like try to like
39:31
fill everything into the context, trying to fetch ahead
39:33
of time. Don't get me wrong, it's still
39:35
relevant, but suddenly we got models that are capable of just
39:40
like navigating a ton of information that are just in
39:43
plain text files and inside the code base, you know?
39:46
And yeah, and a year rag will be out.
39:49
It'll be just, it'll understand the tone of my voice, you know
39:53
what I mean? Exactly.
39:55
So this is this. This is madness.
39:56
Like it's crazy. It's it's fun, yes.
40:01
It's so fun, you know? And I'm, I'm probably gonna,
40:04
maybe this is my challenge for this week.
40:06
I'm gonna start talking to it for real, you know?
40:08
So like, if it makes a mistake, I'll be like, yeah, you're real
40:11
sure about that. Like start being sarcastic and
40:13
messing around with it, you know?
40:17
What I wanted to have for a while is the, the, the, a lot of
40:21
the Frontier labs do not update their voice models, you know, so
40:24
voice models are, are usually behind.
40:26
I actually want to have that chat, you know, and I want to
40:30
have like, you know, just like a casual conversation, because I
40:34
think they're, they're definitely capable.
40:35
And one of the things that we prototyped internally was can we
40:39
make our analysts work that way as well?
40:40
Where you, where you call it and then ask questions in a, in a
40:43
human language, just like by speaking to it and it would
40:46
answer you back. And and then you get your
40:49
analysis in a minute with like a, a Slack DM, for instance,
40:53
right? So I think there's gonna be a
40:55
lot of these different interaction patterns that we're
40:59
gonna see with agents that we're not used to yet, you know?
41:02
So it takes a lot of experimentation.
41:06
So exciting. All right.
41:09
Now when I look at your website, one thing that I noticed is
41:11
there's no pricing. So how do you guys, what's the
41:13
business model and how are you charging customers?
41:16
A lot of our core tech is is open source.
41:19
If you go to if anyone goes to our website, they can see our
41:24
you know all of all the links to our GitHub pages.
41:26
We have recently launched a a new module inside Brune Academy
41:30
that allows you to build your own AI data analyst yourself
41:34
using all of our open source tooling so you don't have to pay
41:36
us for our cloud offering. We do not accept public sign ups
41:41
for now and we are about to get into general availability soon.
41:46
The pricing depends on the size of your data, the types of the
41:50
data that you have, the number of people in the organization
41:52
that will be accessing this data, as well as any other
41:56
regulatory requirements that that the organization might
41:58
have, such as, you know, different deployment models,
42:02
different types of agreements that we would need to sign,
42:04
etcetera. We work with very small gaming
42:09
companies all the way down to large multinational retailers
42:12
with thousands of real stores in the world and each have
42:16
different things that, that they ask for us.
42:19
There have been cases where I have signed my left kidney, you
42:22
know, so I would, I would have to make sure that we get a, we
42:25
get a good deal out of that as well.
42:27
But basically it's a it's a case by case scenario.
42:31
Interesting. And is that because just the the
42:34
the stage of business you're in or is that a kind of a strategic
42:37
design mission right now? I would say it's a stage of the
42:40
business that we're in for now. So I do not from the, from the,
42:45
so we're 2 1/2 years old from the, we, we didn't want anyone
42:50
in the team to spend any time building any building software,
42:53
right. We wanted to put all of our
42:55
attention into building a good product, solving customers real
42:59
problems. And and then we, we believe that
43:03
money will money would follow. It has followed and the way that
43:06
we worked was essentially based on some of these criteria.
43:09
We come up with a, a pricing package that we that we agree
43:14
upfront the customers always pay the same monthly amount and and
43:19
they don't have to be worried about that.
43:20
They got like a very predictable pricing and we don't have to
43:23
build worry about building a billing infrastructure, right.
43:25
So we are likely going to do that.
43:28
It's just that we've been, we've been pushing that down the line
43:31
because to be quite frank, building an analyst sounds a lot
43:34
more fun than building a billing software, regardless how much of
43:38
that can be automated. You know, that has exactly been
43:42
the reason why we're we're doing it.
43:44
That's like a custom pricing. When you're in a heck of a
43:47
position to see what demand is too.
43:49
So if you're building broad analysts across multiple
43:53
sectors, what you see is probably patterns emerge and
43:55
then you're like, OK, actually this connection to this
43:58
connection to this result and outcome is really actually hot
44:01
in demand. So you could almost spin a
44:03
product like that off and send it on its way with like in a
44:06
very specific silo is are you doing kind of product validation
44:10
through all this as well? 100% So we're working with a lot
44:13
of different industries. The good thing about what we do,
44:15
like you said, is it's applicable to a lot of different
44:17
industries. We do a lot of these kinds of
44:20
experiments and validations and we do get our hands very much
44:23
dirty. We almost act like as a, as a
44:25
service company, even though we're not, we're not charging
44:27
the customers for it. You know, we just like getting
44:30
deeper into the problems because that's essentially where the,
44:33
the core, the golden feedback lies, right?
44:37
So we we try to get ourselves into all of these problems as
44:41
deep as we can and then the patterns start demerging
44:44
suddenly. One example is we work with a
44:47
lot of gaming and app companies, you know, and that's an industry
44:51
we understand very well. We work with 10s of different
44:54
companies in the space and suddenly we can now build
44:58
solutions that are specialized for gaming or app companies that
45:04
would that doesn't exist in a market that source that saves
45:08
them like days worth of time every week, you know, But one
45:13
interesting part of that, like that pricing equation is this
45:17
flexibility also allows us to be a lot more flexible for
45:21
different types of teams where, you know, we work with founders,
45:24
Founders go through tough times. Founders go through, they have
45:27
different milestones. They need support in some cases,
45:30
you know, and, and that gives me the flexibility to offer them a
45:34
pricing model that's not just standard, you know, it just we
45:37
can make it variable. We can make it change over time.
45:40
We can make it like, I don't know, conditional based on
45:42
something that you need from us, right?
45:44
Or I don't know, you have, let's be real, like you have cash flow
45:49
problems. It's fine.
45:50
We're going to accommodate that. You know, we're not going to
45:52
die. We're around, you know, we have
45:54
a ton of money in the bank. We're going to, we're going to
45:56
be fine. Let's actually solve your
45:59
problems and then we build a, a, a real relationship, which is
46:03
why I'm really proud that that none of our customers have
46:06
stopped working with us. You know, like companies went
46:09
down, they went bankrupt, but no one actually willing to stop the
46:12
using the, the brewing platform, you know, so that's the kind of
46:17
relationship we want to, we want to build with.
46:20
Yeah. And that the that's the humanity
46:22
that they're talking about that AI can never replace.
46:24
You know what I mean is like to have that customer first ride or
46:28
die mentality, man, it's beautiful.
46:30
Yeah, absolutely. I mean, we're, we're, we're,
46:32
we're trying our best. You know, I think like without
46:36
that, that feedback, those customers make the product what
46:39
it is today, we're 100% doubling down on that.
46:44
We're we're, we're going to try our best to keep it, keep it
46:46
there forever. Like we have dedicated Slack
46:48
channels with all of our customers, pretty much all of my
46:51
customers have my personal phone number.
46:53
You know, anyone knows that if, if something happens, they know
46:57
who to call, you know, and they, they know that there's a human
47:00
being on the other side of the table.
47:02
And, and I am part of every sales call.
47:04
I am part of every onboarding. Everyone knows us, you know,
47:08
and, and that might not be the most scalable way of doing
47:11
business in the world. I don't know it's, but it's
47:13
fine, you know, it's fine. Like I, like, I, I went to an
47:17
event a couple weeks ago, like 6 of my customers were there.
47:20
It was like a party for me, you know?
47:22
Amazing. Friends, you know, it's, it's
47:25
such a great feeling. I love it.
47:29
We got a couple minutes left here.
47:31
Aside from brewing, what else are you playing around with?
47:34
What? What else are you building?
47:37
So we're, we're definitely like a lot of organizations we're,
47:39
we're, we're spending a lot of time internally thinking about
47:42
like software factory, you know, we are, we're trying to think of
47:47
ways of getting ourselves more, more and more productive in
47:50
terms of like building software all around.
47:54
We are, I am personally spending a bit of time these days with,
47:59
with software generated music. I, I play, I play music.
48:04
I, I have, like, I, I have little, I like little
48:08
instruments around, you know, I like, play some guitar and
48:11
stuff. And, and I'm trying to, trying
48:14
to play these days with the, with the help of AI, with some
48:16
procedurally generated music. You know, not much on that front
48:20
yet. I still feel more comfortable in
48:22
the guitar myself, but hopefully I'm going to get better and the
48:27
and trying to find. Stuff you doing arena rock?
48:29
Are you doing trap music? What kind of music are you?
48:32
Making, I'm trying with, I'm trying a bit of trap these days,
48:37
Yeah, just just you know, just playing, just taking a chill.
48:41
Yeah, Rock. This was awesome, man.
48:43
Thank you so much for your time. getbruin.com is the website to
48:49
go check this out. It's open source.
48:51
Go play with it. You know where to find Barack.
48:53
Actually, to that point, what's the best place that people can
48:56
reach out to you online? I would suggest LinkedIn is
48:59
probably the best place to to reach out to me directly.
49:02
We also have a Slack community for our open source users online
49:07
there 24/7 as well, so if anyone jumps in search my name I'll
49:12
I'll be there. Amazing.
49:14
And before we get out of here, I know internetsociety.org is a a
49:19
favorite charity of yours. We want to give a shout out to
49:21
them. Give us a a quick rundown of
49:24
exactly what Internet Society does.
49:26
Internet Society allows basically making Internet
49:30
accessible to the rest of the world across different
49:32
initiatives. They are a a group of very smart
49:36
people, very hard working people trying to make sure Internet
49:40
gets all around the globe, you know, and everyone gets access
49:44
to the same, same chance. They're they're one of our
49:48
customers. I love working with them.
49:50
They're they're they're an actual stellar team.
49:53
So shout out to them. Shout out indeed.
49:56
Again, that's getbruin.com. I'll link Barack's LinkedIn in
50:01
the show notes below. So if you're listening to this,
50:03
go down the show notes and he's a great follow on LinkedIn.
50:06
Really appreciate the time today.
50:08
Ryan, thanks a lot for having me.
50:09
Appreciate the conversation and thank you.
50:13
You are about to spend anywhere from 100,000 to $300,000 on your
50:17
kids college education and that's not a small bet.
50:20
And right now there's a very real chance that they graduate
50:23
and can't find a job. Not because they weren't smart
50:25
enough, not because they didn't work hard enough, but because
50:28
the job market changed and nobody told them.
50:31
Employers are hiring AI native candidates, people who don't
50:34
just know their field but know how to work with AI at a
50:37
professional level. And most colleges are not
50:39
teaching this, not even close. AI Native student.com is the
50:44
program that fills that gap. It takes high school and college
50:47
students and teaches them how to actually think and work with AI
50:50
the way top companies expect on day one.
50:53
So your investment in their degree actually pays off.
50:56
You're already paying for tuition, housing, textbooks.
50:59
This is the one thing that actually makes it all worth it.
51:02
Go to ainativestudent.com because a diploma is a starting
51:05
point. Being AI native is what gets
51:07
them hired. That's AI native student.com.