Latest / AI for Founders with Ryan Estes / Data agents you can trust in production. | Pradnesh Patil from Altimate AI

Transcript

Data agents you can trust in production. | Pradnesh Patil from Altimate AI · AI for Founders with Ryan Estes

0:00
For enterprise, your data team is the most expensive bottleneck
0:03
in your entire company and most founders don't even know it.
0:05
What if the answer isn't hiring more engineers?
0:08
It's agents and the agents you have been flying blind this
0:11
whole time. Let's say you've got a 15 year
0:13
veteran whose entire brain is your company's most valuable
0:16
asset, and the second they leave, if they do, the knowledge
0:20
walks out the door forever. Or does it?
0:23
We need an AI that can watch how our best engineer fixes
0:26
problems, remember it, and teach every new hire, human or agent
0:30
to do it the same way. Discernment.
0:32
Now imagine that happening across your entire data stack.
0:35
Pipelines that build them self infrastructure that self
0:38
optimize governance that doesn't need to chase engineers down
0:41
hallways. You don't need to pay service
0:43
firms millions of dollars to migrate your legacy stack.
0:46
Let agents do it for you for a fraction of the cost in a
0:49
fraction of the time while your engineers finally get to do the
0:53
interesting work. The data doesn't lie, but do you
0:55
know when your company who owns understanding it?
0:58
Welcome to AI for founders. I'm Ryan.
1:00
Today's guest is Prodnish Co, founder of Ultimate AI, and this
1:04
episode is for technical founders, data-driven super
1:06
operators, and anyone building teams where speed of insight is
1:10
a competitive weapon. You'll learn how to use AI agent
1:13
harnesses to 10X your data output, slash infrastructure
1:16
cost by up to 40%, and unlock institutional knowledge before
1:20
it disappears forever. And if you want the frameworks,
1:23
the tools, and the Founder Edge delivered straight to you, sign
1:25
up for our newsletter. At aiforfounders.co.
1:28
Where we turn AI chaos US into your unfair advantage.
1:31
It's the weekly cheat code founders are quietly passing
1:34
around. Don't be the last one to find
1:36
it. And hey, if this episode made
1:37
you think differently about your data stack, please drop us a
1:40
five star review. And bonus points if you write
1:43
something that makes me smile a little bit pranish.
1:52
Thank you so much for coming to the show.
1:54
Welcome. Thanks for having me Blidgett to
1:58
be here. Yeah, You know, it's kind of
2:00
wild doing a podcast on Monday morning.
2:03
I feel like I always need a couple of hours just to get my
2:06
brain going, but I see we're on the same wavelength.
2:08
Why is black T-shirt the best uniform for Monday morning?
2:15
I think because the color is sort of similar to coffee,
2:19
right? And it gives you the energy.
2:22
I can't get started with a couple of cups of coffee.
2:24
So, you know, just more of a boost to get that energy going.
2:28
100%, yeah. I'm on my second latte, so I'm
2:30
feeling very good coming into the week.
2:33
Super excited to talk to you, largely because you've had a
2:36
recent launch. I know we're going to talk about
2:38
that. Also, you have one of the most
2:40
logical applications of AI, which is data, data management,
2:45
data insights, and we're talking about your company, Ultimate dot
2:49
AI. So give us an introduction to
2:51
your company. Absolutely.
2:54
So ultimate AI, very close to our heart.
2:59
We have a team of about 50 people and all of us, we are
3:02
focused on how we can help different data teams in the
3:07
world use AI in their day-to-day work, how AI can be used to
3:13
build things like data pipelines, you know, those
3:16
extractable transform VLD pipelines.
3:19
How we can help their how we can help these data teams manage
3:22
their data infrastructure better.
3:24
For example, people use tools like Snowflake data breaks,
3:29
Google Big Query or AW Redshare. How we can help them manage
3:33
their data infrastructure better and more efficiently so they can
3:37
get more performance out of it at highest efficiency possible
3:42
and not just limited to, you know, data pipelines or just the
3:45
data infrastructure. When it comes to data,
3:48
governance is a big part of it. How we can help teams govern the
3:53
data better as well from the sensitivity perspective, who has
3:56
which axis perspective? There's so much work involved.
4:00
Data teams always struggle to finding the right people to do
4:04
this work, having you know, people to do this work and all
4:08
of us are in luck now. AI is coming into the picture.
4:11
If we use it in the right way, it can solve lot of problems for
4:15
us and help us deliver technics amount of the projects we used
4:20
to do manually a few years ago. Incredible and and talk about
4:24
the launch. I know you we, we spoke
4:26
beforehand of just a little bit that you bundled up a package
4:30
and open sourced it. So talk about that.
4:34
So we have been doing this. We have a bunch of users using
4:39
the product and then we thought, hey, what's the best way to give
4:43
everybody access to what we have in building and at the same time
4:48
give them opportunity to contribute to the project also.
4:52
So we just recently took much of pieces that we have been working
4:57
on and open sourced it as ultimate core project, which is
5:01
basically an agentic data engineering harvest that can be
5:06
used to do data engineering work, data work lot better.
5:11
It offers lot of advantages. It's tech stack and NASDAQ.
5:17
So as I was telling you earlier, you can people usually have
5:20
these hybrid data stacks when it comes to tools etcetera.
5:23
So it works with a lot of different tools and platforms.
5:27
It's a very token efficient. We have been hearing all that
5:30
chapter about the LLMS are costing me so much money, right,
5:37
And and that lot of that happens because we just tried to prove
5:41
force LLMS into the specialized task also like PITA or that it
5:46
goes in all sort of different directions without context,
5:49
without like tools and skills and in some consuming so many
5:53
dockers. And we have well, the word flaws
5:56
inside this HMT card less very token efficient also, and people
6:00
can bring in their own LM. So unlike being forced to say,
6:04
using clock gold for less or something else which can get
6:07
costlier very rapidly if you use moose cutting edge models,
6:11
people can bring in their own NLM, keep the cost down.
6:14
And also it helps from the government's perspective, they
6:16
work with companies in a sort of more regulated industries as
6:20
well, like healthcare, financials.
6:22
They also appreciate that a lot. And then the third part is
6:25
customizability part of it, you know, every, every enterprise is
6:29
little bit different. They have their own work flows
6:32
their own way out of things. And in that situation, you can
6:36
customize these things based on your teams, our nodes, etcetera,
6:40
because it's open source. Incredible.
6:43
Now, I admittedly don't know much about data engineering, so
6:47
when you say harness what, what exactly does that mean?
6:51
So the concept of harness is actually not just limited to
6:55
data engineering. So you might be seeing, I think
6:58
this is a term that has been coined in the space recently and
7:01
people have started talking about it at a very high level.
7:05
Harness is a set of components that you need to provide to LLM
7:10
to make sure the special specialized work that you are
7:14
doing, you are providing the right things to LLM to get that
7:17
work done. For example, the harness can be
7:20
for your finance teams. Also harness can be for the
7:23
marketing teams as well. In a similar way, when it comes
7:27
to harness for data teams, there are few key components to it.
7:31
One is the context that is coming from the data stack.
7:35
So for example when you have this data sitting in a lot of
7:37
different systems, especially when it comes to data
7:40
engineering, how you can put metadata about that data.
7:44
Lot of tribal knowledge is today sitting in peoples head, how you
7:48
can pull it out and provide it to LM.
7:50
So one part of it is that context, the second part of it
7:54
is the governance side of it. I will, I was touching on that a
7:57
little bit up here. Governance is very important
7:59
when it comes to data, how you can give right rules and
8:02
permissions and access controls when you hook into the entire
8:06
data stack of the company and try to do it.
8:09
That second governance components is also very
8:11
important part of that harness. Now then there are other
8:15
components like tools and scales, like when you are
8:17
building ADR pipeline, for example, or when you are
8:20
managing data infrastructure in a certain way, there is a, there
8:23
are certain recipes to it and certain tools you need to
8:26
connect it to. So those two components, tools
8:29
and skills are also considered important parts of that, of
8:33
this. And last but not the least, you
8:36
know, previously we had humans doing all the work and now
8:39
they're going to be in conjunction with humans,
8:41
hundreds of regions doing this work.
8:43
So imagine the simple situation, engineer went and made some
8:46
change. Now they need to test it, right?
8:49
So in that situation, you need to have that engineer use some
8:53
kind of development environment. You can't go and make the
8:56
changes in production environments, etcetera.
8:58
Now extend that to agents. When you have hundreds of agents
9:01
doing the work, we need to think about how I give sandbox
9:05
environments to these hundreds of agents.
9:07
So that is another component of the harness as well.
9:10
To summarize, context, governance, rules and skills,
9:16
and then the last part, the right infrastructure for the
9:18
agents. These are usually considered the
9:21
key components of the agentic harness for the data worker data
9:24
generic. I love it.
9:27
And for that last part, you know, where are we talking about
9:29
alignment, 'cause that's a thing that it's something that comes
9:32
up more and more. Like finding alignment with a
9:35
team of 50 can be challenging, but if everybody has 10 agents,
9:39
now you've got 500 people that need to be working in the same
9:43
direction. Is that kind of built into the
9:44
harness as well? Absolutely, because what
9:49
happened previously was we had a lot of knowledge sort of siloed
9:53
into people's heads as well and just couple of people are
9:56
experts, know exactly how these systems work, a couple of other
9:59
people in the different areas of the company, they know a
10:01
different system, etcetera. Now what is happening is we got
10:06
all those agents also in the wings and they can't, they can't
10:10
remember things in their head. They are not human.
10:12
So there is this whole concept of the memory components of it
10:16
when it comes to infrastructure. And when I say memory, I'm not
10:20
just talking about individual level of memory.
10:23
You can have this giant pool of memory that's shared between
10:26
agents and humans and it gives us an opportunity to sort of
10:31
kind of a way highlight mind where all the tribal knowledge
10:36
is also sustained and this is amazing brain at the enterprise
10:40
level we can build at least that's what we are trying to do
10:43
at the data work level as well that can be used to do the
10:47
things to learn better and in a smart way.
10:49
And for example, even in humans, right, we saw this different
10:52
level of skills somebody has to be in years of experience
10:55
exactly knows how things are done here.
10:57
Somebody just joined six months ago.
10:58
Now there is a skill variation. And then one of the biggest gaps
11:02
for that is the knowledge that has been accumulated over years
11:05
of experience. Now imagine if you can digitize
11:09
that knowledge right here. We build it up in this way.
11:12
Hey, our revenue is calculated in this way.
11:14
Our data infrastructure, you just combine all of that, and if
11:17
it's shared between humans and agents, suddenly everybody is an
11:21
expert. Yeah.
11:23
So let's that's a big challenge though, right.
11:26
So assuming that I'm the senior data engineer, I've been around
11:30
for 15 years, I have all the experience, all the expertise
11:33
about this particular system. How do you set about like
11:37
extracting that expertise and giving it to the agents?
11:41
So there are a couple of different ways.
11:43
So in the framework itself, as people build these data
11:46
pipelines or manage data infrastructure and they are
11:49
doing this work in the in our project itself, in ultimate
11:54
code, we have built a mechanism where the system itself will
11:58
start creating memories automatically.
12:01
Science you are doing the work, it will start creating the
12:04
memory. Maybe Ryan came in and he
12:06
actually corrected the LLM in such a way it will create the
12:10
memory for it. Or for example, you are fixing
12:12
something in your work. When this type of issue
12:15
happened, Ryan came in and fixed it in this way.
12:18
And last four times it happened in a certain way.
12:21
Now that knowledge is digitized. Now imagine somebody who is less
12:24
experienced, but just once of experience hit the exact same
12:27
issue in some other area. The system is going to recall
12:30
that information the last four times my principal engineer
12:34
fixed in this way. Maybe I should help this new
12:37
person fix it that way as well. So lot of it is becoming
12:40
automatic. It's not 100% for proof, but as
12:43
you're correcting that AI to work with you and get the work
12:47
done in ultimate code or in other systems, it will start
12:50
creating those memories and knowledges.
12:52
Beyond that, you know user can give instructions also.
12:55
So for example, you hire an intern, right?
12:58
You got a new person on the team, you're going to take an
13:01
effort to cost that person because eventually that is going
13:03
to be going to make your work lighter.
13:06
So actually this is little bit of the education.
13:09
Also data teams should start coaching these agents actively
13:12
as well so they can help them more as well.
13:14
So there is a mechanism for that too.
13:17
But inherently automated is always the best.
13:20
If it's creating memories for me and capturing things things
13:22
automatically, that works really well and users can update it
13:26
also. You know, yeah, it's not
13:27
perfect. Sometimes it takes around things
13:29
you can go in and correct it as well because all the memory
13:32
repository you can go in and see as well.
13:36
Quick side note for founders who are using AI to write content,
13:39
you've probably noticed this. It's incredibly fast, but it
13:43
doesn't really Dr. meaningful traffic.
13:45
And honestly, that's because most AI generated content is
13:48
just average. There's no real insight, no
13:51
depth, no structure, missing things like internal linking,
13:54
and Google's getting very good at filtering that out.
13:57
So instead of trying to just prompt better, I actually came
14:00
across something interesting called Neena.
14:02
It's a system that combines human strategy and review with
14:06
AI writing to create high quality SEO optimized blog
14:09
content that's actually designed to rank and keep ranking.
14:13
What I like about it is that it's done for me.
14:15
Their team handles the topic strategy in the content review
14:19
and data markups for Google, and turns blog posts into actual
14:22
long term traffic assets. So if you're a founder thinking
14:26
seriously about organic growth, this is one of those things
14:29
worth paying attention to. You can check it out at
14:31
trineena.co. That's TRYNIN a.co trineena.co.
14:39
In three years, AI fluency will be a baseline expectation for
14:43
every college grad, not a differentiator.
14:45
It'll be a requirement. The students getting ahead are
14:48
the ones building that skill right now in high school and
14:51
college before it becomes table stakes. ainativestudent.com is
14:55
the program doing that weekly one-on-one coaching, live cohort
14:59
sessions, on demand courses and the fastest growing AI chat for
15:04
students all tailored to your students specific career path.
15:07
Pre Med they learn how AI is transforming diagnostics.
15:11
Pre law they learn legal AI tools business they learn to
15:14
think like an AI power operator. 1 flat membership less than a
15:18
private tutor. Cancel anytime.
15:20
And if you're a founder listening to the show, you
15:22
already know the cost of falling and behind.
15:24
Don't let your kid pay that price.
15:26
Go to ainativestudent.com, apply in a couple minutes and get your
15:30
student into the cohort that changes their trajectory.
15:33
AI native student.com and send them to the chat.
15:36
It's totally free. Cool.
15:50
So it's the agents are observing you and then it's kind of
15:54
learning your personal discernment so that it can have
15:58
ultimate. Well, I don't know if I want to
15:59
go that far. What this reminds me of though,
16:02
is if I'm feeling particularly irritated and I'm, I'm dealing
16:07
with the team, I will say there's a right way to do it.
16:11
There's a wrong way to do it, and then there's the way that I
16:13
want it done. And that's how we're going to do
16:15
it. So, so in this case, it's like
16:20
this might this agent might be able to say like, Hey, what what
16:23
you're doing here is actually this is how Ryan would do it.
16:26
So if that it influences your behavior is that, is that kind
16:30
of what we're talking about here?
16:32
Exactly, exactly and that and you're talking about an
16:36
individual level, but this actually percolates at the
16:41
company level also as a as a bag in this particular region, for
16:46
example, our data workloads are done certainly This is why we do
16:50
it this way, right. All that knowledge and
16:52
customization that should come in when the work gets delivered
16:56
as well. And this is a way to get it
16:58
delivered. And because because we provide
17:02
these on these components that customization or making agents
17:06
that much stronger for data work has been possible.
17:08
So for this project, if you get a chance, definitely check out
17:12
the website also that we have put around the ultimate code.
17:15
We are testing it at Easter benchmark, which is an industry
17:19
benchmark. It's called Agentic Data
17:21
Engineering benchmark or Ade benchmark in short.
17:25
And we tested our solution again some other promising solutions
17:29
in the industry today. So Snowflake has launched
17:32
something called Cortex Core or DBD also has some other
17:36
components around MCP server, some other fusion components and
17:41
all our favorite cloud core as well.
17:43
Like we have been using very heavy cloud core for software
17:46
engineering work. We use it heavily also
17:48
internally for our development work.
17:51
But log code is not suited for data engineering work.
17:55
It can do software engineering work just better because it's
17:57
missing all those hardness compromise.
18:00
Now when we tested Ultimate code against the benchmarks, right
18:03
now, we are #1 on the benchmark with 75% Cortex code is
18:08
somewhere around 65%. And when I look at log code,
18:11
because it's missing all these hardness components, it's
18:13
somewhere around 45%. So we see a huge difference of
18:20
these hardness components, what it can do for these agency
18:23
cultures. What we are even seeing is it's
18:26
not about using the latest and latest model like opposite the
18:30
latest right now and the one before that is Sonnet.
18:32
And what we are seeing is even if you use Sonnet with the right
18:36
components, you are actually going to get much better
18:39
results. So it's not actually about the
18:41
most cutting edge LLM, it's actually about what kind of
18:45
information you can feed to that LLM and how much you get told it
18:49
for specific tasks like data tasks in your company.
18:53
I love that, yeah, because you're right, like everyone is
18:55
like, well, the last couple of weeks it's like, oh, why is
18:58
Claude code so expensive now? Why am I?
19:00
Why am I burning so many tokens? They're, you know, always
19:03
putting me into the latest, greatest model.
19:05
And so people are figuring out how to put this harness on it so
19:09
that you can use it. Now, what does that tell us
19:11
about kind of market dynamics for the big LLM companies in
19:15
general? Because ultimately we're going
19:17
to have these harnesses that are like, you know what, for 90% of
19:20
your, your tasks, this open source model is totally free.
19:24
Is is best. Let's use that.
19:26
And then we'll bring in Sonic 4.7 or whatever for very
19:29
specialized tasks. So you're just kind of sniping,
19:32
you know, for like your with your tokens.
19:34
What do you, what do you think that's, that spells out for kind
19:37
of the, the, the future of these companies, particularly if like
19:40
open AI is going public this year potentially, I mean.
19:46
Yeah, you are, you are absolutely right.
19:48
And there are a lot of dynamics into play here.
19:52
One this part is you do need latest and greatest model every
19:56
time, right, Because people have been working on that model
19:59
gateway as well. So depending on the task people
20:02
are working on making it even automatic and we we are
20:05
including that component in our solution also, especially for
20:08
data tasks. Only an E&B kind of a task like
20:12
data modelling in specific areas you go use your data stand
20:16
right? But for other simpler tasks,
20:18
maybe it's something simple as I'm just trying to write a
20:21
description for a column or a table or I'm just trying to
20:24
write a simple data task. We can route them automatically
20:27
to a lower tier model as well. So that's one part of it like
20:30
there is a lot of this routing will start happening.
20:33
Second is people will start bringing in their own models as
20:37
well as open source models as the cost start getting out of
20:40
hand. It's a very famous story from
20:42
the Uber CTO, right? They blew through their budget
20:45
in just four months. And when things start happening
20:47
because enterprises start thinking about, hey, how I can
20:50
bring my own model, like can I use some sort of gate?
20:53
We are now going to do these things.
20:57
So in short, I think the party on the unlimited tokens and
21:02
everything is kind of the whole world, right.
21:04
What parties people are going to bring in their own models also
21:08
for a long time open the eyes and entropic of the world
21:11
shoulder the bill for people's tokens like the 100 to $100
21:15
subscription that you get with your board.
21:17
It gives you tremendous amount of tokens for the amount of
21:20
money you are bringing. And that is going to end also
21:22
because how far you can just subsidized the money and also at
21:27
that such a high percentage. I'm trying to guesstimate here,
21:32
I'm not exactly right, but actually the API of Spanthropic
21:36
is almost 30X expensive than the Cloud Code subscription.
21:41
So when there is a bad big of dispatch, things are definitely
21:45
going to change. And I guess all of us, we are
21:47
going to figure it out together what it means for options that
21:51
you can use, how much money we should spend and everything like
21:54
that. You mind if I get a little
21:57
philosophic? Go for it.
22:01
The fun part about AAI right now is we seem to be racing towards
22:05
AGI and no one really has a great definition of what that
22:10
is. But I think that the devil's in
22:12
the details. And ultimately it's, it's the
22:15
folks like you that are engineering the data in novel
22:18
ways and figuring out ways we can use this, this technology to
22:22
achieve the outcomes we want. So when I hear about like the
22:27
agents, like learning discernment as I would do it,
22:30
it's almost as if it's creating a theory of mind of who I might
22:36
be in a way that we do with each other, which is just to say that
22:40
like, I know I'm conscious and I assume you are, but I, I can't
22:43
ever really know that. But I have a theory of mind of
22:46
who you are that comes through our interactions and learning
22:49
your patterns. And you learn my patterns in the
22:51
same way that the agents are doing this.
22:54
And it's doing this at on a sample rate of billions of
22:56
people all at the same time in billions of actions, you know.
23:01
So is there anything that sticks out or is surprising when you
23:05
start to have agents that are creating discernment of people's
23:09
habits and patterns that maybe gives you pause or at least just
23:14
like piques your curiosity? No.
23:18
So make them straight, even doing theaterial work, right?
23:24
At least I'll share my experience with you, right.
23:26
I'm in the product building, figuring out how our UI should
23:30
look like, how our functionality should be like or how the
23:33
workload should be like. And when we start discussing
23:37
that with say cloud or other solutions, it's, it's
23:42
fascinating. It's like actually working with
23:44
a Co worker with sometimes the kind of ideas also it can come
23:48
up with and even I can't think of which it's sort of
23:55
developing. It's part of view and
23:57
personality. And the more you do it and more
23:59
as I was talking about earlier, the more context and information
24:02
you feed into it. It's very, very interesting.
24:06
Now the other side of it is sometimes it can turn into echo
24:10
chamber also because it doesn't know what it doesn't know.
24:14
It's dependent on you for the kind of information you are
24:17
viewing in. Of course through that larger
24:18
LLM training it knows a lot of information and can go and
24:22
search web etcetera. But what I was talking about
24:25
earlier, some of the things that you just know, those were just
24:29
some conversations you had with people at least of people's
24:32
solution and things like that and then agent wouldn't know
24:35
those. Now in that situation, the
24:38
hardest part is validating what it doesn't know.
24:42
So when you came up with this situation, hey, have you thought
24:44
about these six other things also?
24:46
Because those were hardly conversations for me.
24:49
So when we talk about this traits and sort of personality
24:53
development, that's, I think that's how like when you talk
24:56
with that, you'll feel it as it's doing things along those
24:59
lines, etcetera. But it's, it's not perfect.
25:02
It has its gaps, but the hardest part is figuring out where the
25:05
gaps are. Otherwise it just turns into a
25:07
big economy. Yeah, yeah, it is kind of like a
25:11
Co worker, you know, it gets lazy sometimes.
25:13
It'll start giving me tasks. I'm like, no, you do that and
25:15
it's like, oh OK, it goes. Doesn't.
25:18
Yeah, cool. Let's kind of take it back to
25:22
the inception of the company. I know the company's only a
25:25
couple of years old. What was the central problem
25:27
that you were coming up against that kind of inspired the the
25:31
company in the first place? Yeah, no, it's a it's a pretty
25:35
interesting story. So Alan, my Co founder and I, we
25:40
started the company together. We have a trends for 10 plus
25:43
years. So we always did a lot of these
25:45
different projects together including Autonomous crypto bot
25:49
which did training for us in the crypto exchange.
25:52
Now my background is I have been a product leader for a long time
25:58
building products which brought in hundreds of millions of
26:02
dollars in revenue. Being the product leader, data
26:05
was always important. How many customers we have, how
26:08
are we trending functionality, how it is being accepted by the
26:12
sales or are we getting too many support requests for this
26:16
particular functionality, etcetera, etcetera.
26:19
And whenever I presented this information to the leadership,
26:21
there's always this culture like, hey, why this number is
26:25
looking different? Why?
26:26
And then nobody knew and go back, try to figure it out.
26:30
Or can we get a couple more insights on this boom, You are
26:34
waiting for two months to get those insights.
26:36
And a lot of times they have been in these situations.
26:39
We are then unfortunate of window has passed because we
26:41
couldn't get the data analysis done at the time.
26:43
Sometimes you just you need to make decisions quickly in a
26:46
week. So then I was chatting with
26:49
Anand like, hey, in my previous companies I worked for like
26:54
really amazing companies, Fortune 5 hundreds, etcetera.
26:56
We haven't like a bad data team. Why it's taking so much time and
27:00
other's background is he did engineering, ML engineering work
27:04
for many, many years as a leader and I was like Pradesh.
27:08
It's the situation across the world.
27:10
You don't have a bad data team. It's it's everybody's problem of
27:14
data world takes time and it's because it's complicated.
27:17
You need the Super expert people and usually we get pulled across
27:21
all the different departments of the company.
27:23
So there is no, not enough people always.
27:26
And that was the genesis that the IV was started coming in,
27:29
you know, GPT models, etcetera. And we thought, hey, let's,
27:32
let's marry this together. We need to marry data, work with
27:35
AI, accelerate these things tremendously so we can increase
27:39
the output, increase the velocity.
27:41
So actual consumers kept their things very fast.
27:44
And at the same time, for data teams also, it's not crazy hours
27:48
of work and you are not always chasing some escalation and
27:52
stuff. Rather than that, you can
27:54
deliver all their projects in your pipeline, more things much
27:58
faster and don't do boring work, right.
28:00
They, for example, engineers don't like documentation,
28:03
engineers don't like governors, thinks a lot of times that yeah,
28:07
I take care of it. You don't.
28:09
Interesting data modelling work and complex works.
28:12
Yeah. So that was the.
28:15
It's great, you know, and it kind of points to like a, a
28:18
future where like you can unlock some of this brilliance into new
28:22
creative ways because, you know, data engineers, there's always
28:26
2, not enough of them, and they always have a backlog of tasks
28:30
that's like 25. And so they're just like putting
28:32
out fires as opposed to like having wrapped up all of their
28:36
tasks to be like, OK, now what can we do that's new?
28:39
You know, how are we doing something productive as opposed
28:42
to something that's reactionary? Exactly.
28:45
Exactly. Then once we figured out, hey,
28:50
this is something interesting, we did some validation with
28:53
people being in San Francisco area or you know, sometimes
28:58
outside, just call it Silicon Valley.
29:01
We showed just a few slides to raise some money, didn't even
29:04
write single line of code at night time.
29:06
And we started building what we built.
29:10
Initially we built for use cases which most of the engineers
29:14
hated, like documentation or writing test.
29:17
Nobody likes writing data test and that got really a lot of
29:23
interest like yeah, I gave off a lot of this work to AI and we
29:27
did a free offering at that time.
29:29
We have read the product and the product got downloaded.
29:34
Now I think we have almost like million downloads now.
29:37
So even then like it was rising like very rapidly it's still we
29:43
are getting a lot of downloads. So then our free solutions
29:46
started getting used in high plus countries and also the
29:52
people using it. A lot of feedback started coming
29:55
in because as as people say like hey, in the world of AI,
29:58
building a POC or building a demo is very easy.
30:01
Replacing people by showing demos is very easy.
30:03
But the hard part is when thousands and thousands of
30:06
people are using it in production and trying to solve
30:09
all the other cases for you. So then we expanded product
30:12
rollback from there. We started doing more of
30:15
production use cases and then we made our enterprise offering in
30:19
America and some of the biggest names in the industry like some
30:22
other Fortune Finder because big customers started using our
30:26
product. You can see a bunch of logos on
30:28
our website etcetera, very cutting edge teams.
30:31
And then we raised our second round of funding and started
30:34
doing it at more and more scales.
30:36
So now the Ultimate code is our latest offering where we have
30:40
combined bunch of pieces we are doing at the smaller scale for a
30:43
couple of ecosystems. And now we have expanded it
30:46
basically for a very large hybrid data stacks and trying to
30:50
help people with migrations like legacy workload, migrations,
30:54
data workload migrations, data infra management, data pipeline
30:59
development. And just imagine all the world
31:01
that usually data generic teams, data analyst teams to we are
31:06
trying to help them automate much of it through this ultimate
31:09
code agent offering behavior. Thank you bet, Eddie, is are
31:13
most people using it for the governance and security
31:15
features? They're like, OK, we have this
31:17
buttoned up so that they can move quickly.
31:21
So highest use cases we have seen is data pipeline
31:24
development, like the DLD pipeline development and
31:27
debugging #2 is around data infrastructure management and
31:31
optimization. You know that data
31:33
infrastructure cost millions and millions of dollars of Vita,
31:37
right? Whether it's software or data,
31:38
etcetera, people use our solution to automate bunch of
31:41
tasks, save engineering time and at the same time use it to
31:45
optimize data infrastructure. Also many times it brings down
31:48
the cost by up to 3040% and saves people millions of dollars
31:52
in there. Third part where it's very
31:56
popular is on the governance side of the things.
31:58
A lot of times you know your data stores or data governance,
32:02
they are not super technical. They have to go and chase data
32:05
engineers. So there is a lot of coupling
32:07
that needs to happen. Now they and they need to
32:11
generate these reports etcetera within the sensitive data has
32:14
been copied. Are we are we deprecating tables
32:17
in a right way? Now you can do a lot of that
32:20
using ultimate code very easily. People are using for that as
32:23
well. And last but not the least, data
32:26
migrations, you know, there are always legacy text tags.
32:29
People are trying to move away from those.
32:31
And traditionally that used to cost millions of dollars
32:34
depending on your scale. And you used a services company
32:37
to do it because migration is a complex, nasty task that usually
32:42
cost things. Don't want to do it, right.
32:44
So you go find, yeah, you find some services company pay
32:48
millions of dollars, try to do that whole dance of coordination
32:52
with somebody external. And now with the AI coming in,
32:55
even in migration, a lot of those things can be automated.
32:58
People are using it heavily for it.
33:01
Yeah, you know, and for enterprise IT, obviously there's
33:04
so many applications, you know, because they have so much data.
33:07
What about for a team like yours?
33:08
You know 5050 team members? How do you use ultimate?
33:13
So even in our product, I was talking about that context that
33:17
we need to pull from people's hybrid data stack, right?
33:20
So it's a data pipeline. You're pulling context from
33:22
different system, massaging it and making it available for AI
33:27
agents. And as part of that, we
33:29
ourselves have to maintain a data infrastructure.
33:32
We have to build data pipeline. Once you build data pipeline,
33:35
they're going to break. You have to debug those.
33:37
And we have been using Ultimate code extensively for it.
33:41
And our data engineers, our software engineers love it when
33:45
it comes to data work. And that's a validation for us
33:48
also that when we are using our own product and how much it is
33:51
helpful for us and whenever you know, it's a way for us to find
33:55
the issues as well. Whenever we do a new release,
33:57
sometimes something goes off. You find it right away because
34:02
we are using it internally and we get it fixed right away so.
34:05
Yeah, Yeah. You know, I, I, I'm thinking of
34:08
a personal use case that I would love to use it for largely
34:11
because, you know, if I'm evaluating a start up investment
34:14
opportunity or something like that, I'm, I'm always looking to
34:17
like LTV to CAC. You know, and people like I got
34:20
a four, I got a four LTV to CAC. I'm like, Dang, well, how are
34:24
you like determining LTV and how you determining CAC?
34:27
And it's some like, it's some calculus of like, who knows how
34:31
it's all mess moved around. And of course, one, they want to
34:35
make their self look, look great.
34:36
And I, I totally understand that.
34:38
But also it's like you have all these desperate data sources
34:41
that makes it difficult, you know?
34:43
So if you do the, the beautiful thing about AI is like, hey,
34:46
give me a reasonable LTV, give me a reasonable CAC and tell me
34:50
like how you're determining that.
34:52
And then you can massage it from there.
34:54
It makes it a lot easier because usually it's probably maybe half
34:57
a dozen kind of functions for each of those, you know?
35:00
Would that be a good use case? Yeah, it can be.
35:04
So for example, I'm assuming at your scale, of course, the data
35:08
is not as big as enterprise level.
35:10
So you can have something like in WPB is one of the very
35:14
popular, you know, data warehouse sort of a solution for
35:17
this. You get, you put all that
35:19
information in WPB, pull in the right context through the
35:23
ultimate code and it will fill that in as well.
35:26
And then you write some of the tools and skills around
35:30
specifically how you calculate LTB or what kind of data you
35:33
need. It will get the work done for
35:35
you. And in fact, actually it was
35:37
built what we call as a data apps for you, which is like a
35:41
nice UI in the React and all these components and chart
35:45
because even you probably want to track how their LTV or CAC
35:50
had changed over time as well as you are considering investing in
35:54
that particular opportunity. So that entire thing based on
35:57
your directions, it won't build it out and give you a very nice
36:01
data that you can just click and you could choose different
36:04
companies. And even if you might want to
36:06
compare two different companies or three different companies in
36:09
the same industry segment, it will allow you to compare it.
36:13
If you go to our website, there is a documentation in which we
36:17
have put some of these examples already.
36:19
And the beauty of this is I talked about that PB or some of
36:23
the other components, right, which are data infrastructure
36:25
components. Now, usually it's challenging to
36:28
installing those components, configure those, you know,
36:31
maintain those, but as part of Ultimate Core, it will install
36:35
at the field of those components also for you.
36:38
So you just show up with your machine and then it will create
36:41
that in potential layer for you, that software layer on top of
36:45
it, and ultimately that presentation layer as well where
36:47
you just go and click things and keep it up today.
36:51
I love it. Maybe we'll bring you back and
36:53
just have to do a live demo because that sounds awesome.
36:55
I love the live builds. Yes, yes, lifetime has are
36:59
always fun and getting those questions and figuring it out
37:03
for yourself. That's all it's for.
37:05
It's the best. OK, cool.
37:06
We got a couple more minutes here, so I'd like to take the
37:10
conversation into the quantum round.
37:13
Are you ready for this? Let me try all.
37:17
Right. Talking about security and
37:19
governance, we're a week out of that.
37:21
The the Mythos kind of release and the press release around
37:25
that and everybody's super scared about 0 day events and
37:28
what's going to happen to cybersecurity.
37:30
What's your take on Mythos? I think it's still the model has
37:36
been in the racks, few, few people have access to it.
37:41
But the findings that are coming out are very, very interesting.
37:44
For example, one of the findings was, is actually finding
37:48
vulnerabilities better than 99% of security researchers.
37:53
And in the security space, finding vulnerabilities is
37:56
actually one of the most complex and intelligent tasks.
37:58
And that's why those folks are called researchers and not, not
38:01
something else, right? So you have to actually analyze
38:04
the software very deeply to figure out those
38:06
vulnerabilities. Now, if Mitos is doing it, it's,
38:12
it's it's like I'm completely blown away.
38:17
Then it's just a matter of time what else it can do.
38:20
And we are looking at it from the protection and security
38:23
perspective, from finding vulnerability side.
38:25
But now think about the other side.
38:27
If hackers also have access to micros, how crazy it can get,
38:32
They're going to do like super sophisticated access.
38:36
Yeah. And how we were talking about,
38:38
right, everything is going to be 10X.
38:40
We talked about, we saw with the software development, we talked
38:43
about data engineering and data work projects, 10X, right?
38:46
Like next level is going to get unlocked people.
38:49
The same freaking thing is going to happen on the security side
38:52
also your attacks are going to be 10X super sophisticated.
38:56
And at the same time, the systems you are going to need to
38:58
defend going to be next level as well.
39:02
It's good. Oh, really?
39:03
I suppose the one thing that helps me sleep at night is
39:05
they're saying that like, Mythos is so expensive, like the, the
39:09
the tokens are so outrageous that it's like there's no way
39:14
right now for that to be a commercial model.
39:16
So I think really they're just kind of teasing information to
39:19
get the proper infrastructure in place to, you know, hopefully
39:23
they can begin all the cyber stuff before all the attacks
39:26
come because, yeah, it's about to get to Star Wars through the
39:29
Internet. It's going to be need some
39:32
Jedis, you know what I mean? Yeah, yeah.
39:35
And my perspective is usually 'cause it's just matter of time,
39:38
right? And usually when the new
39:40
technology launches, it's expensive, it's exclusive.
39:45
But in a year, you know, everybody has access to it.
39:48
It comes down, it becomes commoditized.
39:51
And by then, you know, super miters comes on the horizon.
39:55
That's that's how the cycles work.
39:58
Yeah, it's a wild time. OK, Aside from Ultimate, what
40:03
are, what else are you doing with AI?
40:04
What kind of little projects are you running?
40:06
What kind of experiments are you doing?
40:07
How are you having fun with it? So as a start up, it gives us a
40:11
lot of flexibility and freedom to move very fast.
40:14
Even within our teams, we are working on being completely
40:18
active company. So a lot of these roles are
40:21
pushing together right between the designer and product
40:23
manager, even between software engineer and product managers.
40:27
Our product managers are going to keep creating prototypes.
40:29
Now the engineering team is taking and implementing and not
40:33
just on the technical side, even on the business side of the
40:36
things. On our marketing side, sales
40:39
side, we are using AI very heavily.
40:41
So even on the sales side, all the different roles are coming
40:44
together between you see business development people,
40:47
your account executives, even how we are reaching out to
40:50
people to our sales cycle or how we are showing them the devil.
40:54
We are nopping in AI so heavily and we are trying to automate it
40:58
so that we can operate as a company of the size of 500.
41:02
And next thing, how we can that work for the sales and marketing
41:05
side also whether it's cottage generation, whether it's running
41:08
campaigns or even it's just doing demos to our prospects to
41:13
see it, which automate. Describe the the feeling with
41:18
the team and there's there. Enthusiasm is through the roof.
41:23
Creativity is building. There also is this like kind of
41:27
terror of the mythos and the hackers and all this other
41:30
stuff. How is that all coming together?
41:32
What's culture like? So at least in our teams, since
41:37
it's a very close knit team, we have been able to educate people
41:42
that hey, AI is the next thing, either you adopt AI or somebody
41:47
who has adopted AI is going to come and replace you.
41:50
And I have heard stories from big companies, they tried to
41:54
educate everybody and still there were like 10% of people
41:57
who never listened and never believed in AI and they had to
42:00
do some restructuring, etcetera. He's in our company.
42:02
That's not a problem. Everybody's bought in.
42:05
And we did really could show them of convincing people on why
42:09
AI is very important and people are understanding that almost
42:14
every role is going to change fundamentals and then AI is
42:17
going to be an integral part of it.
42:19
So we are seeing actually a lot of openness on adopting it.
42:23
But the challenge is not everybody is at the same skill
42:26
level when it comes to adopting the AI.
42:28
So we are doing internally these training sessions and workflows
42:31
that we are standardizing at the company and peak level.
42:35
And then at the individual level, we are helping people to
42:38
adopt them rather than every individual trying to figure it
42:40
out on their own. That's that's helping us out.
42:44
Incredible Pradesh. This has been so much fun.
42:46
I want to give a shout out to the Humane Society of Silicon
42:49
Valley, hssv.org, because I know that's a important organization
42:55
for you. And did you adopt A dog or Kitty
42:57
from that organization? My my dog is adopted, but not
43:02
from that particular organization.
43:04
But I used to have all that deer in the in the organization to
43:08
take care of some of the pets and dogs, etcetera.
43:10
It's it's very close to my heart.
43:12
And that operation is amazingly run by the people responsible.
43:17
And the best thing about their website is they have like the
43:19
little yearbook. You can see the little puppies
43:21
on there. And there's one right here named
43:23
Sam who looks kind of stumpy with like a like a tan and black
43:27
face. Very cute.
43:29
So shout out to them. This has been amazing.
43:32
It's ultimate dot AIALTIMATE dot AI Pradesh.
43:37
What's the best place for people to connect with you personally
43:40
online? The best place is LinkedIn, so
43:45
if you want to stay connected, just send me a link to request
43:48
and I tend to post a bunch of things there.
43:51
I'm I'm wrapping up my ex Twitter game as well, but let's
43:54
stay right now. Let's stay connected all
43:56
LinkedIn. You got it.
43:58
I will put all the links into the the bio here.
44:02
So wherever you listen to this podcast or watching it, just
44:04
look down and and connect with Prodness.
44:06
This has been amazing. Congratulations on such a really
44:09
cool company. Thank you so much.
44:12
It's it's been a pleasure being here and discussing amazing
44:15
things with you. Really enjoyed the time.
44:19
Every week I sit down and ask myself one question.
44:22
What do founders actually need to know about AI right now?
44:24
Not the think pieces or the stuff that sounds impressive,
44:27
the real useful, put it to work today stuff.
44:30
Then I write it up, send it to you in your inbox, free.
44:33
That newsletter is called AI for Founders and I built it because
44:36
I was tired of. Wading through hit pieces and
44:38
people. That are afraid of the future.
44:39
Each issue I breakdown the tools worth your time, the strategies
44:43
that are working, and the ways you can use AI to grow your
44:45
company without burning out or getting left behind.
44:48
Thousands of founders read it every week and I'd love for you
44:50
to be one of them. Subscribe for free at AI for
44:53
founders.co. 3 seconds to sign up and I promise it's worth it.
44:58
Every founder I know is posting on LinkedIn, grinding on X,
45:02
maybe gabbling in short form video.
45:04
And The Dirty secret nobody wants to say out.
45:06
Loud is you don't own any of it. One algorithm change and your
45:09
audience evaporates overnight. The only audience you actually
45:13
own is your e-mail list. That's why the smartest founders
45:16
in the world are building newsletters, and Inbox Alchemy
45:19
builds it for you. They create your weekly founder
45:22
newsletter, design a beautiful newsletter website, and grow
45:25
your audience to 25,000 subscribers in the first six
45:28
months. All of it is managed so it
45:30
drives real pipeline. It builds your reputation.
45:33
It nurtures the relationships that turns into deals,
45:36
partnerships and opportunities you can't even predict yet.
45:39
Social is rented. e-mail is owned.
45:41
The founders who figured that out early win for decades.
45:44
Go to inboxalchemy.co. That's Inbox alchemy.co and
45:49
start building the audience that's actually yours.
45:54
While you were recording your last episode, Casey Taylor was
45:57
out there pitching your podcast to brands and securing
45:59
sponsorship deals. Well, she's not yet, but she
46:02
could be. Casey is Howdy's AI partnership
46:05
agent and she does something pretty wild.
46:07
She actually secures sponsorships for your show.
46:10
Not a match list that you have to work through yourself, not a
46:13
template you have to send. She handles the outreach to
46:16
follow up the whole thing and brings back partners.
46:18
You started your podcast because you had something worth saying.
46:21
Casey makes sure that you get paid for
46:23
sayingitgomeetcasey@gethowdy.com. That's GETHOW di.com. 2 minutes
46:30
to get started, then Casey takes it from there.
46:33
That's howdy, partnershipsgethowdy.com.