Palantir's Danny Lutkus demos AI-driven supply chain consulting: from vague problem to proposal in a day

the future into the pre-training data. So be very careful. Don't leak anything. But uh but but introduce yourself. Tell us what you're going to show us. Yeah. Absolutely. Uh microphone. Oh yeah. My bad. Uh what's going on guys? Um my name is Danny. Yeah. Let's see here. Uh I'm an engineer at Palunteer.

I've been a Palunteer for about 12 years. In terms of like my role It's hard to describe. Like I'm sure everyone at Palunteer said that. Uh I guess like if I had a role or a title, I I uh do a lot of our business in the Midwest at this point. So first six years at Palunteer, I was on the government side.

I did work with Department of Justice. Yeah. US Special Operations, CIA, National Counterterrorism Center. Sure. After my wife and I had our first kids, she was like, "Hey, could you not go to weird places in the world anymore? " And I was like, "Totally reasonable. " Yep. Reasonable request.

We we moved back to the Midwest and I switched over to the commercial side and that's kind of like what I do now is like grow our business in the Midwest. Yeah.

What what's like a what's a uh like just line drive uh solution that you like just total wheelhouse uh solution for uh you know I imagine like a large enterprise customer in the Midwest. Yeah. Uh what I focus on a lot is manufacturing in the Midwest.

So you can like there's huge manufacturers in the Midwest whether that's like Johnson Controls or Eaton or um Molson Kors uh Cummins Engine. So it's a widgets factory. Yeah. They're making widgets. They're buying parts. They're assembling them and you have to understand the flow rate.

Where's the where's the rate limiting factor? How can we increase flow?

This is where I think we have the most differentiation from a product perspective because it's like like I can actually affect the physical world and then I can measure how I affect it and then I can learn and improve how I affect the physical world the next time right whether that's like hey I'm in supply chain and I'm short on inventory like how do I solve that problem in the most effective and optimized way versus like I'm trying to manufacture something and like how do I make sure my machines are running I have the right labor I'm trying to do the right thing and so like the the real magic behind all this too is like these yes they start off as like singular use cases that are like pretty great like straight shot but then like when you start to connect these workflows together and it's like oh the machine's down like and I have this material like what do I do and how do I go do it uh what do you want us to show us today I can kind of hold this for you if you want we getting good sound on this okay cool yeah walk us through it what I was going to demo is I think like one of the interesting things and I'm I'm sure you've like talked to a lot of different palunteerans today is like we are never going to purport to be like a strategy consulting type of thing when we engage with customers like we're never going to purport to be like oh like a hey we're experts in x y or z and the great thing about that right is like we're true to like who we are.

Or the bad thing about that right is like companies will identify and the organizations that we ident like that we work with will identify like hey I know this is a problem right but like there's a huge amount of time between like hey there's a problem and then let's go like implement a solution and the dependencies on actually getting to that faster are like uh I have the internal SME that can actually like understand the problem and come up with the right solution and do the feasibility and all that great stuff or I go work with like strategy consulting, I pay millions and millions of dollars to get a deck that tells me like, hey, this is the solution that we think you should employ with the right like ROI in this approach and we've done this feasibility study and we think that you should go do that.

And so like we find that as a huge impediment to like our own growth, right? Like why should I wait months? Yeah. You don't want them to go spend millions of dollars with some random group to then recommend a palenteer product. That's 100% right.

And so like what we've been exploring more is just like well why can't I use AI to do that?

Like why can't I like give a fairly haphazard business like a a description of business problem and use agents essentially to like structure that into a better business problem description to do the necessary research about like what are the potential solutions of of of things that I could and should deploy to go solve this problem.

Can I generate ideas with all the requisites of how I actually employ those ideas and actually generate a proposal where then I also have like agents as critiques on that proposal to be like is this technologically feasible? Is this like financially feasible?

All the things that you would expect like strategy consultants to do for you like I should just be able to do that in a day and come up with a proposal. But then like I don't know if you guys have talked to anyone about AI FTE, but then like I should just then be able to use the output of like this to then go build it.

Yeah.

Like I should just be able to say like cool here's the solution I need to go build input into AI FTE build it right and go from like you know what would have taken six or nine months until we ever get engaged to like well I think I this is a problem like let's just go do it like in the next week right okay does that make sense yeah it makes sense um I I have some follow-up questions but maybe maybe jump into the demo first cool I think like like my my immediate I guess question maybe it's relevant uh is like how do how do you ensure kind of quality right because like you didn't say this but like someone else in another context might call this like vibe coding right sort of like generating like a deep research report on like a problem and a potential solution and then like you know sort of prompting your way to an implementation totally and uh today uh you know just like code quality and product quality ends up popping up.

But I'm sure that you're already think about that. Like my take on this is like when you start doing anything with AI or large language models like it there has to be a human in the loop, right?

not only to make sure that quality is coming out of the other side but also to ensure feedback loops are occurring and right and then and then you can take that context and start getting closer and closer to a Jesus take the wheel moment where like um where like you actually have built trust because like part of this is not actually like I think a technology problem it's like a people and process problem where like people actually build trust in it and also you get all the tribal knowledge that's not in any system um actually incorporate in some knowledge context that you can start to build off of over time.

But I think that's like that's the that's the trick is like humans always have to be in the loop, right, that to begin, but then like you build trust until you actually do the Jesus take the wheel moment. Yeah.

So yeah, with this demo, what is the uh is it designed as like an internal tool or something that you would actually? A lot of our customers are starting to use this to start to shorten the the the cycle time of going from like initial problem identification to implementation.

So like and is that for is that for customers that are already using Palunteer? Yeah. Um so like we we've started using this primarily with like a lot of existing customers, right?

But then the cool thing about it is I don't know if you guys have heard where like all of the things I'm going to show you are kind of like native components of the platform. But then we've developed this capability where we can say like hey this is actually a really repeatable workflow.

What if we package this up and then just it's way easier to deploy where we can just like deploy there deploy there deploy anywhere basically. Uh cool. Yeah. So walk us through pull it up and maybe bring it a little bit closer so you can see it. You share your whole screen. Oh yeah. Yeah. Go ahead. Ready? Yeah.

Okay, let's do it. No saying text messages or anything like that. All right, cool. Um, I used to work in the aviation space a lot and I fly in and out of Newark. Um, which like if you guys do that, you know that's a real pain in the ass. Yeah. So, let's let's start there. Let's just say like um redesign.

Hey, I'm a um Oh, yeah, for sure. Go ahead. So like the problem the problem that I'll type in basically is like hey I'm an aviation expert like um we're seeing significant delays around like Newark airport because there's not enough runways and the runways are too short. Uh like what should I do to optimize my flow?

Okay. Basically to to solve this problem. Sure. So like uh you now you guys get to see me type which is always fun. Yeah. This is interesting. Um I yeah a ton of questions about I've always wanted to redesign the LAX uh like uh like streets like uh the flow of traffic.

Yeah, that is a wild choice by LAX just constant constant traffic. Uh wasn't too bad uh this morning fortunately but we did have a funny incident with a member of our team who uh first day John arrives got through security. Oh yeah.

and almost managed to miss his flight because he was getting a breakfast by a former guest and friendly. I would call I would called and texted and said you know this is no time to take shots at the dyslexic.

He had missed he had he had made a mistake and and confused uh gate nine for uh for gate six right and and there is no gate six at this particular terminal headed to a different terminal. Thank you for covering. So everyone Yeah. So it doesn't have to be.

So right now I just I typed in I like pretty rough problem statement. I'm an aviation expert. I want to solve problems around EWR airport. Yep. Uh there are too few runways and the runways are too short. How do I optimize traffic flow around it to minimize disruptions? So that's kind of like the first point.

And what's happening here is like the first set of agents is basically taking that as a problem description and actually like putting more structure around it.

So it's not like my um you know my like missper effectively working right and so you on the left side of the screen you can actually see some of the logic of like what happened the train of thought here of like hey here's the problem statement I can see the system prompt like what the task prompt is what the LLM like responded to when they saw this to them actually then creating and structuring this problem which is like hey the core objective is I want to optimize air traffic flow around uh Newark Liberty International Airport to minimize disruptions, de delays, and efficiencies.

It puts out like key requirements. Yep. Like prioritize aviation safety standards. It gives out restraint uh constraints. Nathan Fielder would be happy to hear that you're Yeah. Yeah. Right. Um it gives out constraints like limited number of existing runways, restrict uh simultaneous operations, etc. , etc.

So like this looks pretty good to me like as the initial problem description. um way better than like the garbblegook like two sentence thing that I did.

So now I want to like start to get into the phase of um like actually starting to do research on this to say like what are potential tools, what are potential approaches to actually solve this problem. Yep. And so what's happening right now is like now we're going into kicking off into more of like an agent. Yeah.

Just branching a bunch of agents to go do deep research. So yeah, exactly. Yeah. So like now on this screen I can see that same like core objection uh objective function over on the left what it's working towards. Yep.

And then I can start to see as it's running on the left like research topics as it's doing research popup and modeling. This is all built in like native foundry tooling. Sure. Um how how inferenceheavy is this? Because it feels like it's going to town right now. Yeah. Yeah.

I'll I'll show you I'll show you kind of like the under of how we're actually doing the research. Yeah.

It is a unique uh it is a unique like like I don't know like problem set because it's like going to town is something we worry about when we're talking about like oh yeah you have a billion consumers and $10 really adds up. Yeah. But if it's like a problem as important as this.

If you're talking about if you're talking about you know optimizing an airport I think I can I think I can deal with a $100 inference bill. You know I'm gonna be okay with that for sure.

Um, so the other thing that I think is interesting here is that like I think agent is like a very there are a lot of definitions for what an agent is.

I think at this point in time like one definition is like uh and this was like kind of our first approach was like hey let's let's build a set of logic that an LLM actually orchestrates different parts of that logic between and it can use tools like you know deterministic tools or it can write back uh or it can access and query things to ultimately do some type of automation.

I think the other definition of like what an agent right now is like more of a chat interface. Um and then in that regard, right, like I want to be able to give that um chat interface like access to tools. Yep. Right. And so in this case, like what I've given um the agent access to is a bunch of different tools.

First, like I can see the model that I'm using behind the screen here. And like for our from our perspective, like we think the models are mostly like commoditized at this point. There might be certain models that are better at different things.

And you actually probably want to use these things interchangeably and actually have an evaluation framework that based on the task that you're asking it to do will like select the right model for that particular task.

Uh but in this case right I'm using gro 4 and then like for the tools in particular like I've given it access to like conduct research.

So I've given it some ways in which it can actually reach out and use different either internal or proprietary information uh of the organization that we're working with or reach out and use something like perplexity to do like more AI based search.

I've given it the ability to like generate like create code blocks if it's like coming up with an ROI and it needs to do napkin math like I want to say like I want you to allow you to actually like generate the code but also then run the code to see like what what the result is.

And then I mean it seems like all of this is all of this is kind of like frontier level but available broadly but the palunteer that you actually have like data that isn't just available on the web and so like if I'm actually an airport and I actually have specific data about you have the thing stands out to me is like if you're a large enterprise you want to work with with foundry and and have that ability to be model agnostic and like where does the leverage flow in that situation when Foundry can just sort of decide on the fly what what form of intelligence do I want to use for this problem set.

Very cool. So I can see like kind of like the train of thought on the right like what it's doing. Um and so it's going to go it's already using the um research um kind of tool and you can already see the research topics starting to like pop up here.

So like this is an example of an application right that like a user would use. They would they know nothing about Foundry, right? they're logging in to an application. Their job is to like go do this thing, right? But then behind the scenes, you have a lot of different options for how you're setting up this logic.

I don't know how much you guys have seen Foundry, but this is an example of what we call AIP logic. I could write all of this orchestration in code if I wanted to. I'm fairly lazy, so I use the lower code tool, which is AIP logic.

And so here, I can just like set up a bunch of different orchestrations for how I want a function to run. In this case, I'm I'm putting in inputs for what I want the query to be, which is around like that problem statement we talked about.

And I'm setting up functions for how it can like reach out to different types of sources. So like the first one is like if I had internal kind of like proprietary information on schematics of a runway or planes or what types of runways planes can land on, things like that.

Like that's all information that then I can make available to the LLM to go do a combination of like semantic and keyword search against it to find the right information to go do research against. But then like as a backfall then I'm just like also giving it access to go and query perplexity, right?

And go say like hey go find what's out what else is out on the internet to actually go do this research about this particular problem, right? And then bring that back. And then the last part of this is like an action then to like go capture all that information and store it back into the ontology layer in Foundry.

Awesome. So this is kind of like what it's doing live is like um it's still working. It's working like and it's and it's writing as we like as it's doing research, right?

So it's like what is the current runway configuration, operational capacities and key limitations at EWR including details on runway lengths numbers and how they impact uh aircraft operations. Sure. And so then it actually gives me like this is this is pretty good information.

It will site the sources where it's coming from and everything like that, right? Yeah. What are effective non-infrastructure strategies for optimizing airport throughput, right? Uh and so in this case, right, it's actually saying like, hey, there's this performance-based navigation as a cornerstone, right? Yeah.

I remember hearing that if you if you have the plane board from the back to the front, it'll load way faster, but no one wants to do that because the it's a it's a business model thing. Yeah. because people pay to be at the front of the plane and they want to get on the plane first.

But if uh there was another proposal that was like uh load all the passengers that have window seats, then all the passengers that have middle seats, and then all the passengers that have aisle seats, and they all kind of just flow in.

Um no one's quite figured that out, but yeah, I mean, I can imagine that it could come up with a bunch of different proposals for uh you know, similar just kind of like rethinking of this the flow of traffic. I think we're getting short on time here. One question one me let me like zoom forward.

I'll show you kind of like an end product here which is like let's go I already ran this today. I was like hanging out with the American Airlines guys cuz like we were making fun of EWR which is not their hub.

Um but yeah, this is like an idea that it generates and then like I get a summary of what that idea is and then it automatically develops critique agents that are like looking and evaluating on different type of like uh different criteria right which is like hey can I what's the risk assessment and mitigation evaluation what's the economic feasibility of actually doing this like what is the safety and regulatory compliance evaluation and then it's going to run like those evaluations using that agent as a like a task criteria to actually then say like I can see the the guidance that we gave the agent right and its task and then it has to go evaluate to see if it makes sense from that perspective y right and it even like generates its own models and its own code to say like hey is this feasible from like can I do basically nap uh like napkin math and say like can I come up with like how I could calculate this and actually go and like run and see how close Does this output do you think to what a larger it's like pretty I think it's like pretty aligned right because like they're not they in in normal times like these strategy consulting firms aren't getting access to all the data and so they're like being like okay come up with the idea do the research generate the idea for a little bit then like I need to do some napkin math on like how I would think about actually like critiquing this idea and then ultimately like I need to come up with a proposal right and here's like the end proposal for what I think you should go do same framework where I have agents then writing portions of that proposal and then uh from there right it's just like copy paste that proposal in the AI FTE and like start building right last uh last quick question are you feeling the reindustrialization yet are you seeing new entrance into the Midwest building things or is it more legacy players just trying to trying to increase I think it's legacy a lot of what I work with are companies like Eaton which are like hundred-y old companies or like Johnson Controls 100 euro companies um that are saying like how do I actually use this as an advantage to do to do better right like and and that's like where I think is interesting is that like maybe five years ago this was really hard like people were like yeah I don't trust it or I don't believe in it I think now what's interesting is they're like I trust it let's go like it's just you can give them you can sit down and give them a demo that's right that's Well, thank you so much for coming on.

Thanks so much for joining. Thanks for having me, guys. Brave to do a live demo next guts. Great. Hey, great work listener. So, thank you. Love it. Have a great rest of the conf. You're the man. And we will bring in our next guest, Jonathan Web from the nuclear man himself. Welcome. Sorry to keep you waiting.

Good to meet you. I'm John. Today is a great name to have a company that starts with the I don't know if you saw the browser company. The free press sold for $200 million. The browser company sold for uh $620