Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AI

0:01 [music] >> Hello everyone. So, my name is Ayush Bhardwaj and I did applied AI for a hedge fund. And now I do everything tech plus applied AI for a

0:23 pharma tech startup cuz you know the way startups are. You have to do everything, wear multiple hats. So, before I start the session, I would like

0:31 to do a small survey. Can I get a raise of hands for all the engineers in the room? Okay, that's a tough room. Now, can I get a raise of hands for managers?

0:41 Okay, just to be clear, managing AI agent does not count. You have to manage people. Okay, we have few managers as well. Interesting. So, they will help me like

0:49 fine-tune my talk a bit. So, today my aim is to take you through the journey of how do you actually build and iterate in applied vertical AI? And my experience is from the hedge fund

1:00 and the pharma tech company. So, before delving deep into the recipe, I'll just like take you through what do I even mean by applied vertical AI cuz I don't know if it sounds like a very weird

1:10 term. It's like the vertical word is kind of forced. I won't lie, it is. I coined this term probably. So, applied AI is like built for So, what applied vertical AI is essentially applied AI

1:21 but built for one very specific industry. It's It's aim is to simulate a job of a person in that particular industry in a sense. So, an example of applied AI is

1:31 Google Translate which is like general purpose, helps you translate. It could be used in education tech and it can have like tons and various sorts of uses is whereas Elos, which is my employer,

1:41 the pharma tech company, we specifically build drugs with AI. So, that's a very specific use case. Another examples of applied vertical AI field could be the legal tech firms that

1:50 are now coming up with. You must I'm sure you must have heard about them. So, those are like another the examples of applied vertical AI. So,

2:00 when I left the hedge fund, right? So, I was expecting that the world would change for me cuz you know, hedge funds are like really fast and really pressure sensitive. Whereas, pharma is like,

2:09 "Okay, we're going to take 15 years, but we're going to do it right." Hedge fund was all about like, "You need to do it fast and mostly right. It does not matter if we lose at one paradigm as

2:18 long as we are overall winning." Whereas, a pharma firm is like, "We have to be absolutely right. You can take a week more." And it was true. It's it's a completely different world. But,

2:29 to your surprise and to mine as well, nothing changed, actually. My job increased, but the core part of my job, applied AI, remained the exact same and I cannot uh

2:40 express how surprised I was cuz I thought that it'll be a complete different thing, but apparently it was not. So, uh So, then I spoke to other people as well

2:49 across legal AI and the people coming up with the prop tech firms, which is essentially the real estate tech firms. And I realized that everyone is kind of building the applied vertical AI in a

2:58 very similar way. I could see some steps that could be essentially abstracted out. And that's what we'll do today. So, uh before again delving the deep into that, I received a few reach outs

3:11 saying, "Are people actually putting agents into production?" And I was like, this is such a wrong question to ask. Everyone is putting

3:18 agents into production, even like 15-year-old 15-year-old kids these days. The question to ask is whether they actually work, whether they actually make or save

3:26 money, whether they justify their ROI, whether uh they're making way more than the amount we are investing into it like end-to-end. And I can say from my anecdotal experience, yes. At the both

3:38 places I worked, the agent either saved the money or made more money. So, with that, let's get started. So, the recipe I'll take you through a

3:49 series of seven steps, roughly, and try to like make this process as simple as possible. So, the first step is formulate the problem. So, this is sounds like very trivial, but a lot of

4:00 people, specifically startups, get this wrong. They just try to do too much at once. Whereas, from what I have learned and what I think a lot of colleagues would agree, you need to pick a very

4:10 narrow task. You just cannot ask it to do everything. A good example for this could be, let's say if you build something in finance, you won't ask it to like, "Hey, can you fetch me top

4:19 three market opportunities that I could invest in?" No, that won't work. You have to be like very specific. Like you pick a market, you say, "Let's take the US equities." Then you pick an

4:28 industry, let's take IT. And then you ask it to like rank stocks based on some parameters like capital expenditure or let's say the AI um uh investments. So, you pick like very

4:41 specific things, and then you uh sort of formulate a very narrow job for the AI agent to do. And you can build like n number of AI agent. Last I checked, there was no tax on building

4:50 more AI agents. So, why do you want your single agent to do everything? So, this is important, and this is in the same uh in the pharma context is the exact same. We just break down the

5:00 process into steps, and then ask really pointed questions with the agent. We model our agent for a task. So, once we have our problem right off the way, we know what we're trying to

5:09 solve, the next step is identify the data. And I cannot stress this enough. This is a really, really, really important step, cuz everyone has news data. Everyone has like seller side

5:21 reports from JP Morgan, Morgan Stanley. Uh everyone has the arXiv preprint server or PubChem or your research papers, right? But what actually makes your application better than let's say

5:32 ChatGPT or Claude? It is your proprietary data. So, the thing with proprietary data is it's really expensive to buy, and most people won't sell it to you. So, you

5:41 need to curate it by yourself. Imagine your organization has been working for 3 years, right? They already have a lot of data. It's just unstructured. And in the age of LLMs, I think this is a very

5:51 fairly easy task to make unstructured data into structured data. Like a LLM workflow could do it overnight. So, to give you a great example of the proprietary data that finance industry

6:04 has, it's the trade thesis, which is like what trade work and why it worked. And in pharma, it is the data for failed experiments. For successful experiments data, yes, you can get it, but failed

6:16 experiments, that's relatively hard to get. So, now we have the problem, we have the data. What's the third step? That is to model the problem, like write the prompt.

6:27 So, while writing prompt, you like what we should aim is to model it after the person who you are trying to replace. I mean, that's the hypothesis, but yeah, no offense, we're

6:38 not trying to replace anyone with AI, but that's the ideology behind writing prompts. Encode how a person would solve this job into multiple steps. So, it's just like a like a mental model. So,

6:48 this is again fairly simple. Next thing, observability, I'm sure you have been in this conference at 3 years and this word, I think I don't know, you'll be hearing about like a thousandth time.

6:58 There are tons of observability provider. If you can't see it, you can fix it. So, you need observability to see the traces, understand what your

7:08 uh AI application is doing, and debug it. So, sorry, but all of this was the easy part, to be honest. All of this fits one screen. The mythical 10x engineers can

7:19 do this stuff in minutes. Like literally, this is the code you precisely need to build an AI agent. So, that's why it's not the moat. Uh of course, except your proprietary data.

7:29 So, what do you do now? What do you do after doing the first four steps, which is observability, and prompts, and like uh getting the

7:40 data right, and everything? UI trade. Now, the thing with iteration is like when I joined the hedge fund, I thought how hard it can be. I mean, everyone can iterate. I mean, we have been iterating

7:51 our whole life for each of the task. But, to be honest, I could build it, but I just could not tell if it worked cuz

8:02 I'm not a trader. I'm not someone who has a PhD in biology or chemistry. I just don't understand what the model is saying, what is the output of my AI agent is. And since most of you are

8:13 engineers, you would relate. You can instantly tell that Sonnet 5 sucks because you have your own training. You understand, okay, this code is not great code. Whereas, some X model, let's say

8:24 Fable 5, you see, okay, this is great but not as great as the high base cuz you've been trained for this for life. You have a mental model to judge these things.

8:33 But, you just do not have the same kind of mental model when it comes to like predicting trade thesis is or doing like really specific task that vertically our industry does. And this is also the

8:45 place where like a lot of vertical AI projects quietly die because on the surface it looks like you have made it, you have built it, let's put this into production and start selling it. But, no

8:55 one would buy it the same way you won't use an inferior coding model. So, as an engineer when I ran into this, I just couldn't accept honestly. I thought, no, there's certainly more that

9:05 I can do. We don't need other people. So, I thought I could LLM as a judge my way out of it. >> [sighs and laughter] >> And this was a really, really stupid

9:16 mistake to be honest cuz what LLM is essentially doing, it's it's predicting the next probable word. So, if you see, it's just like jargoning its way out. It does not understand what alpha means. It

9:26 does not understand how to actually create value unless you have like taught it some way. And whereas a human can just tell it instantly what's and what's not.

9:36 So, I'll just try to dwell a bit more deeper on why you can just iterate. So, first thing is that model cannot verify itself, specifically in these

9:46 fields, because reinforcement learning via verifiable rewards is really good at math and code because you have like answer keys, you can verify your code is uh compiling or not, and there are tons

9:56 of stuff you can just model uh the complete thing around this. But, when in these fields, there is just no way to model it. And And let's say if any error gets in, it's just compounds with every

10:07 stuff. And that's what LeCun seems to think as well. And now, the more important part that we touched upon previously, the data. So, the interesting thing with pharma

10:19 and finance is the data was never there. And I'll explain to you why. So, any institutional manager holding over $100 million in qualifying US equities are forced to publicly file

10:35 their holdings, long position holdings, every quarter. And once a hedge fund does this, this is the percentage decrease in their returns because everyone just sees those

10:45 reverse engineers and takes away their moat. And when it comes to pharma, right? So, this is the number of uh so, by law, you are like required to disclose every

10:55 clinical trial pass or failure you have done. But, 30% of the funds, which is like nearly 1/3 of firms, never do. And in like 2026, FDA had to like publicly remind over, I don't know, about 2,000

11:07 sponsors that they are, I mean, doing injustice by not uh releasing unfavorable results because this is the exact data which helps the model thing, which helps your

11:18 LLM actually reason through these complex and niche industries. And they hide it because for them, it's like a chicken laying golden eggs. Why would they sell their chicken? So, naturally,

11:28 neither OpenAI nor Anthropic has that has this data because it's like gatekeeper. You just cannot hire a trader for $100 an hour and have them annotate that stuff because there's like

11:37 lots of NDAs and they definitely earn more. So, okay, now I have told you about tens of problems, right? Now, you would naturally think, okay, yeah, right, then what do we do?

11:48 How do we build a startup in like a vertical space space? So, very self-explanatory, you hire the person who you want to sell it to cuz

12:00 there is, to be honest, no other way around. I have tried a lot of stuff. You just need to hire the user. In finance in a hedge fund, this was very easy because the user was kind of

12:10 like my boss, the trader. We worked together, but in the PharmaTech startup, it was very weird. We were like a bunch of young engineers and we were like, oh, we need a 20-year-old scientist in our

12:20 company to tell us what to do? Yeah, I guess we do. And then we hired someone, right? And that someone actually changed the trajectory of our tools. Our tools started making sense. When we pitched to

12:31 the other pharma companies, the big ones, the big pharma, they started liking our tools because it's kind of spoke their language versus the normal jargonish LLM language. So,

12:42 once you have hired the user, let's say, then what would you make that user do? You try to build a learning loop out of it. The domain expert can start at the like

12:52 a very, very low level, the ground level, where they just think about prompts. Okay, yeah, I mean, let's not ask LLM to do this. Let's ask a very

13:01 specific query again. They'll help you curate data. Just like engineers know which conferences are which are not, which research paper sites are great, which are not, which are like top

13:12 leaders in engineering, which is which are just like influencers. Similarly, a pharma expert or let's say a trader knows which sources are more reliable than the other. So, they help you create

13:21 their data. They help you like refine your prompts better, and they try to create like thinking models of how they would think about a problem. Cuz I mean, let's say if you if you follow five

13:32 steps to solve a problem, right? You just cannot do it in in any random order. There has to be a logical flow. There has to be a natural flow. That's So, that's what they

13:40 uh try to curate like decompose a problem, gradually refine, and then finally judge. So, the person who sort of has lived through the complete of the industry that they're

13:52 trying to revolutionize, their judgment is now like turning into agents. So, that's what's happening behind the loop. So, uh to do this there are like again multiple ways. I mean, each of these

14:03 could have been a hour-long session on its own. And I wish I could take, but these are like few ways that I identified. Uh I'll just like take uh you through them like really quickly in

14:14 the interest of time. So, supervised fine-tuning I think most of you would know where like model mimics human nest demonstrations. Uh reinforcement learning from human feedback is like a

14:23 kind of uh a very efficient way where human preferences train a reward model. Then rubrics as a reward is I I like to call it reinforcement learning from AI feedback.

14:34 This is because that you can human can just create a rubric, and then AI will just like grade itself based on that rubric, and that improve its own processes. But again, there is a slight

14:46 chance that you might run into an echo chamber with rubrics as rewards. And the cheapest of all, and I think the highest ROI is the error analysis. Whereas the observability part that you set up

14:56 earlier, you just analyze the logs plain and simple. You understand where model is going wrong, and then you just try to correct it. So, this is where you have like don't have

15:05 to touch any weights, and the most highest ROI way to get the impact from a like start on. And once you understand like uh what more you could do, or if error analysis is solving or not, you

15:18 can just gradually climb up the ladder, and probably uh later on go to the ultimate reinforcement learning from human feedback cuz that's I think in our industry kind of the golden standard

15:28 these days that you need to do RLHF to actually get some edge. But uh certainly there are some pitfalls of it. Like for example, now there's GLM 5.2, right?

15:39 You fine-tuned it, right? Uh Alibaba Cloud or let's say Deep Seek will release a newer model, then you have to fine-tune that too as well. So there is a cost. It's not cheap.

15:49 So once you have done all this, you just create a loop and you just like go on to that loop. You hired one user, you hire more users, they ask more queries, the scoping increases, the data increases.

16:00 At this point you're kind of generating your own data. The exercise you have been doing in loop, right? That exercise itself is generating a very I would say a crazy data set of what works and what

16:10 does not work. And this loop never stops. Once you feel confident enough in your application, you just ship it, provide it to the external paying users, and then you see the magic of it that it

16:20 actually works. So I just pulled the stat from Stanford AI Index report cuz it's a really nice report that gives you an idea of what the state of AI is. And this says like

16:31 80% 89% of enterprise AI agents never reach production. Again, I disagree. Every AI reaches production, but it just fails to work or like justify its own cost. So that's the real thing. You can

16:43 just build and ship AI agents whenever you want, but you need to justify ROI. And finance and pharma are two such industries where if it does not make money, it's shown the door.

16:54 Simple. They won't like wait and say, "Okay, maybe it'll work in 2 years. Maybe the cost will be lower in by the third year." No. It has to instantly make money. It has to like hit the

17:03 ground running. And if it does not, shown the door instantly. So just to summarize the seven steps that I feel are like good enough to give you an abstraction of how the vertical AI

17:14 industry moves. You formulate the problem statement, you source your data sources, you prompt it well, you define those prompts, you observe how your tool is performing, you

17:24 don't iterate yet, you hire the user. And this user or users now play with the tool as much as possible. They like kind of form a learning loop, an endless learning loop that goes on and at a

17:36 point when you feel yeah, it's it's really delivering that alpha over let's say Claude and ChatGPT, you just ship it, you start earning money. So,

17:45 one more interesting thing. So, HITL is like kind of a thing everyone is like yeah, let's add human in the loop. I would say not yet. Finance and pharma are still those two industries where

17:55 it's AITL, AI in the loop cuz everything is like done by the expert, but the AI assistant really helps save time. Like for example, uh it may take an X amount for a trader to

18:07 form different trade thesis, and AI can just give him five candidate trade thesis, but which one would actually work in the market and which won't is the discussion the discussion still lies

18:17 with the trader. And same for pharma when you're like picking drug candidates, which one to pick, the expert still does it, but you just like reduce the time of expert by a lot lot.

18:29 So, and and it will stay this way for really long. So, uh for the models to actually make good decisions, they don't need to do correlation, they need to do causation.

18:39 And as Ya as Jan LeCun puts it, these are like text statistics, not real-world models. You cannot just pattern match with past and use future to predict to

18:48 it. And so, we are like kind of not there yet. That's what I call as the AGI line. Once we are there, yeah, probably then models will just like make drugs. You will have vibe coded drugs. Someone

18:59 would be vibe coding market, but yeah, not yet. So, a final takeaway that I would call if if if there's one thing you are taking away

19:07 from this talk, this is it. Model infra ecosystem, everyone selling you tons of stuff at this conference is just commodity. Everyone has it. If you have it, everyone has it. Everyone can

19:18 pay X number of dollars for a subscription. But, what is moat and no one will come and sell it to you. You won't have to curate it on your own is the domain expertise. You need your

19:28 data. You need other people's data. That is just not out there on the internet. And that that's what will form your moat. So, thank you for your time. I think you enjoyed the talk and yeah, let

19:38 me know if you have any questions. We can meet outside. Thank you. >> [applause] [music]