--- schema: 1 type: video title: >- Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AI description: >- Ayush Bhardwaj could build the agent. What he could not do was tell whether it was any good. He moved from applied AI at a hedge fund to a pharma tech company expecting a different world, and found the job identical, including the wall. An engineer glances at generated code and knows instantly that it is weak, because years of training built that judgment. Nobody on his team had the equivalent instinct for a trade thesis or a drug candidate. He calls this the point where vertical AI projects quietly die, because the thing looks finished and then nobody buys it. Judging his way out with a model was, in his words, a stupid mistake, since it jargons its way through without knowing what alpha means. Verifiable rewards work for math and code because answer keys exist. Worse, the data that would teach a model to reason in these fields is deliberately withheld. Funds must file their holdings quarterly, and their returns drop once competitors reverse engineer them. Disclosing every clinical trial is legally required, yet roughly 30% of firms never do, and in 2026 the FDA publicly reminded more than 2,000 sponsors. The frontier labs do not have it either. So his answer is to hire the user, which for a team of young engineers meant hiring a senior scientist, after which their tools started speaking big pharma's language. That expert curates sources, sharpens prompts and does the judging, starting from error analysis as the cheapest rung. The moat is never the model or the infrastructure. Both are commodities. Speaker info: - https://x.com/aybh08 - https://www.linkedin.com/in/aybh/ - https://ayushb.me/ Timestamps: 0:00 - Reading the room 1:04 - What applied vertical AI actually means 2:43 - Leaving a hedge fund for pharma and changing nothing 3:34 - The wrong question about agents in production 4:27 - Step one, make the task narrow 5:18 - Proprietary data is the only real differentiator 6:57 - The easy part fits on one screen 7:50 - Why you cannot iterate on what you cannot judge 9:30 - Trying to use a model as the judge 10:21 - The data was never there, by design 12:05 - Hire the user 12:55 - Building the learning loop around a domain expert 14:39 - From error analysis up to preference training 16:20 - Reaching production is not the same as working 17:10 - The seven steps 18:00 - AI in the loop, not human in the loop 19:00 - Your moat is domain expertise and data canonical: 'https://liquidgold.cc/watch/Yphdry8ttAQ' markdown: 'https://liquidgold.cc/watch/Yphdry8ttAQ.md' llms: 'https://liquidgold.cc/watch/Yphdry8ttAQ/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T05:13:23.937Z' language: null children: [] videoId: Yphdry8ttAQ durationSeconds: 1202 publishedAt: '2026-08-19T17:00:31.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 80 endSeconds: 170 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=80' heldReason: rights - ordinal: 1 text: null startSeconds: 80 endSeconds: 170 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=80' heldReason: rights - ordinal: 2 text: null startSeconds: 182 endSeconds: 214 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=182' heldReason: rights - ordinal: 3 text: null startSeconds: 225 endSeconds: 270 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=225' heldReason: rights - ordinal: 4 text: null startSeconds: 325 endSeconds: 336 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=325' heldReason: rights - ordinal: 5 text: null startSeconds: 336 endSeconds: 425 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=336' heldReason: rights - ordinal: 6 text: null startSeconds: 336 endSeconds: 425 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=336' heldReason: rights - ordinal: 7 text: null startSeconds: 472 endSeconds: 482 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=472' heldReason: rights - ordinal: 8 text: null startSeconds: 482 endSeconds: 573 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=482' heldReason: rights - ordinal: 9 text: null startSeconds: 583 endSeconds: 630 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=583' heldReason: rights - ordinal: 10 text: null startSeconds: 630 endSeconds: 642 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=630' heldReason: rights - ordinal: 11 text: null startSeconds: 642 endSeconds: 729 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=642' heldReason: rights - ordinal: 12 text: null startSeconds: 642 endSeconds: 729 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=642' heldReason: rights - ordinal: 13 text: null startSeconds: 739 endSeconds: 784 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=739' heldReason: rights - ordinal: 14 text: null startSeconds: 794 endSeconds: 883 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=794' heldReason: rights - ordinal: 15 text: null startSeconds: 895 endSeconds: 983 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=895' heldReason: rights - ordinal: 16 text: null startSeconds: 983 endSeconds: 994 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=983' heldReason: rights - ordinal: 17 text: null startSeconds: 1041 endSeconds: 1088 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=1041' heldReason: rights - ordinal: 18 text: null startSeconds: 1100 endSeconds: 1145 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=1100' heldReason: rights - ordinal: 19 text: null startSeconds: 1145 endSeconds: 1156 watch: 'https://liquidgold.cc/watch/Yphdry8ttAQ?t=1145' heldReason: rights --- # Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AI Ayush Bhardwaj could build the agent. What he could not do was tell whether it was any good. He moved from applied AI at a hedge fund to a pharma tech company expecting a different world, and found the job identical, including the wall. An engineer glances at generated code and knows instantly that it is weak, because years of training built that judgment. Nobody on his team had the equivalent instinct for a trade thesis or a drug candidate. He calls this the point where vertical AI projects quietly die, because the thing looks finished and then nobody buys it. Judging his way out with a model was, in his words, a stupid mistake, since it jargons its way through without knowing what alpha means. Verifiable rewards work for math and code because answer keys exist. Worse, the data that would teach a model to reason in these fields is deliberately withheld. Funds must file their holdings quarterly, and their returns drop once competitors reverse engineer them. Disclosing every clinical trial is legally required, yet roughly 30% of firms never do, and in 2026 the FDA publicly reminded more than 2,000 sponsors. The frontier labs do not have it either. So his answer is to hire the user, which for a team of young engineers meant hiring a senior scientist, after which their tools started speaking big pharma's language. That expert curates sources, sharpens prompts and does the judging, starting from error analysis as the cheapest rung. The moat is never the model or the infrastructure. Both are commodities. Speaker info: - https://x.com/aybh08 - https://www.linkedin.com/in/aybh/ - https://ayushb.me/ Timestamps: 0:00 - Reading the room 1:04 - What applied vertical AI actually means 2:43 - Leaving a hedge fund for pharma and changing nothing 3:34 - The wrong question about agents in production 4:27 - Step one, make the task narrow 5:18 - Proprietary data is the only real differentiator 6:57 - The easy part fits on one screen 7:50 - Why you cannot iterate on what you cannot judge 9:30 - Trying to use a model as the judge 10:21 - The data was never there, by design 12:05 - Hire the user 12:55 - Building the learning loop around a domain expert 14:39 - From error analysis up to preference training 16:20 - Reaching production is not the same as working 17:10 - The seven steps 18:00 - AI in the loop, not human in the loop 19:00 - Your moat is domain expertise and data Canonical: https://liquidgold.cc/watch/Yphdry8ttAQ · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 1:20) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:20 to 3:02) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:02 to 4:30) > Held: transcript rights do not permit this chunk's text.