Summary
Ayush Bhardwaj, who applied AI at a hedge fund and now leads AI development at pharma-tech startup Allos AI—which uses AI to build drugs—presents at AI Engineer on a seven-step recipe for applied vertical AI: AI built to simulate expert roles in specific high-stakes industries like finance, pharmaceuticals, and legal tech. The talk draws directly on first-hand experience across both domains and identifies surprising parallels in how the same core engineering challenges appear regardless of industry.
A central theme is the data problem unique to regulated industries. Hedge funds holding over $100 million in US equities must publicly disclose quarterly positions, which erodes competitive moat the moment they file. In pharma, while clinical trial outcomes must be disclosed by law, nearly a third of firms never do, leaving significant gaps. This data scarcity and sensitivity makes standard ML iteration loops impractical and forces practitioners to rely more heavily on domain expert judgment.
Bhardwaj is candid about failure modes, particularly the temptation to use LLM-as-judge for domain evaluation in fields like trading or drug discovery. He argues this fails because LLMs predict probable language, not actual alpha generation or chemical validity—and without verifiable rewards analogous to math or code, errors compound silently through multi-step agent pipelines. His prescription: aggressively narrow the task, involve domain experts early and continuously in the evaluation loop, and resist the urge to declare success before someone who actually understands the domain’s ground truth has validated the output.
📺 Source: AI Engineer · Published August 19, 2026
🏷️ Format: Keynote Launch
