Ayush Bhardwaj finds similar development problems in hedge-fund and pharmaceutical AI despite their different timelines and tolerance for errors. He starts with a narrow task, relevant data, prompts that reflect how an expert works and traces that make behavior inspectable. Broad requests are harder to evaluate than clearly scoped responsibilities.
Ayush Bhardwaj argues that proprietary knowledge can distinguish an industry-specific application from a general assistant. Examples include the reasoning behind past trades and records of failed experiments. He describes the gap between building a functioning agent and knowing whether its output has value when the engineer lacks the relevant trading or scientific expertise.
Ayush Bhardwaj recommends bringing intended users or experienced domain specialists into the development team. They can judge sources, refine prompts, decompose work into a sensible sequence and evaluate outputs. He warns that an ungrounded model-based judge can reward plausible terminology without establishing whether an answer is useful.
Ayush Bhardwaj places error analysis early in the improvement process, before more expensive training methods. He also discusses demonstrations, human feedback and expert-written rubrics, noting that fine-tuning creates continuing costs as base models change. Repeated expert interaction generates a growing record of successful and unsuccessful approaches.
Ayush Bhardwaj describes the near-term role of AI in these settings as assisting expert decisions. An agent can propose candidate trade theses or help evaluate drug candidates, while the specialist retains judgment. He argues that deployment should be assessed by practical benefit and cost, with domain expertise and accumulated data providing lasting differentiation.
Watch on YouTube
