Ayush Bhardwaj is the tech lead at Allos AI, where he oversees applied AI for pharmaceutical research, including multi-agent systems, infrastructure, evaluation, and observability. Previously, he built agentic systems at D. E. Shaw that surfaced macroeconomic signals and market research for traders. His work addresses a shared problem in both industries: making general-purpose models useful when critical knowledge is proprietary and only experienced specialists can judge the results.
Bhardwaj started programming in high school and developed experience in software engineering and cloud infrastructure before specializing in applied AI. Through Google Summer of Code 2024, he contributed to LORIS, an open-source research and imaging platform, with the International Neuroinformatics Coordinating Facility. He also developed a knowledge-curation system for collecting information from multiple sources and producing tailored summaries.
Building AI around professional expertise
His work across finance and pharmaceutical research centers on several practical convictions:
- Applied vertical AI needs narrow, measurable tasks. An agent should address a bounded professional problem, such as ranking companies against specified investment criteria, and demonstrate concrete business value.
- Proprietary data creates the durable advantage. Internal trade theses and failed pharmaceutical experiments contain knowledge that public datasets and widely available models cannot reproduce.
- Domain-expert evaluation must shape the entire workflow. Traders and scientists should help define problems, assess sources, structure reasoning, and review outputs. Fluent responses and automated grading cannot reliably substitute for their judgment.
- Error analysis should precede expensive optimization. Inspecting traces and correcting specific failures can improve systems before supervised fine-tuning or reinforcement learning becomes necessary; fixed evaluation rubrics can reinforce their own blind spots.
- AI in the loop preserves expert accountability. Models can propose investment theses or narrow drug candidates, but professional users remain responsible for consequential decisions.
Bhardwaj also created chat-history, a Rust command-line tool for searching, inspecting, and exporting conversations across Claude Code, Cursor, and OpenAI Codex. He traced its origin to the practical frustration of session limits and context scattered across coding assistants.
