Harsh Chauhan

About

AI/LLM engineer who has shipped a production RAG platform at Ticker Inc. on GCP (Qdrant + Postgres) delivering sub-second retrieval over 550k+ items, with measurable gains in latency and answer quality (HNSW optimization, MMR re-ranking). Also built an asynchronous LangChain/LangGraph multi-agent research system (10x faster cycles) and partnered with Indiana University doctors on synthetic patient records and ML error analysis using clinician-friendly F1/loss dashboards.

Key Strengths

• Built and deployed production RAG system on GCP for 200+ users over 500k+ financial datapoints and 40k articles
• Improved vector search efficiency 40% and reduced query latency 60% via HNSW parameter optimization (A/B tested vs IVF)
• Designed hybrid retrieval architecture with metadata-rich Qdrant + optimized PostgreSQL pre-filtering
• Improved answer quality ~40% by implementing MMR to increase retrieval diversity and reduce redundant chunks
• Built multi-agent research platform with asynchronous orchestration, accelerating research cycles 10x
• Hardened multi-agent workflows using Pydantic schema validation for reliable structured handoffs
• Applied rigorous evaluation practices (A/B tests, expert-in-the-loop validation, F1/loss monitoring) with reported +15% accuracy improvement
• Selected domain-aligned models (BioClinicalBERT/DeBERTa) for healthcare NLP, boosting classification accuracy 35%
• Collaborated with non-technical clinicians; translated metrics into dashboards and led cross-functional error analysis (6-person team)

Experience

Education

Skills

Languages

Certifications

Publications

1 publication