Building Agentic AI Assistants: Lessons from productionising agents in Preclinical R&D
summary
Built system called PRINCE unifies 20,000+ PDF reports (vector DB/OpenSearch) and Athena tables via retrieval-augmented generation plus text-to-SQL and agent orchestration (LangGraph/LangChain). Researcher agent retrieves top-seven chunks, reranker orders relevance, dynamic few‑shot prompting fixes SQL accuracy from <40% to near zero errors. Supervisor, planner, researcher, reflection and writer agents loop with K limits and human‑in‑the‑loop; search effort drops ~90%, first drafts of 100‑page regulatory summaries generate in minutes.
AI-written summary · about
chapters (15)
- 0:00Building AI agents in healthcare
- 1:28Drug discovery problem and Prints
- 3:52Duplicate studies waste time
- 6:39RAG to multi-agent systems
- 10:57Simple to multi-agent workflows
- 12:01Building Prints multi-agent system
- 14:52Researcher agent and hybrid search
- 19:24Few-shot prompting for SQL
- 21:54Reflection agent and data validation
- 22:15Productionizing agent workflows
- 23:43Evaluating agent responses
- 26:52Trust with explainability
- 29:01Handling failures with fallbacks
- 31:08Memory and context engineering
- 33:38Key impacts and future
related talks
45mAI agents in practice - beyond the prototype
01h 01mKeynote Panel: Future of AI agents
22mAgentic AI for Engineering Leaders | Strategy & Execution Guide
35mCan AI solve the biggest problem in drug discovery? Time!
49mAI Agents: Collaboration in Real-Life Applications
34mAI Agents: Transforming Autonomous Decision-Making
32mBuilding an Advanced Knowledge Assistant
28mAgentic AI