AI System Design: From Idea to Production
summary
Framework for taking an AI idea to production with concrete decisions not platitudes The talk claims a measurable goal to reduce processing time from 2 days to 1 hour in 90 days and shows a working design using MongoDB for claims storage Confluence PDFs chunked and embedded for vector search RAG retrieval a controlled workflow that routes complex cases to human review and escrow rules for denials It names metrics to track like missing citation rate override rate token cost and proposes optimizations such as RAG + control flow + human-in-loop and semantic caching and structured outputs
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:30Specs as new code in AI
- 2:35Four-phase AI design
- 3:38Health claims AI requirements
- 5:43Business constraints and performance
- 6:48AI in product design
- 7:30Success metrics for AI
- 8:35Data strategy for AI
- 10:42Data processing techniques
- 12:27Claims review workflow
- 14:12AI design patterns
- 16:23Applying patterns to claims review
- 17:48End-user feedback
- 19:58Stack and tooling
- 21:01Evaluating AI systems
- 23:51Production monitoring
- 25:14Cost, latency, reliability
- 27:22Key takeaways and closing remarks
related talks
56mAI War Stories: Real world learnings delivering enterprise AI solutions
36mBeyond the Hype
27mBuilding AI applications: Successes and Lessons
32mFrom Three Days to One Minute: Architecture Trade-offs in Production Enterprise AI
28mAI at Work
32mAI Engineering Anti-patterns
39mAI as a Catalyst, Not a Replacement
14m15 Months of building AI agents in 15 minutes