Descriptions:
Apoorva Joshi, a data scientist turned developer advocate at MongoDB, presents a four-phase framework for designing AI systems end-to-end — from initial idea through production — using a health insurance claims review system as a running example. The framing is deliberately counterintuitive: in an era of AI coding tools, Joshi argues that product requirements and system design specs matter more than ever, not less, citing Anthropic and OpenAI practitioners who describe specs as “the new code.”
The four phases are: product requirements (define the business problem with specific users, measurable baselines, and solution-agnostic constraints — e.g., medical reviewers at MDB Health spend 2 days per claim, four times the industry standard); system design (choose data architecture and AI patterns — RAG, agents, or agentic pipelines — starting from the simplest viable design); evaluation and monitoring (establish success criteria before shipping and maintain them in production); and optimization (iterate on cost, latency, and reliability alongside accuracy). Joshi walks through concrete retrieval decisions for the claims use case: hybrid search with metadata pre-filtering for medical terminology like diagnosis and procedure codes, exact-match lookup for patient history, and staged human escalation for complex or denial cases.
The talk also covers MongoDB and Voyage AI integration points, agentic vs. multi-agent architecture tradeoffs, and why jumping straight to an agent architecture — however tempting — typically produces over-engineered systems that are harder to evaluate and maintain.
📺 Source: AI Engineer · Published June 28, 2026
🏷️ Format: Deep Dive
