Apoorva Joshi presents a framework for AI system design, beginning with a measurable business problem, constraints and success metrics. Her hypothetical health insurance claims example includes requirements for human review and keeping sensitive data within an approved environment.
She walks through data freshness, document preparation and hybrid retrieval before choosing a controlled workflow with retrieval and human oversight. She stresses beginning with a simple architecture and changing it in response to evaluation.
The talk covers input and output guardrails, faithfulness and domain-specific metrics, then monitoring reviewer overrides and other production signals. Joshi closes with cost, latency and reliability techniques including reranking, caching and structured outputs.
Watch on YouTube
