01Architected a six-agent Amazon Bedrock system where AWS Lambda stream processors fire domain watchers (LocationWatcher, MedicalWatcher, FireWatcher, CrimeWatcher) per transcript word — triggering specialist agents asynchronously with InvocationType:Event so nothing blocks the live transcript pipeline.
02Implemented a two-model strategy: Claude Haiku 3.5 as a parallel verifier resolving location ambiguities and severity corrections mid-run, and Claude Sonnet 4 as the Coordinator synthesizing all agent outputs into a single dispatcher recommendation card — right model for the right reasoning tier.
03Built a Bedrock Knowledge Base over FEMA Emergency Response Guidebook, AHA cardiac protocols, and MPDS triage guidelines — Amazon Titan Embeddings V2 producing 1536-dimension vectors, top-5 semantic retrieval grounding every medical and hazmat recommendation to a cited source document.
04Enforced human-in-the-loop at the architecture level via Bedrock Guardrails: no unit is ever dispatched, no hospital alerted, without explicit dispatcher approval — override logging captured to DynamoDB for post-incident audit.
05Delivered a WebSocket-driven dispatcher dashboard (React + Tailwind, Mapbox GL JS, API Gateway WebSocket) where transcript words, agent status, partial approvals, and the final recommendation card each push independently the instant data arrives — zero full-page spinners at any point in the pipeline.