Justin Dumas-Carr

Built AI call classification on Claude (AWS Bedrock) for a hospital — ~600 calls/day

categorized by intent and quality, withschema validation preventing bad LLM output from reaching the database(an eval/guardrail at the write boundary). Owned the surrounding production system: event-driven ingestion (webhooks → API Gateway → SQS → Lambda classifier → RDS) with idempotency, DLQ failure isolation, and HIPAA controls (PHI masking, KMS, pgAudit, cdk-nag). Built GraphQL EHR integrations (Healthie) and a billing rollup engine for CMS-1500 claims. *Claude

• · Python

◦ · AWS Bedrock/CDK/Lambda/SQS/RDS

◦ · Drizzle

◦ · Postgres*

Built and operate a multi-agent orchestration framework — agent teams own tasks from research through execution and self-verification, auto-deploying above a 0.85 confidence threshold — reaching ~95% ideation-to-deploy autonomy across 9 concurrent projects. Built a 21-tool Model Context Protocol (MCP) server (Python, 2,260 lines) with an FTS5/BM25 search engine, cross-project constraint validation, and execution-trace logging — the context backbone the agents run on. Made evals first-class: a 6-layer confidence-scored verification gate per change. Cut LLM/inference cost 98% ($120→$2/mo) by self-hosting on serverless GPU; built a

4-stage daily pipeline (3.75× speedup via asyncio). *Claude/OpenAI/Gemini

• · Python

◦ · TypeScript

◦ · asyncio

◦ · Convex

◦ · Supabase

◦ · Sentry*

Set up Azure Delta Lake; Spark in Synapse for clean-and-query at scale. Migrated Montreal Airport's (YUL) Data from SQL Server to Azure for Data Analysis Machine Learning.