CTO @ Clarity
2020 — Present
- Architected and built a multi-tenant healthcare RAG platform on AWS: a five-stage Step Functions pipeline (ingestion → canonicalization → vectorization → LLM enrichment → delivery) with per-tenant state machines, vector indexes, and API domains.
- Designed the retrieval stack: hybrid BM25 + KNN search over OpenSearch fused with weighted Reciprocal Rank Fusion, Cohere Rerank v3.5 cross-encoder reranking, and MMR diversity selection — plus exact and semantic near-duplicate suppression at index time.
- Built end-to-end data provenance: every LLM answer resolves to deterministic evidence atoms (source document, XPath/page, quote), so clinical answers are grounded and auditable back to the original ingested file.
- Engineered healthcare data ingestion and interoperability: an HL7 ↔ FHIR canonicalization change feed, a containerized C-CDA ↔ FHIR conversion service on autoscaling ECS Fargate behind a private VPC Link, PDF extraction with AWS Textract (forms/tables), and Carequality/TEFCA patient discovery and document retrieval.
- Built Epic EHR integration: OAuth2 authorization-code flow with CSRF state validation, refresh-token lifecycle management in DynamoDB (TTL + streams), an authenticated FHIR API proxy, and event-driven patient queries over EventBridge.
- Drove LLM cost optimization: model tiering on Bedrock (routing lightweight passes to cheaper models), rerank-skip heuristics, token budgeting, batched embeddings, and a usage-metering/billing subsystem with idempotent events and versioned rate cards.
- Built prompt optimization workflows: S3-hosted Jinja2 prompt templates hot-swappable without redeploy, multi-pass citation repair, and an evaluation harness with ground-truth grading.
- Built PHI protection and clinical NLP services: reversible PII de-identification (tokenized redaction with secure re-identification) and clinical entity extraction covering conditions, medications, and PHI.
- Built a wearables health-data platform: webhook-driven ingestion from ~30 wearable and health-device brands (Garmin, Fitbit, Apple, Oura, Whoop, CGMs) into Amazon Timestream for time-series analytics, with LLM-agent insights built on Langchain.
- Designed the multi-tenant identity platform: a FastAPI/PostgreSQL service mapping external IdP tokens to stable internal principals with per-tenant roles, DB-backed JWKS/RS256 verification, Auth0 provisioning with saga-style compensation, and internal-only service auth over VPC Lattice with SigV4.
- Owned the cloud architecture end to end: AWS CDK (TypeScript), Python Lambdas and FastAPI on Fargate, DynamoDB single-table design, Timestream, S3, SQS/SNS/EventBridge alerting, and Cognito/Auth0 authorization.
- Python
- AWS Bedrock
- OpenSearch
- Step Functions
- CDK
- Langchain
- AdalFlow
- FHIR/HL7