Medical evidence retrieval and citation provenance
Built medical RAG pipelines across textbooks, clinical guidelines, and research literature. Combined document ingestion, OpenSearch/Qdrant hybrid retrieval, biomedical search, and source deduplication with passage-level provenance and traceable citations so readers can inspect the evidence supporting an answer.
LangGraph and LangChain runtime
Engineered multi-provider LLM workflows using LangGraph and LangChain for clinical research and medical education. The runtime supports parallel tool execution, structured outputs, citation validation, token budgets, streaming, and bounded retries and fallbacks.
Evaluation and observability
Built Langfuse tracing, versioned prompt management, evaluation harnesses, and retrieval replay/profiling tools. These tools investigate source grounding, answer coverage, latency, token costs, and provider failures.
Durable operations and reconnectable streams
Designed a FastAPI/PostgreSQL backend with independently deployable API and worker services, transactional outbox dispatch through Cloud Tasks, idempotent operations, replayable server-sent events, cancellation, workspace authorization, rotating sessions, and usage metering.
Clinical workflows across web and mobile
Developed Next.js/React web and React Native/Expo iOS/Android clients with evidence-linked answers, OCR-backed file uploads, clinical document workflows, and editable medical mind maps. Built administration tools spanning users, operations, analytics, and spend.
Deployment and infrastructure
Built containerized infrastructure on Google Cloud Run and Cloudflare Pages, provisioned with Terraform. CLI-driven release pipelines support immutable artifacts, independent component releases, validation gates, smoke checks, and rollback workflows.