Designed and built a production-grade, multi-agent LLM copilot that answers complex product questions by grounding responses in millions of customer feedback entries using a hybrid retrieval pipeline.

Implemented a retrieval strategy combining pgvector semantic search, keyword search, and metadata filters, followed by a cross-encoder reranker to deliver the most relevant evidence to GPT-4/Claude for final response composition.

Securely integrated live tool access (Jira, Slack, GitHub) via an MCP server, allowing the agent to perform actions and fetch real-time context beyond static documents.

Ensured answer trust and traceability by building a citation grounding layer and using LangSmith for full agent observability, debugging, and performance evaluation.

At Simform, I led the backend development of OrthoAssist, a real‑time, voice‑first AI copilot designed for orthopedic urgent care clinics. The system listens to clinician‑patient conversations, drafts structured SOAP notes and orders in real time, suggests follow‑up questions, and retrieves relevant patient history from the EHR—all while maintaining strict HIPAA and PHI compliance.

This solved two critical problems for the clinic network: documentation overhead (clinicians spending hours after each shift on manual charting) and AI adoption skepticism (fears around hallucinations, workflow disruption, and data privacy). By building a retrieval‑first, clinician‑in‑the‑loop system, we reduced note completion time by [X%—insert your real number if available] and achieved zero critical PHI violations across pilot clinics.

Key technical challenges I owned included: orchestrating a multi‑agent runtime with real‑time streaming transcription, grounding all model outputs in FHIR‑based patient data to eliminate hallucination risk, standardizing EHR and imaging tool calls via the Model Context Protocol (MCP), and building an evaluation harness with regression gates to ensure clinical safety before every deployment.