Agents, ranking models, and full-stack products, shipped and tested, not just demoed.
Deterministic rules for the obvious cases, confidence-gated LLM escalation for everything ambiguous. Text, image, and voice, unified into one pipeline.
XGBoost ranking model evaluated on chronologically split data to eliminate lookahead bias. Rankings and SHAP explanations exposed as MCP tools an LLM can call directly.
Full-stack writing platform, designed, built, and shipped solo. React and TypeScript frontend, Supabase backend, Stripe subscriptions, plus an MCP server for live production data queries.
Turns meeting transcripts into owner-tagged action items, every task backed by a verbatim source quote. Rebuilt around async queues to beat Slack's 3-second response window.
Real-time multilingual interpreter for patient-clinician conversations. Five coordinated agents (Intake Conductor, Follow-up Proposer, Coverage Tracker, Red-Flag Monitor, History Scribe) in a shared Band.ai room.
Declarative DSL and compiler generating pseudo-realistic IoT sensor streams with causal event chains, for training downstream ML models at production scale.
Form a frame with both hands, thumb and index fingers, and the interior live-restyles through an AI video model. Real-time hand tracking (MediaPipe) compositing a live Decart video stream, fully client-side.
Assembles Bharatanatyam dance compositions synced to a song's beat structure: audio phrasing analysis, pose extraction from reference performances, and a Three.js 3D avatar rendering the composed choreography live.
Open to new grad AI/ML, backend, and full-stack roles.