Vision pipelines, LLM tooling, and embedded systems — built at the boundary of hardware and intelligence.
6-stage vision pipeline automating dataset generation for 16 blood tube classes across 3 tube families — RealSense depth camera, MobileSAM segmentation, and Pydantic-versioned reproducible exports. Delivered under industrial client contract; 3 ambiguous class pairs resolved via depth fusion.
Local reverse proxy that sits transparently between any LLM client and the upstream API — intercepts calls, tracks token usage per session, and auto-compacts context at configurable soft/hard thresholds before limits are hit. Supports Anthropic, OpenAI, and any OpenAI-compatible endpoint with zero client-side changes.
Physics-meets-ML digital twin for a combined-cycle power plant — MATLAB/Simulink model paired with Random Forest achieving 99.1% prediction accuracy on 9,568 operational data points across 24 months.
Live Simulation
Combined Cycle Power Plant
Output Predictor
Full-stack IoT platform built in 48 hours for IBM TechXchange — Granite-3-8B LLM integration, WebSocket architecture delivering sub-100ms API response under concurrent device load. Hackathon Finalist.
Event-driven home automation engine on Raspberry Pi — 20+ heterogeneous sensors over MQTT, rule-based Python logic, and a Node-RED dashboard validated over 6 weeks continuous operation.