projects

I trained and released small BDH language models, comparing byte and BPE training, streaming inference, and recall.

Wireless physics sensors (ronto) plus a live dashboard, simulator, and graphing app. Began as kineti-kit, 1st at IdeaHacks 2026.

Plugins for DeepSeek Harness that manage context size, limit concurrent requests, and adjust timeouts for local models.

Turns floor-plan images into walkable 3D homes, with chat-based restyling and furniture imported from Amazon links.

Records AI agent runs, identifies likely failure points, and replays corrected steps to test what went wrong.

A handmade, standalone music mixing station with real faders and knobs. Web app, firmware, electronics, and a 3D-printed enclosure

An expert-labeled benchmark for evaluating how well LLMs write apologies, rated on accountability, specificity, and warmth.

A low-cost infrared scanner that maps veins in real time with a Raspberry Pi and NIR LEDs. Built at MakeMIT x Harvard Hackathon 2026.

A GraphRAG system for grounded foreign-policy analysis. Independent research paper and interdisciplinary project.

An AI agent that creates, queries, and updates Notion database tasks from natural-language speech or text.

A Python agent that manages Google Calendar from plain-language requests and retries failed scheduling steps.

An Arduino sensor system that monitors temperature and humidity and detects falls, reporting wirelessly for worker safety.

An electronic whistle with a haptic wearable that delivers in-game signals to athletes with hearing loss.

I led Quarry Lane's team in the MIT BWSI Build a CubeSat Challenge 2024. We presented our work at the program's final event.

A computer-vision model using YOLO and Darknet to detect cars, pedestrians, and other objects on the road in real time.