Gym rep counters don't say much. RepSense is a real-time, on-device coach that uses an ESP32-S3 and IMU data to analyze bench-press form and provide feedback - rep count, imbalance, tempo, rest quality - via the OpenAI API.
Tested whether an AI can be made to forget or correct specific facts without full retraining, a core question for privacy and the 'right to be forgotten.' Led the evaluation of two leading editing methods (ROME and MEMIT, Meng et al.) across GPT-2 model sizes, measuring whether edits worked, stayed contained, survived rephrasing, and kept the model fluent. Found that the smallest model corrupted unrelated facts and the largest failed to hold edits when they were rephrased, while mid-size models performed best
A Database Systems build with a five-person team. I owned the seller side - inventory CRUD, order fulfillment, stock - plus the UI/UX. Flask + PostgreSQL with complex queries; search, carts, checkout end to end.
A convolutional neural network that sorts waste imagery into recyclables, organics, and landfill - trained for sustainable waste-management workflows.
A Flutter + Firebase to-do app with authentication and real-time sync, a playful Dark Side aesthetic, and Android + Web support. Team project.
Retro-terminal, monospace, no-scroll. Next.js + TypeScript + Tailwind, embla drag-to-pull, live Open-Meteo weather, static-exported - built on a main/prod branch strategy with semantic versioning.