A video captioning pipeline that samples frames from a video using ffmpeg, then sends them to VLMs to generate captions in four styles: formal, sarcastic, tech-humor, and general humor. Optionally runs multiple models in ensemble and uses a judge model to pick the best caption, in addition to testing pipeline.

Security-hardened team-matching platform for university students — matches teammates by skill similarity, verifies members via university email domains, built with Next.js, FastAPI, and Supabase.

Agentic personal medical assistant that reasons over medical data using multi-agent orchestration, with leveraging mutiple ML/DL pre-trained models, with addition to relational and vector databases.