Bikee Prajapati.
Full-stack developer, AI/ML engineer.

A full-stack learning management system where admins publish video courses and students browse, purchase, and learn, with Q&A threads, reviews, real-time notifications, and an analytics dashboard. Built as a Dev Weekends Fellowship project. I containerized the backend, built a GitHub Actions CI/CD pipeline, and fixed a handful of production issues along the way: a Stripe checkout hang, cross-site cookie failures, and a logout race condition from stale RTK Query cache state.

A retrieval-augmented chatbot for querying PDFs, built from scratch as a deep-learning exercise. I deliberately avoided LangChain abstractions to actually understand how retrieval pipelines work. Text is extracted with PyMuPDF, embedded with a HuggingFace sentence-transformer (MiniLM), indexed in FAISS, and served through FastAPI with a Mistral LLM for generation. Dockerized and deployed to HuggingFace Spaces.

A multivendor e-commerce marketplace supporting three roles (buyers, sellers, and admins) with product and event listings, discount coupons, real-time chat, order lifecycle management, refunds, and Stripe/PayPal payments. The harder part was the debugging: Mongoose 9 breaking changes, a MUI v4→v5 migration, async race conditions, and CORS/Socket.IO configuration across three separate services.

Backend and socket server run on Render's free tier and spin down after 15 minutes idle. First request after a while can take 30 to 60 seconds to wake up.

A tool that reads any public GitHub repository's structure and writes a professional README automatically, using Flan-T5 and Mistral through a LangChain pipeline. The backend is a fully async FastAPI service with fallback logic built in, so the app stays up even when a model endpoint goes down.