An image classifier built on DenseNet-121 to detect whether a face is real or AI-generated. Trained on 65,000 images (32,500 real / 32,500 synthetic) — achieved 99.42% accuracy on the test set. Includes a React frontend where users can upload an image and receive a live prediction.
A two-stream video classifier — one stream processes per-frame spatial features; the other captures temporal motion via frame differencing. Trained on 3,600 videos (1,800 real / 1,800 synthetic), achieving 92.6% test accuracy. Demo is live on Hugging Face Spaces.
Using react native and SQLite, I built a mobile app to track workouts and gym progress. Users can log exercises, sets, and reps, view workout history, and analyze progress over time. The app stores data locally on the device itself ensuring offline access and quick load times in cases where users do not wish to be distracted by using internet while working out.
Built a backend-driven AI system that generates structured CVs from GitHub profiles by integrating GitHub REST APIs with Large Language Models via OpenRouter (Gemma/DeepSeek). The system extracts and normalizes user repositories, activity, and metadata, then uses prompt-engineered LLM inference to generate professional summaries, skills, and project descriptions. It features a Node.js + Express architecture with secure API handling, rate limiting, CORS protection, and a custom request queue for managing concurrent AI calls. Deployed in a serverless-ready setup (Vercel), it acts as a scalable middleware between frontend clients, GitHub data sources, and AI services for real-time CV generation.