I build software that keeps working when real people use it.

Led a 5-person team to ship a fitness platform with real-time form checking from the camera.

• Led a 5-member team building a fitness platform with three modules: the web platform, LangChain-based nutrition and diet recommendations, and AI workout form tracking. Contributed hands-on across all three.
• Users had no way to check gym form at home, so we used Google's MediaPipe to track body landmarks from the camera and check exercise form in real time.

Built the exam datesheet generator the institute runs on; 3x faster across 400+ constraints.

• Made the institute's exam datesheet generator: a multithreaded scheduler (Flask + MySQL) that cut end-to-end generation time 3x under high constraint density.
• Wrote its constraint solver, which handles 400+ variables (subjects, rooms, faculty) with real-time conflict detection and automated resolution.

Fine-tuned a quantized LLaMA with QLoRA and RLHF as the safety layer of a mental-health chatbot; tested against 50+ adversarial prompts.

• Built a multi-agent safety pipeline for mental health chatbots: emotion, sentiment, and safety modules analyze each input independently, then a deterministic aggregator decides the final response. Tested it against 50+ adversarial prompts for edge cases.
• Fine-tuned a quantized LLaMA with QLoRA and RLHF for safety alignment, layered behind toxicity filters and fallback rules so no single model failure reaches the user.