Things I built because I was curious, bored, or mildly annoyed something didn't exist yet.
“MUJ needed honest college talks without the social baggage.”
Anonymous 1-on-1 chat, swipeable confession wall, and group rooms for MUJ students. Real-time via Firebase RTDB, shadowban system, and a content moderation layer.
“Students needed brutal financial honesty, not another pretty banking app.”
PWA that uses Llama 3 (Groq) to generate monthly spending report cards, flag bad habits, and surface savings opportunities — designed for broke Indian students.
“I wanted to actually see what happens inside a neural net while it trains.”
Browser-based tool to design, train, and observe neural networks. Color-coded activations, live loss/accuracy charts, dataset playground (XOR, moons, spiral) — no backend required.
“My first end-to-end ML deploy. Picked Twitter data because tweets are hilariously extreme.”
Logistic regression + TF-IDF trained on 100k Sentiment140 tweets, achieving ~77% macro F1. Deployed on Hugging Face Spaces with a Gradio interface.
“Curious whether you could replace an insurance adjudicator with OCR + LLM logic.”
Upload a medical bill → Gemini Vision extracts text → Groq Llama adjudicates through a 6-step rule engine (eligibility, fraud, limits). Surprisingly accurate on real receipts.
“Wanted to deploy a real PyTorch model and watch predictions stream live in the browser.”
FastAPI + PyTorch energy load forecaster backed by PJME dataset. React dashboard with a 2-second polling loop, glassmorphism UI, and confidence meter.
“MUJ needed a chatbot that talks like a helpful senior, not a corporate FAQ bot.”
RAG chatbot backed by FAISS + LangChain that reads MUJ PDFs (academic calendar, mess menu, syllabus) and answers in Hinglish tone via Llama 3.3 on Groq.