I'm a final year IT student at VIT Vellore. I like building systems where AI handles the fuzzy decision support while physics and hard math make the actual calls.
- Samsung PRISM : Wrapping up a multimodal framework to detect AI generated content. The text detector hits 95% accuracy.
- CAPS-FCL: Early research phase for an IoT and computer vision waste segregation system.
AirWatch India Built for the ET AI Hackathon. Takes live pollution hotspot data and figures out exactly which facility an inspector should visit. Instead of an LLM guessing, it scores 7,900+ real emission sources using a Gaussian plume dispersion model with live wind data. (Fun fact: we caught that 80 out of 84 WAQI city feeds were serving years-old data during this build. Switched to OpenAQ v3 with strict staleness checks so the API fails honestly rather than faking urgency).
Eco-Loop Building Agent Honeywell Eco-Loop hackathon project. Uses a local Llama 3.1 8B model to control EnergyPlus (a building physics simulator) in a live closed loop. It cut cooling energy by 22% and the total bill by 9% in a two week simulated run, while keeping human comfort levels perfect.
Enterprise AI Copilot for Google Sheets A production grade AI copilot built natively on Google Apps Script and Gemini. Uses a strict pipeline (Context Builder to Planner to 7 Specialist Agents to Formula Verification) instead of a basic chat wrapper. Includes full undo stacks, cost tracking, and dry runs.
Python (FastAPI, PyTorch, scikit-learn), React (TypeScript, Vite, Tailwind), Azure OpenAI, Ollama, MCP.