Reach me at arushkarnatak1881@gmail.com

• SMARTFLOW: a reinforcement-learning traffic-signal controller that reads live per-lane demand and learns to choose signal phases in the SUMO simulator — cutting average vehicle waiting time ~55% vs a fixed-time controller and matching strong heuristics like max-pressure, with a multi-agent 2×2 grid extension (shared-parameter PPO) and emergency-vehicle preemption. Built with Python + SUMO + Stable-Baselines3 (DQN/PPO) + FastAPI + React/TypeScript → Code

• TraceRAG: a GraphRAG engine you can see — answers questions over real codebases by fusing vector search with knowledge-graph traversal, then renders the full retrieval path live on a React Flow canvas so every claim is traceable back to its source. Built end-to-end with FastAPI, LadybugDB (single-file vector + graph store), GLiNER entity extraction, and a Vite/React frontend. → Code · Live Demo

Tech stack I work with:

Languages
ML / Computer Vision
LLM & AI Agents
Backend
Database
MLOps & Tools

I'm a B.Tech IT student (GGSIPU, Class of 2027) who ships AI systems that run in production — not just notebooks.

In the past year I've architected ML pipelines processing 10,000+ aerial images a day at Kanchan Drones, built a real-time exam-proctoring system holding sub-200ms latency on GCP, and shipped a policy-aware RAG chatbot that answers in under a second. I've taught an 8-week NLP course to 20+ students, and built SmartFlow — a reinforcement-learning traffic controller that cut vehicle waiting time in half and won a national hackathon.

I live at the intersection of computer vision, LLMs, and MLOps: RL agents, GraphRAG engines, real-time CV pipelines, the works. My favorite thing is taking research-grade ideas and turning them into products people can actually use — and I share the journey on my portfolio.

I contribute to open source along the way. Outside of models and code, you'll catch me sharing what I build on my portfolio and socials.

I consider myself curious, hands-on, and always looking for problems worth solving.

5 national hackathon finals, 2 first-place wins, and counting. Always hunting for the next problem worth solving.