AI engineer. Production LLM agents, computer vision, and the GPU infrastructure underneath them. IIT Madras.
I build AI systems that run on real work. At Centaq India I built the AI function from one workstation to a multi-GPU platform, and I now lead a small team shipping LLM agents to real clients.
- Edward: a tool-using LLM agent that interviews clients and files a structured creative brief into a partner's order workflow. Live on real orders. 25 typed tools, guardrails enforced in code, a fallback chain across four model providers, and an evaluation setup with simulated clients and regression checks.
- AI infrastructure and vision models: in-house models for portrait and face retouching, colour correction, hair removal and background removal. Multi-GPU training, TorchServe serving, and a production pipeline that processes 20,000+ images a day. Cut editing from 8-10 minutes per image to seconds.
- SnapStand: an on-device AI photobooth for iPad using CoreML.
- Vissoo Portal: a multi-tenant platform for creative collaboration and asset review.
- SeriousDB: fixed write-ahead-log replay to reject corrupted entries, with regression tests. Merged upstream (PR #344).
- InterviewGraph: an open-source interview question bank, with a planned tool that predicts likely questions from a job description. In progress.
- Multi-ball football tracking: tracks four balls through occlusion and labels possession from player pose. A YOLO11m detector distilled from Grounding DINO labels reaches 0.966 mAP50, and the write-up covers where the student beat the teacher.
- MethaneGuard: methane monitoring and leak localisation with a hybrid RAG copilot.
- Cricket biomechanics: real-time pose analysis extracting 20+ kinematic metrics from athlete motion.
Python, PyTorch, FastAPI, TypeScript, PostgreSQL, Docker, Kubernetes, Azure, GCP. Gemini and open-weight models.