// projects
Production ML systems, LLM applications, and open-source tooling.
Retrieval-augmented generation system for the NetSuite ERP: document ingestion, vector search, and grounded summarization/Q&A over enterprise documents, orchestrated with LangChain and LangGraph.
Real-time voice agents for a law-firm product built entirely on open-source VAD, STT, and TTS models running locally on GPU machines — hosted on Google Cloud Run, handling 20–50 concurrent voice streams per GPU.
Fine-tuned LLMs (PEFT) that convert user stories into complete test cases. Deployed on local models (LLaMA, Mistral, DeepSeek-R1) instead of the OpenAI API — saving $1,400/month while keeping data in-house.
A Model Context Protocol server exposing enterprise tools to LLMs in real time, plus agentic applications built with LangChain, LangGraph, and AutoGen on top of it.
Smart India Hackathon 2022 national-winning project for the Ministry of Shipping and Ports — a virtual training simulator for crane operators.
Vehicle-to-everything communication combined with deep learning models to detect and prevent collisions autonomously.
Hands-free computer control: a virtual mouse driven by an AI voice assistant, combining speech recognition with gesture-free accessibility.