At Tata Consultancy Services, I architected an Excel-to-Java transpiler that converts workbooks into typed Java while preserving calculation semantics. Its multi-agent LangGraph pipeline includes a deterministic transpiler for 28 Excel functions and a correctness system hardened with 1,700+ automated tests.
I also studied why a graph neural network mispredicts shortest-path costs. I traced errors to opposing mechanisms in large- and small-diameter graphs and used controlled synthetic-edge experiments to cut loss by up to 9.6×.
For Catalog AI, I architected an LLM-powered product-catalog enrichment pipeline using FastAPI and React, with a human-in-the-loop approval workflow. I also developed a LangChain/LangGraph framework that parses seven file formats into OpenAPI specifications and coordinates specialist extraction agents.
My other work includes benchmarking LlamaIndex agent architectures for generating database data and test cases, and co-authoring ClauseBandit, an LLM and SMT-solver framework that achieved 74.1% accuracy across 27 industrial tasks. I trained a TensorFlow/Keras heart-murmur classifier that reached 66.2% mean F1; the method was filed as a patent.

