At Mammoth Analytics, I architected routing for Source Interpreter so only ambiguous inputs go to the LLM, while interpreted sources are saved as reusable recipes for future uploads. I also built a retrieval layer for a multi-agent support system using PostgreSQL and pgvector.
I replaced an LLM-based pipeline optimizer with a deterministic graph-based engine and shipped a production missing-value imputation pipeline after benchmarking MICE against LightGBM. I also designed and owned a PDF-to-CSV extraction microservice, including queue-based processing and response caching.
