At Maruti Empire – Pavansut Inc., I build LLM, ML and data features for operations and leadership. I built a guarded LangChain tool-calling assistant that returns structured JSON to fill parameterized SQL templates.
I improved RAG top-3 hit rate from 71% to 89% by addressing metadata pre-filtering, hybrid search and re-ranking. I also built document ingestion for roughly 150–250K chunks, using structure-based chunking and embeddings indexed in BigQuery Vector Search and FAISS.
I evaluate LLMs with hand-labeled test sets and tune production serving; this included lifting Gemini classification from roughly 70% to the mid-80s and reducing vLLM latency by 35% and token cost by 25%. I ship Python services with FastAPI, Docker and Cloud Run, and support reliable data pipelines with Airflow and Databricks.
At Mercor, I built a golden set for evaluating models and a FastAPI/Pydantic validation service, and review AI-generated Python and SQL. I also fine-tune LLMs with PyTorch and LoRA, and hold the Google Cloud Professional Data Engineer certification.
