At GobbleCube, I built a production human-validation layer for ML attribute extraction, spanning 14,053 SKUs. I also recovered 125,807 records by fixing a join defect and resolved schema drift with a Pydantic gate and canonicalisation engine.
At NIELIT, I achieved 80%+ emotion-classification accuracy across 50,000+ audio samples and deployed the model as a Flask API. I also engineered a multi-agent research assistant that synthesized reports from 20+ real-time web sources and cut research-synthesis time by 70%.
On my QuickCommerce Copilot project, I shipped a live agentic RAG assistant that raised tool-call success from 20% to 100% and citation coverage from 56% to 100% against a retrieval-only baseline. I also built a dbt and LightGBM pipeline that cut day-ahead electricity load-forecast error by 52%.
