At Simpplr, I evaluate and optimize production LLM and search models for accuracy, latency, and cost. I replaced a paid LLM API with self-hosted search intent classifiers, improving accuracy from 95.7% to 99.1% and lowering p50 latency by 3–6×.
I also optimized the RAG answer engine, reducing latency by 60% and cost per token by 40% while supporting 8× more concurrent users. Prompt restructuring raised capacity 26×, and I cut a 33-hour job to 12.7 hours while reducing GPU cost by 17–21%.
In my projects, I built an experimentation and causal inference toolkit tested on real ASOS experiments and Criteo users, and a demand forecasting service for Walmart M5 data. The forecasting system used LightGBM, a dbt and DuckDB feature pipeline, and an MLflow model gate; its winning model reduced item-level WAPE by 18.5% versus seasonal naive.

