ML / AI Engineering · Kiran Kumar
I spent a PhD asking that about human attention. Now I ask it about LLM routing, marketing attribution, and the production ML systems that answer it at scale, most recently at Chewy, EquipmentShare, and Convoy.
Currently building
A common interface evaluation harness for comparing agentic LLM routers (LiteLLM, RouteLLM, vLLM Semantic Router) across live benchmarks. Scores on a deployment focused metric suite (cost, latency, tool call accuracy, Pareto frontier) instead of a single leaderboard number.
View on GitHub →A Bayesian media mix modeling package for measuring marketing effectiveness across geographies and KPIs at once: hierarchical geo×KPI pooling, adstock/saturation curves, and lift calibration from real incrementality experiments.
View on GitHub →Experience
Marketing measurement and causal MMM/incrementality tooling behind $500M+ in annual media spend; architected an LLM/agent prototype that helped set the org's AI roadmap.
Architected an LLM powered RAG search & recommendation engine and a production LLM gateway on Kubernetes handling 10K+ daily requests at 99.9% uptime.
Built a cost forecasting platform on Tab Transformer foundation models and adapters; prototyped a human in the loop SHAP based explainability loop that lifted revenue 3%.
Defined and shipped AI driven renewal underwriting, cutting loss ratio 5%.
Improved vehicle failure prediction recall 12% through telematics feature engineering across multiple automotive OEM product lines.
Led Industry 4.0 and explainability initiatives across operations, finance, and procurement, including an NLP driven order status chatbot.
PhD research on visual attention, the throughline for the routing and attribution work above. See Research.
Full role history, metrics, and skills in the CV (PDF) →
Research · 2013 to 2019
Plus 3 conference presentations at the Cognitive Science Society and CEMS · Google Scholar →