Yash Dharmadhikari
Senior AI Product Manager — agentic systems, evals & GPU-accelerated ML
I ship production agentic AI end to end: the model, the evals, and the roadmap. I came up through data science, so I'm as comfortable in a fine-tuning loop or a CUDA migration as in a PRD. Co-inventor on a pending U.S. patent. I write here about AI/ML product and engineering, most weeks.
Selected work
Self-Improvement in Production Isn't One Flywheel
Recursive self-improvement as an operational architecture: failed queries are banked, mined into specs, auto-built by a coding harness, and deployed only through a certification gauntlet. Pre-registered and scored against ground truth on a public benchmark: a 25.4% mean drop in unanswerable traffic, with the remaining ceiling traced to routing, not building. Read the write-up →
Latest writing
-
Self-Improvement in Production Isn't One Flywheel
Recursive self-improvement (RSI) as an operational architecture: one shared skeleton, a family of certified axes (three of them mapped in this series), and, in this first part, the axis I would ship first.
-
Productizing Loop Engineering
An enterprise blueprint: an agentic harness that builds your automation, and a Control Plane that proves, deploys, and safeguards it.
-
Hello — what I write about here
A quick intro: I write about both AI/ML engineering and AI product management — agents, evals, fine-tuning, and what it takes to ship them.