Hi, I'm Saran
RSS FeedI build production data pipelines by day. Here I document building evaluation and reward infrastructure for clinical AI: a verifiable clinical extraction environment on MIMIC-IV, GRPO fine-tuning against real verifiers, judge calibration, and the transformer stack rebuilt from scratch.
The rule is one public artifact every Friday. The posts below are the receipts. Code lives on GitHub .
Featured
-
What a Data Engineer Misses About Transformers (Until You Build One)
I move data through pipelines for a living. Then I built the transformer stack from scratch, from autograd to attention. Five places my instincts pointed the wrong way, and three bugs that never crashed.
-
Why I'm Spending 8 Weeks Building Toward Verifiable Clinical AI Environments
One clinical extraction environment, built once and reused three ways, as an evaluation suite, an RL reward signal, and a governance instrument. The plan, the constraints, and the rule that everything ships on Fridays.
Recent Posts
-
Four Days of RL: One Gradient, Dressed Four Ways
Bellman as a fixed-point computation, watching behavioral cloning fail on schedule, and why REINFORCE is the ancestor of everything in modern post-training. Notes from four days of foundations.