The Rad Journey by Bargava

Using AI to make better decisions

By Bargava Subramanian

· Launched 3 months ago

Did ORPO Actually Improve Our $50 Healthcare SLM?

The raw evaluation results: large reasoning gains, uneven specialty performance, and two safety regressions we cannot ignore.

Jun 18•Raghotham Sripadraj

Evals Are Where a Small Language Model Proves It Is Not Lying

How we built an evaluation system that could catch convincing answers, biased judges, and hidden clinical failures

Jun 11•Raghotham Sripadraj

Reinforcement Learning is where SLM learns what it got wrong

How we used the simplest preference optimization technique that could possibly work to teach a 1.7B parameter clinical model its own failure modes

Apr 23•Bargava Subramanian and Raghotham Sripadraj

Supervised Fine-Tuning is where a Small Language Model learns to stop guessing

How we trained a 1.7B parameter clinical reasoning model to think like a Doctor

Apr 14•Bargava Subramanian and Raghotham Sripadraj

Data Curation is the true moat: Here’s how we built a high-signal training data

Most AI teams over-invest in compute and under-invest in data. We did the opposite. And it worked.

Apr 5•Bargava Subramanian and Raghotham Sripadraj

We built a Healthcare SLM for $50 that matched frontier models for our task

Narrow specialized intelligence can be compressed.

Mar 24•Bargava Subramanian and Raghotham Sripadraj

Who am I and why subscribe?

About Bargava

Nov 4, 2019•Bargava Subramanian