Shreyas Fadnavis

In a past life I worked at the intersection of imaging and machine learning, and picked up some applied math and theory along the way. Now building at Bioscope AI, doing all things AI.

coverage

Conformal Prediction

Your model gives you a number. How wrong could it be? This sequence starts with that question and traces a path through distribution-free prediction intervals, the impossibility of conditional coverage, variance stabilization, adaptive methods, and what happens when the foundational assumption of exchangeability breaks down.

8 notes · From first principles to open problems
Your Model Is Confident. Should You Be?
Browse all notes →
influence

Leverage Scores

A single diagonal of a matrix tells you where your model is extrapolating, how sensitive each prediction is to the training data, and why some points matter more than others. This sequence builds from the classical hat matrix through regression diagnostics, the sign flip that most methods get wrong, and onward to randomized algorithms and high-dimensional extensions.

7 notes · From classical regression to neural tangent kernels
The Hat Matrix
Browse all notes →
collective

Agentic AI

One model hallucinates. Multiple models, organized into a collaborative architecture, can cross-check, correct, and refine each other. This sequence traces the theory from ensemble error decomposition and Condorcet's Jury Theorem through Mixture of Agents, social choice theory, reasoning architectures, agentic retrieval, and the scaling laws that govern when more agents help and when they don't.

6 notes · From ensemble theory to multi-agent scaling laws
Why One Model Isn't Enough
Browse all notes →