Skip to content

Interpretability research

CLIP-H

Clinical hypothesis verification on MIMIC-IV using sparse autoencoders and an LLM ensemble.

Role
Software Engineer, AI Research
Organisation
Purdue University
Year
2026
Status
Research
Interpretability research2026

0.844

AUROC against a synthetic oracle

Problem

Language models will generate clinical hypotheses all day. The hard question is which ones survive contact with the data, and whether you can show your work well enough for a reviewer to check it.

Approach

I built CLIP-H with Purdue and Harvard Business School faculty. The goal was hypothesis verification you could actually audit: sparse features you can name, an ensemble that disagrees usefully, and a validation setup that doesn't quietly grade its own homework.

What I built

A hypothesis verification pipeline over MIMIC-IV using Top-K sparse autoencoders to surface interpretable features, with a GPT and Claude ensemble scoring candidate hypotheses.

Architecture

  1. Top-K sparse autoencoders over MIMIC-IV representations, producing features sparse enough to be named and inspected.

  2. A GPT and Claude ensemble scoring candidate clinical hypotheses, where disagreement between models is signal rather than noise.

  3. Validation against a synthetic oracle with known ground truth, so verification accuracy is measurable rather than asserted.

Impact

Reached 0.844 AUROC against the synthetic oracle and certified 14 hypotheses with Purdue and Harvard Business School faculty. The work is being prepared for a NeurIPS submission targeted for September 2026.

0.844

AUROC against a synthetic oracle

Contact

Building something in this world? Let’s talk.

I’m always up for a conversation about agents, developer tools, or a product you think should exist. The fastest way to reach me is email.

baliutkarsh2@gmail.com