Yash Ravipati

Most of my research is here. I build deep learning models that extract clinical biomarkers from routine CT and MRI scans.

One focus is body composition: muscle loss (sarcopenia) strongly predicts survival in cancer patients, so I helped build a pipeline that segments neck muscle on head and neck cancer CTs and links it to survival and treatment toxicity. It is now being tested in clinical settings. I also worked on automated temporalis muscle measurement from brain MRI, which we used to build growth charts from over 23,000 scans.

More recently I worked on BrainIAC, a brain MRI foundation model trained with self-supervised contrastive learning. It adapts to seven downstream tasks, from brain age estimation to IDH mutation prediction, and outperforms supervised baselines when labeled data is limited. I've also worked on PET/CT tumor segmentation and on survival models that combine tumor and body composition features.

I worked on machine learning for brain-computer interfaces, with the goal of restoring communication for patients with brain stem injuries.

In one project, I benchmarked classical and deep learning models for P300 classification on EEG from 75 subjects. EEG-Inception, a CNN, significantly outperformed the standard SWLDA approach when trained across subjects, which points toward BCIs that don't need to be calibrated for each patient.

I also worked on predicting freezing of gait in Parkinson's disease and detecting epileptic seizures, and presented some of this work at international conferences.

I worked on subscription lines of credit, the loans private equity funds take out against their investors' commitments so they can invest now and call the capital later. The question was how much funds have come to rely on them, and what that changes for the pensions that back those funds.

I built the dataset behind the project in STATA, linking each fund's cash flows to its credit line balance over time, and used it to track how that borrowing has changed across fund vintages and sizes.

Give a language model a few labeled examples in its prompt and it can often pick up a new way of sorting things. I'm asking whether the geometry of its internal representations predicts how well it does.

The starting point is a theory from Sorscher, Ganguli, and Sompolinsky that predicts few-shot error from a handful of measurable properties of a representation, first tested on vision networks and the macaque visual cortex. I apply it to an open-weights LLM's hidden states, layer by layer, and compare its predictions with how the model actually does in context.

I'm also asking what happens when the examples regroup the same items by a different attribute (mammal or bird, domestic or wild): do the representations reorganize toward that grouping, and does the shift predict accuracy on new items? And with Anthropic's Jacobian lens, I read out the rule the model seems to have inferred and test whether geometry in that space predicts generalization better than the full hidden state.

Most work on RL for negotiation trains agents to negotiate. We flip it: LLM agents turn a negotiation case into a rehearsal environment that trains people.

The environment plays the other side with behavior we can steer, from how hard it anchors to how fast it concedes, and it adapts difficulty to the person practicing. It also infers a negotiator's strengths and weaknesses from how they play, and picks the next scenario to work on them.

Feedback comes after each session, through counterfactual credit assignment: at the turns that mattered, we fork the negotiation, try other moves, and roll each one out to see how the outcome would have changed. There are no live suggestions mid-negotiation, so the tool builds skill instead of doing the negotiating for you.