Federated learning across institutions
Institutions that cannot hand each other their data still need to learn from each other, so the model travels to the data instead of the other way around. Each site trains on what it holds and sends back only what the model learned, never the records. The hard part is that sites are not interchangeable: their populations, their record-keeping, and how well they follow up all differ, so naive averaging quietly buries the ones that matter. Hospitals are the case I work on. Banks and insurers sit under the same constraint for their own reasons.
Related publications
Survival analysis and time-to-event
Survival analysis asks not whether something will happen but when, and it has to handle the cases whose outcome nobody has seen yet because they are still running or stopped being observed. That is usually the question people need answered: when will this patient get worse, when will this borrower default. Insurance pricing and equipment maintenance run on the same question. My own work is on clinical data, with Cox proportional hazards models and deep survival networks trained across many centers at once.
Related publications
Trust and privacy between institutions
To know whether another institution is helping your model, you would like to inspect what they sent. That is exactly the thing federated learning exists to prevent, and any federation of near rivals runs into it. My work builds reputation between sites from privacy-protected versions of their updates: enough signal to tell whether a peer helped, not enough to learn anything about a patient. Hospitals are where I measured it.
Related publications
Decentralized and edge systems
A central server is a convenient assumption and a bad single point of failure, and plenty of networks have nobody that everyone would trust to run one. I study peer-to-peer training where the network of who talks to whom is sparse, keeps changing, and has to earn its connections, along with what it takes to run these models on small hardware at the edge of a clinic rather than in a data center.
Related publications
Language models and agent systems
Most agent work assumes somebody else is paying for the compute: a large model behind an API, context to spare, a network that is always there. Take those away and most of what we have built stops working. That gap is the interesting part. The setting I know best is clinical, where the work is mostly text and language models are finally good enough to be useful on it, as long as the reasoning can be shown and checked. I work on fine-tuning, retrieval, and agent tooling for clinical decision support.
Related publications