Seven research groups, ten active projects

This project develops triage-support models that surface the specific evidence behind a recommendation, so clinical staff can agree, override, or ask for more information rather than simply trusting a score. The team is now running structured evaluations with partner clinicians to test whether the explanations actually change decisions for the better.

Home and campus IoT devices are cheap, numerous, and rarely patched. This project builds a lightweight detection layer that learns a device's normal behavior and flags deviations without needing constant vendor updates. A working prototype is deployed on a segment of the campus residence network under controlled conditions.

Hospitals and clinics often can't share patient data directly, even for research that could help their own patients. This project's federated learning framework trains shared models across institutions without the raw data ever leaving its source. Two partner hospital research networks are using an early version in production.

Formal verification is powerful but expensive to run continuously. This project investigates lightweight verification checks that can run on every code change, catching a meaningful subset of critical defects without the cost of full formal proofs on every commit.

People calibrate their trust in a conversational system based on cues that have little to do with its actual reliability. This project prototypes interface patterns that communicate a system's real confidence and limitations, then tests whether users' trust shifts to match.

Municipal planning offices need to know how neighborhoods will actually experience heat, flooding, and infrastructure strain under different climate scenarios. This project's simulation models are now in active use by two regional planning offices to evaluate proposed interventions.

Edge computing only pays off if the network gets requests to the nearest capable node fast enough. This project evaluates routing strategies against real regional traffic traces, in partnership with a local internet service provider that supplied anonymized data for testing.

Most consensus protocols assume more memory and bandwidth than a cluster of low-power devices actually has. This early-stage project is scoping a consensus approach designed from the start for constrained hardware rather than adapted down from a data-center protocol.

Automated hiring tools are widely used and rarely audited by anyone outside the vendor. This project builds an independent auditing toolkit that labor and civil-rights organizations can run themselves, without needing access to a vendor's internal model.

Strong cryptography fails in practice when people can't use it correctly under normal conditions. This project redesigns encryption prompts and key-management flows for ordinary applications, then tests the redesigns against the defaults most software ships with.