We design and deploy AI that lives where it's actually used: clinics, sensor networks, and field settings, built to stay reliable long after the benchmark is over.
So we don't stop at the model. We build the full system around it: the hardware, the data pipelines, the evaluation, and the monitoring that keeps things working over time, and treat them as one system, not separate problems.
Models trained on local, real-world data, and calibrated to hold up across the populations and sites where they'll run.
AI that runs on low-cost hardware in the field: ultrasound probes, sensors, and devices with real compute limits.
Systems tested where they'll actually be used: clinics, communities, and field settings, then monitored so they keep working.
Five active programs, each deployed and tested where it will actually be used. Filter by area, or swipe through.
Health
Chest X-ray models trained on Ugandan clinical data and evaluated across multiple screening sites, tuned for the sensitivity a screening tool actually needs.
Why it matters: missing true cases isn't acceptable, even when overall accuracy looks good on paper.
Health
Evaluating and deploying AI-assisted screening in Uganda, fitting it into existing workflows, structured data capture, and clinical reference standards.
Current focus: moving beyond model performance into regulatory readiness.
Health
AI built around portable ultrasound, from data collection in real scanning sessions to interpretation models running on low-cost hardware.
Constraint: it has to work with limited compute and variable image quality.
Datasets and speech systems for underrepresented languages: recognition, text-to-speech, and radio or voice-based interaction built to run in real environments.
Impact signal: several datasets are open and already used by outside researchers.
Sensing
Sensors, data pipelines, and models that watch real-world environments: traffic and mobility, environmental sensing, real-time processing.
Recurring challenge: systems degrade unless they're actively monitored and maintained.
Five programs shown here. More are running in the lab right now.
See the full research indexReleased for global research use.
None of these are endpoints — they are part of an ongoing process of building and improving systems in the field.
Most systems break the moment conditions change. So we treat deployment as the start of the work, not the end.
This is where most systems fail.
It's also where most of our work sits.
Where possible, we move projects along a path: from an idea to something running in real workflows.
Not every project reaches the final stage, but that is always the direction of travel.
We're deliberate about this. The goal isn't to force commercialization; it's to make sure useful systems continue to exist and improve.
Supporting external teams to evaluate and deploy their AI systems
Building reusable platforms for monitoring and validation
Exploring partnerships where systems can be sustained beyond research funding
What matters is how these pieces come together under real constraints.
Shared GPU infrastructure for large-scale model training.
Curated, governed datasets for research and deployment.
Applied research at the intersection of AI, health, and surveillance.
Training the next generation to build and deploy clinical AI.
Some partnerships begin as research projects. Others start with a practical deployment problem. Either way, the goal is the same: to end up with something that works in practice.
Trusted by partners across research, health, and industry
If that sounds familiar, we should talk.
Start a conversation