I'm Sam Osei, a researcher working across two areas: Edge Medical AI — diagnostics that run entirely on-device — and Maritime Intelligence — perception and forecasting for vessels, ports, and coastal waters.
Based at the Center for Applied AI Systems. Previously trained in electrical engineering and machine learning.
Both parts of my work ask the same question in different settings: can a model still perform when it can't rely on a stable connection, a powerful chip, or a clean signal? In Edge Medical AI, that means diagnostic models that run entirely on a handheld device in a clinic with no signal. In Maritime Intelligence, it means detecting, tracking, and forecasting vessels using the sensors that actually exist on a ship or a shoreline — not an idealized feed.
I build and test these models against real deployment conditions rather than curated benchmarks, since that's where they actually have to work.
Models built to run entirely on handheld or point-of-care hardware, for clinics and field settings with unreliable connectivity.
A model that screens chest imaging for early signs of silicosis on portable X-ray hardware, aimed at mining and quarry communities where specialist radiology review is scarce.
A smartphone-camera-based urinalysis model that reads standard test strips and reports results without a lab reader, calibrated to work under variable ambient lighting.
A model that classifies heart sounds from a low-cost digital stethoscope, distinguishing common murmurs from normal variation directly on the recording device.
Detection, tracking, and prediction systems built around the sensors that actually exist on ships, shorelines, and satellites.
Forecasts a vessel's near-term path and estimated arrival time as it approaches busy port waters, using historical traffic patterns alongside live position reports.
Identifies vessel type and likely size class from satellite and coastal camera imagery, tuned for the resolution and weather conditions typical of operational feeds rather than clean benchmark sets.
Estimates a ship's distance from a fixed shoreline camera using monocular depth cues and known camera geometry, paired with the same classification model used on satellite imagery.
Reconstructs common shipping routes from historical position data and flags vessels whose current path deviates from expected behavior for their type and location.
Recovers usable detail from satellite imagery degraded by haze, cloud cover, or rain streaks, so downstream vessel detection and classification models have a cleaner frame to work from.
Illustrative comparison — drag to compare.