Back to our workCase studies

Medical AI that runs on the device in front of the patient.

Three screening systems for clinics with no specialist, no lab reader, and no connection to rely on. In each case the model runs on the hardware already in the room.

01Silicosis

A preventable lung disease that is easy to miss early

The problem

Silicosis is preventable, and mining and quarry communities carry most of it. Catching it early means a radiologist reading a chest X-ray, and the sites that need that reading most are the least likely to have one on hand. Referral adds weeks, by which point the early window has usually closed.

How we solved it

A compact convolutional network distilled from a much larger radiology model, small enough to run on the portable X-ray unit at the site itself. Training combined public chest X-ray data with imaging collected at the pilot sites, because portable hardware varies enough between locations that a model trained only on clean public data does not survive contact with the real thing.

0.4 seconds per read on-device and 81% sensitivity on early-stage cases across three pilot sites. It flagged early cases later confirmed by specialist review, while keeping false positives low enough not to swamp referral capacity.

Full write-up on faisel.in

02Urinalysis

Cheap test strips, expensive readers

The problem

Urinalysis strips cost very little and are available almost everywhere. Reading them accurately is the part that normally needs a lab instrument, which is exactly what a rural clinic or a screening camp does not have. Read by eye, the answer depends on the light in the room and the person holding the strip.

How we solved it

A smartphone camera reads the strip. The hard part is colour constancy: the same pad photographs differently under a fluorescent tube, daylight, and a phone torch. We use a reference patch on the strip itself to normalise for whatever the ambient lighting is doing, then compare each test pad against calibrated references.

Ten parameters in under two seconds, matching lab-reader results closely enough to support triage decisions across a range of handsets and lighting conditions. Now running in pilot.

Full write-up on faisel.in

03Heart auscultation

Hearing a murmur takes training primary care does not have

The problem

A murmur is often the earliest audible sign of structural heart disease. Picking one out reliably takes years of practice with a stethoscope, and most primary care settings do not have that on hand. The patients who would benefit most from early referral are the ones who get missed.

How we solved it

A digital stethoscope costing around $25 and a small convolutional model that runs on the recording device. Audio is converted to a time-frequency representation and classified into common murmur patterns, normal heart sounds, and background noise. Nothing is uploaded, so it works without a connection and the recording never leaves the room.

92% accuracy on murmur classification, holding up on noisy recordings made well outside a controlled clinical setting.

Full write-up on faisel.in

These are screening and triage tools. They are built to get the right person in front of a specialist sooner, not to replace the specialist.

Have a problem worth solving?

Tell us what is difficult, repetitive, inefficient, or technically challenging. We will help determine what should be built.

Or write to contact@lojits.com with a short description of the problem.