Most AI in healthcare assumes three things: a stable internet connection, a qualified specialist nearby, and patient data you're free to send to the cloud.
Rural northern Nigeria has none of those.
That's exactly why I led the engineering team behind the Cervical Cancer Screening and Continuous Improvement System (CXCA) — an offline-first, AI-powered diagnostic tool built for frontline health workers in Kano State, deployed across 39+ facilities, and designed so that zero patient data leaves the clinic without consent and de-identification.
The Problem Nobody Was Solving
Cervical cancer is one of the most preventable causes of death among women in sub-Saharan Africa. The word "preventable" is doing a lot of heavy lifting there — because prevention requires screening, and screening requires specialists who aren't there.
In rural Kano, the nearest gynecologist or pathologist can be 60-70 miles away. The current method — Visual Inspection with Acetic Acid (VIA) — is manual, subjective, and dependent on the observer's training and experience.
There were existing AI models for cervical cancer detection. The problem? Every single one was trained on Caucasian patient data. They didn't generalize to African populations. That's not a minor gap — it's a clinical liability.
We weren't going to adapt a broken tool. We were going to build the right one.
The Architecture: Local Inference, Human-in-the-Loop, Data Flywheel
→ Local YOLOv8 Inference at the Point of Care: The model runs on-device. No API call. No internet dependency. A community health worker captures a cervical image, clicks analyze, and gets a result in seconds — whether the facility has connectivity or not.
→ EXIF Stripping Before Anything Leaves the Device: Every image is processed in-memory through Pillow before upload. GPS coordinates, camera serial numbers, device metadata — all stripped. The image that reaches our database carries no PII. Data sovereignty isn't a feature. It's the foundation.
→ Confidence-Based Escalation: Results below 90% confidence trigger an automatic escalation to a qualified clinician via email — with the image attached. High-confidence results are logged immediately. This protects patients without flooding the clinical review queue.
→ Blind Review Protocol with Round-Robin Assignment: Clinicians review images without seeing the AI's diagnosis first. This is intentional. Anchoring bias — where a clinician defers to the AI result rather than using their own judgment — would compromise the ground-truth labels we need for retraining.
→ The Data Flywheel: Every screening, every clinician label, every disagreement between AI and clinician feeds a retraining pipeline. When we accumulate enough locally-sourced Nigerian patient data — with a deliberate 70/30 negative/positive split — the model retrains. A new version is validated, versioned, and deployed. The system gets smarter with every scan.
The Stack
→ YOLOv8 (Ultralytics) — image classification, local inference
→ Streamlit — multi-role web interface (Provider, Reviewer, Admin)
→ Supabase — PostgreSQL + Auth + Object Storage with role-based access control
→ Python / Pillow — in-memory EXIF de-identification pipeline
→ Gmail SMTP — async clinical escalation notifications
→ Offline Queue — JSON-based local queue with exponential backoff retry for intermittent connectivity
Three distinct roles. Three distinct access levels. One shared goal: earlier detection.
Why This Matters Beyond Nigeria
The architecture is the real story. We built a system that can be replicated in any low-resource setting, for any diagnostic task — because the design principles hold regardless of geography.
→ Local inference removes cloud dependency
→ Blind review removes AI bias from ground-truth labels
→ The data flywheel removes demographic bias from the model over time
→ De-identification removes the legal and ethical barrier to data sharing
Most AI health tools are built for hospitals in San Francisco. We built this for the nurse 70 miles from the nearest specialist. That's a different problem — and it demands a different kind of engineering.
What's Next
The system is live. The facilities are configured. The pipeline is running. If you're serious about building AI for low-resource settings, I want to hear from you.