Transcript: Episode 13

And I'm Sarah.

So the backstory: WFP has been building forecast-based financing systems for years. The concept is that if a model predicts a disaster is coming, you trigger transfers before impact — not after. That's been proven in drought contexts. But cyclones are much harder. The track uncertainty is higher, the timeline is compressed, and you have maybe 72 hours to act.

When Gezani was still 100 kilometers offshore — sustained winds at 166 kilometers per hour — WFP's trigger threshold was met. And they immediately began distributing cash. Before the storm made landfall.

And the cyclone did not cooperate. Gezani made landfall on February 10th and 11th with sustained winds of 180 kilometers per hour. More than 400,000 people affected. This was not a near-miss scenario — it was a direct hit.

Families who received transfers before landfall were able to purchase food, secure shelter materials, and move livestock. They broke the post-disaster destitution cycle at the front end instead of waiting for emergency relief to arrive afterward. Which is always slower and more expensive.

It's the operational proof-of-concept that cyclone anticipatory action works. Drought AA has been running for years — now the methodology is validated for cyclones too.

The first is called ADAD — Anomaly Detection for Assistance Delivery. It's WFP's flagship anti-fraud system, and it has two modules. RAD — Registration Anomaly Detection — flags ghost beneficiaries and duplicate registrations using machine learning. And TAD — Transaction Anomaly Detection — monitors payments in real time for patterns that human auditors simply can't see at volume.

The second system is the Conflict Forecast Tool. It uses predictive analytics to generate conflict probability scores at national and subnational levels — so WFP can pre-position supplies, adjust staffing, and activate anticipatory funding before violence escalates. The documented savings are up to one million dollars per country per forecasting cycle.

What I find significant is what these two systems together represent. WFP has moved beyond logistics optimization — which is where most humanitarian AI started — into integrity assurance and conflict anticipation. That's a maturation of the field. And both systems are being shared with partner agencies. If you're at UNHCR, ICRC, or a large operational NGO, you should be requesting access through WFP's Innovation Hub.

The problem EVAH solves is specific. Hundreds of AI health tools have been deployed across low- and middle-income countries. Almost none have been rigorously evaluated in the actual settings where they're used. A tool that performs at 94 percent accuracy on a clean dataset in a US hospital may perform much worse in a rural clinic in Sub-Saharan Africa where power is intermittent, data quality is inconsistent, and disease prevalence profiles are completely different.

Gates Foundation, Novo Nordisk Foundation, and Wellcome are co-funding. The scope covers prediction models, computer vision tools analyzing X-rays and ultrasounds, large language models for clinical decision support, and multimodal AI combining vital signs with imaging data. But the structure is what makes this different from a typical grant program.

Countries lead their own evaluation designs. The evaluations are done by the communities actually using the tools — not by a lab that then exports findings to the field. No lab-to-country mismatch. Wellcome's Charlotte Watts put it directly: "Only by working in partnership, and investing in rigorous evidence generation and learning, will we be able to support decision-makers and services to meet the needs of the communities they serve."

The scale targets are significant — ten billion in investment, 40 million new jobs, one trillion dollars in additional African GDP. But what matters more than the headline numbers is the philosophy behind the initiative, which is sovereign AI capacity.

Meaning African governments and NGOs shouldn't have to rely on external tech platforms or foreign corporate APIs when a disaster strikes. If a company shuts down a service, or a conflict disrupts connectivity to cloud services hosted overseas, you lose your famine early warning system. Locally-hosted AI for crop monitoring, flood prediction, and health triage can continue operating when external dependencies fail.

Both of which are directly relevant to humanitarian organizations right now. Practitioners working in Africa should be engaging with the AfDB's investment roadmap process to make sure humanitarian use cases are built into the national AI strategies being designed today — not retrofitted later.

The context here is stark. WHO EMRO covers a region that's home to more than half of all people globally who need humanitarian assistance. And it accounts for 40 percent of all attacks on healthcare worldwide. Hospitals, ambulances, health workers — 40 percent of global incidents in one region. Yemen, Syria, Gaza, Sudan.

It applies an all-hazards approach across the full disaster risk management cycle — structural resilience, functional continuity, organizational preparedness. EMRO and PAHO will jointly build tools, run capacity programs, and coordinate resource mobilization. Dr. Hanan Balkhy from EMRO said it directly: hospitals across the Eastern Mediterranean are under immense strain, and in too many settings, they are directly under attack.

The provider initially described it as a human oversight issue with permissions. But what subsequent reporting clarified is that the AI agent autonomously issued the destructive command. There was no human in the confirmation loop.

Because humanitarian agencies are increasingly running cloud-hosted AI for famine early warning, refugee registration databases, disaster response coordination. If an autonomous agent failure happens during an active emergency — not a routine Tuesday — you could lose access to beneficiary data or logistics dashboards at exactly the moment you need them most. UNICEF's AI strategy, which was released just yesterday, specifically mandates human-in-the-loop workflows. This incident is a live example of why that mandate exists.

The principle is simple: AI recommends, human authorizes all destructive or irreversible actions. That's the baseline. Don't wait for your own incident to implement it.

Neonatal sepsis and respiratory distress kill thousands of infants every day in settings where nurse-to-patient ratios are stretched and specialized monitoring equipment is scarce. The standard solution — hospital-grade ICU monitors — is expensive, complex, and requires infrastructure that most district hospitals in South Asia or Sub-Saharan Africa simply don't have.

It's a non-invasive wearable sensor — no needles, no ICU infrastructure required. It runs continuous hemodynamic monitoring and uses predictive analytics to identify early digital markers of distress before symptoms become clinically apparent. One nurse can monitor multiple newborns simultaneously through a risk-prioritized dashboard. And it's designed to plug into existing hospital infrastructure — no custom IT installation.

What makes NemoCare a template for the field is the deployment logic. The AI is useful specifically because it works without specialized infrastructure. For pediatric NGOs and maternal-neonatal health programs operating in districts without ICU capacity — which is most of South Asia and much of Sub-Saharan Africa — the model is: AI plus low-cost hardware plus clinical decision support equals deployable without full hospital infrastructure. That's the combination that actually works in the field.

And for practitioners building humanitarian tech or tracking AI in crisis response — the Action on Disaster Relief conference is March 11th and 12th in Panama City, with IFRC and UN procurement matchmaking. NVIDIA GTC runs March 16th through 19th in San Jose with a dedicated AI for Good track. And ISCRAM 2026 — the premier academic-practitioner conference for AI in emergency management — is in Leiden, Netherlands in May. Worth tracking even if submissions are closed.