Transcript: Episode 12

And I'm Sarah.

DIMER stands for Democratized Intelligent Model Exchange Repository — which is a mouthful but the concept is straightforward. Instead of every agency in the Philippines building its own flood detection model or damage assessment algorithm, they can pull pre-built, field-tested models directly from a shared national commons. NAICRI also includes NAIRA, an AI-as-a-Service hub — so agencies in Mindanao or the Visayas don't need ML engineers on staff to access these tools.

Because it's a model for solving one of the most persistent barriers in humanitarian AI deployment — the build-from-scratch problem. Every time a typhoon hits, someone somewhere is spinning up a new damage assessment tool from zero. NAICRI institutionalizes what's been ad-hoc. And the Philippines is the right place to test this — it sits in one of the most disaster-exposed corridors in the world. If DIMER works here, it's a template other disaster-prone countries can adapt.

Right, most LMIC AI health strategies are aspirational documents — vague language about equity and safety, and no operational infrastructure to back them up. SAHI is different because India is governing AI that is already running at population scale. The eSanjeevani telemedicine platform has logged 282 million consultations using AI-generated differential diagnosis recommendations. The Ayushman Bharat Digital Mission has national health IDs deployed. They're not imagining a future — they're building governance for a present.

Exactly. BODH lets any organization deploying health AI in India run their system through a standardized testing process before deployment. For humanitarian health NGOs working in India — Médecins Sans Frontières, Partners in Health, government health programs — this reduces procurement risk. You're not buying a black box, you're buying a system with a verified benchmark score. The governance principle that's getting attention is SAHI's third sutra: responsible innovation should be prioritized over cautionary restraint. That's a deliberate break from the WHO and EU precautionary models, and it'll shape how other countries in the Global South approach this.

Horacio is a live benchmark for AI-augmented tropical cyclone forecasting. The Joint Typhoon Warning Center is running AI models alongside traditional ECMWF and GFS models, and DeepMind's weather AI ensemble is in the mix. Rapid intensification — when a storm gains 35 or more miles per hour in 24 hours — has historically been the hardest forecasting problem. Horacio is intensifying rapidly, and the question being watched closely is whether AI models are catching that earlier than traditional numerical weather prediction. PDC Global is providing real-time dashboards for disaster managers in Madagascar, Mozambique, and the Mascarene islands.

The programme is called Strengthening Disaster Resilience of Health Care Facilities in Remote Islands. It's funded by Japan and delivered with UNICEF. The numbers are striking: 25 percent of Vanuatu's health facilities have no reliable electricity. That means no vaccine cold chain, no medical equipment running at night. 15 percent have no water access. 89 percent lack essential equipment for newborns. And this is a country that's been hit by multiple Category 5 cyclones and a 7.4 magnitude earthquake in the last two years.

It upgrades 20 primary healthcare facilities to withstand cyclones physically — structural hardening, safe water systems, improved sanitation, generator backup — and adds emergency preparedness training for healthcare workers. This will serve 30,000-plus people in remote islands. It's also relevant right now because Vanuatu is simultaneously managing an active whooping cough outbreak. Seven deaths so far. The ministry says transmission is declining, but vaccination campaigns are ongoing. When you're managing an outbreak, having functional cold chains and clean water in your clinics is not a nice-to-have.

Absolutely. Fiji, Tonga, Samoa face similar conditions. The programme's three-year implementation timeline is realistic for small-island contexts. It's not a one-time donation — it's a sustained upgrade.

This is primary data, not think-tank extrapolation. The consistent finding across similar surveys is that interest in AI tools is high but training gaps and organizational policy uncertainty are the two biggest brakes. What makes the HLA data valuable is the geographic breadth — 120-plus countries means this isn't just capturing well-resourced Northern NGOs. And on Thursday — that's February 26th — HLA is hosting a free webinar presenting the results. If your organization is building a case for AI investment or skills budgets right now, this data is what you bring to that conversation.

This is a smart pre-positioned shelter model. Airbnb.org — that's the non-profit arm — is integrating residential units into Mexico City's emergency housing infrastructure before a disaster hits. When an earthquake or flood activates the protocol, brigadiers and volunteers get immediate free housing, and affected populations can be placed using predictive demand modeling that identifies which neighborhoods need surge capacity. In October 2025, this was used informally during Veracruz floods and landslides. This formalization means the protocol is pre-negotiated — no scramble, no delays.

Right. And the model is replicable. Manila, Lima, Dhaka — any dense urban area in a seismic or flood zone could adapt this framework. The key is formalizing the partnership before the disaster, including AI-assisted demand forecasting for shelter placement. The Mexico City model is the clearest example I've seen of that working.

Copernicus provides real-time satellite data on floods, wildfires, and droughts. The European Commission's Joint Research Centre presented it as a cross-border tool — critical for West Africa, where floods and droughts don't respect national borders. ECOWAS is trying to integrate it into a harmonized regional structure so national agencies can act earlier and coordinate better.

There's always a catch with satellite tools. Copernicus is only as good as the ground-truth data being fed back into it. If national agencies aren't consistently uploading their ground-level measurements, the satellite projections lose accuracy at the local level. The last-mile data gap is still the bottleneck. The webinar participants flagged this explicitly. It's a useful reminder for any organization deploying earth observation or AI tools in the field: the model is the easy part. Building the data pipeline that feeds it is the hard part.

It's being called the AI slop crisis. AI-generated code contributions are flooding open-source projects — pull requests and bug reports generated by agents like GitHub Copilot, often without meaningful human review. The cURL project, which powers a huge amount of internet infrastructure including humanitarian data systems, shut down its bug bounty program after receiving 20 invalid AI-generated security reports in 21 days. Another project implemented a permanent ban on AI-generated submissions.

OpenStreetMap processors, crisis NLP tools, humanitarian data exchange utilities — these are volunteer-maintained projects. When a maintainer is spending 80 percent of their time rejecting AI spam, they have 20 percent left for real contributions. Some projects are moving to restricted contribution models that will slow development. For practitioners: if your technical staff are using AI coding tools to contribute to these projects, require rigorous human review before any submission. And if your organization benefits from open-source humanitarian tools, consider supporting the maintainers financially — Patreon, GitHub Sponsors, direct grants. They're absorbing a new cost that didn't exist a year ago.

Stay ready.