LastMile

Connecting Geospatial Intelligence to Frontline Community Emergency Alerting

Comment

Inspiration

ICPAC — the IGAD Climate Prediction and Applications Centre — already runs the region's early warning infrastructure: Drought Watch, Flood Watch, Hazards Watch. What they don't yet have, by their own admission through the EarlyWarning4IGAD initiative, is a way to turn a hazard forecast into impact-based guidance — who specifically is affected, and what they should specifically do. The UN's own 2026 review of the Early Warnings for All initiative says the same thing at the global level: prediction is the strong pillar now, thanks to AI. Warning delivery and turning forecasts into action are what's lagging.

LastMile doesn't compete with ICPAC's forecasting. It builds the layer everyone agrees is missing.

What it does

LastMile takes live rainfall data for any location across the IGAD region, classifies drought or flood risk using a transparent, published-standard threshold method, cross-references real OpenStreetMap infrastructure data to identify exactly which hospitals, schools, and water points sit inside the affected area, and uses the Gemini API — constrained to only rephrase real, structured facts, never invent them — to generate short, specific, actionable alerts in English and Swahili for three different audiences: farmers, pastoralists, and urban flood-prone residents.

Every alert is auditable: the app shows the exact live API calls, timestamps, and underlying numbers behind each generated message, so nothing is a black box.

Dissemination goes beyond a web dashboard: alerts can be converted to spoken radio-broadcast audio and sent live via SMS through Africa's Talking, for communities without smartphones or reliable data.

How we built it

  • Data layer: live calls to Nominatim (geocoding), Open-Meteo / NASA POWER (rainfall), and the Overpass API (OpenStreetMap infrastructure) — no cached, seeded, or fabricated data anywhere in the pipeline.
  • Risk classification: a transparent percent-of-normal rainfall calculation against historical averages, bucketed into documented threshold bands, shown on-screen with the underlying math.
  • AI layer: the Gemini API, with a system instruction that explicitly forbids inventing any fact not passed to it, generating per-livelihood, per-language alerts as structured output.
  • Dissemination: Web Speech API for instant local audio, Gemini audio for studio-quality narration, and the Africa's Talking SMS sandbox for real telecom delivery.
  • Built end-to-end in Google AI Studio's Build mode, deployed to Cloud Run.

Challenges we ran into

Two, and both taught us something worth mentioning: an Overpass query initially returned zero infrastructure results for real towns because of a radius unit mismatch — a good reminder that "live data" isn't automatically correct data, and impact-grounding claims need to be tested against known ground truth, not just trusted because the API responded. And getting cloud-based TTS to route through the server-side runtime (rather than failing against a client-side route) took debugging — a useful lesson in how differently AI Studio's server vs. client execution contexts behave.

We also made a deliberate choice not to fake cell broadcast integration, which no hackathon team can actually access — that's presented as a clearly labeled concept screen, not a live claim.

What's next for LastMile

A real telecom partnership for cell broadcast delivery — the dissemination method that reaches every phone in range without needing a contact list or app install, which is the actual gap in "weak communication infrastructure" regions. Expansion beyond drought/flood risk into the impact-based forecasting ICPAC has only piloted for two hazard types in one country. And direct integration with ICPAC's own Hazards Watch and IGAD's IDDRSI cross-border coordination mechanisms, so this becomes a translation layer on top of existing regional infrastructure rather than a standalone tool.

Built With

Share this project:

Updates

Submission history