Colombia · La Mojana · Putumayo
Humanitarian flood response is information-bound, not resource-bound. The knowledge to save lives already exists in hundreds of PDFs that nobody can read in an emergency.
humaid puts a small AI on a satellite to spot the flood, sends a tiny alert to ground, and unlocks a pre-synced, role-specific Q&A on a local device that works without internet.
Two live model systems on this page: ask the knowledge base (471 bilingual Q&A pairs · Nomic + DuckDB) or run flood detection on a before/after Sentinel-2 pair (fine-tuned LFM2-VL-450M).
The problem
The information exists. The technology exists. People still drown, lose their homes, and re-learn the same lessons every cycle. The gap isn't knowledge — it's delivery.
UNGRD, IDEAM, the World Bank and Copernicus produce hazard atlases in GIS, SAR and hydraulic-modelling jargon. A campesino in Caimito or a JAC president in Sincelejito can't use them.
The lessons of Mocoa 2017 and La Mojana 2021 are written down — in hundreds of pages of reports. So WASH gaps get re-discovered, triage protocols get improvised, the same lesson is re-learned every cycle.
La Mojana is >50% cloud-cover during the wet season. The bandwidth from satellite to data centre to response team is the bottleneck. By the time the imagery is processed centrally, the actionable window is gone.
"Río Putumayo nivel crítico 11.5 m" tells the IDEAM duty officer something. It tells the family in the zona ribereña nothing — about whether to leave, with what, where the albergue is, who to call.
Cell towers fail, internet drops, roads cut — exactly when the most actionable decisions are being made. Cloud dashboards and web portals are unreachable in the first 6–72 hours.
The solution
SPACE
┌─────────────────────────────┐
│ small VLM on satellite │
│ (LFM2.5-VL-450M, ~450 MB) │
│ Sentinel-2 / Sentinel-1 │
│ image → ~200 B JSON │
└──────────────┬──────────────┘
│
▼
┌─────────────────────── EARTH ─────────────────────────┐
│ ┌─────────────┐ ┌──────────────────────┐ │
│ │ Community │ offline │ Local app │ │
│ │ station │ LAN │ - role-personalised │ │
│ │ (sync hub) │ ◄────► │ - phase-aware │ │
│ │ │ │ - region-specific │ │
│ └─────────────┘ │ - works offline │ │
│ └──────────────────────┘ │
└───────────────────────────────────────────────────────┘01
A 450 MB vision-language model fine-tuned on Colombian flood events runs on the satellite itself. It ingests Sentinel-1 SAR + Sentinel-2 imagery onboard and emits a ~200-byte JSON alert instead of downlinking the raw tile.
02
A solar-tolerant Raspberry-class node at a school, clinic, JAC or alcaldía. Receives JSON alerts via radio, satellite link or SMS gateway; serves the synced knowledge base over local Wi-Fi.
03
Phone, tablet, or laptop client. Works without any network once pre-synced. The same JSON alert routes to a different first screen based on who you are and where you live.
Both systems are running on this page.
The knowledge base side uses Ollama + Nomic embeddings + DuckDB cosine search — the same stack that runs on a Raspberry Pi at a community station.
The flood-detection side runs the fine-tuned LFM2-VL-450M directly via llama.cpp — the same compact build we want to run on a CubeSat.
The Zenú already solved this — once
La Mojana is a 500,000-hectare wetland in northern Colombia where the Cauca, San Jorge and Magdalena rivers meet. For roughly two millennia before the Spanish arrived, it was home to the Zenú— pre-Columbian hydraulic engineers whose canal-and-camellón system embraced the seasonal floods rather than fighting them.
The faint herringbone patterns still visible from satellites today are theirs. So is the population estimated in the hundreds of thousands — denser than today.
Then European contact, depopulation and centuries of neglect collapsed the system. The drainage logic was forgotten. By the 20th century, the response had inverted: build dikes that try to keep water out instead of channels that move it through. The Cara de Gato dike has broken in 2021, 2024 and 2025.
The Zenú had landscape-literacy delivered by living in the landscape. We won't get that back. But we can build a system that delivers an analogous form of literacy — fast, verified, localised — through a pipeline that doesn't depend on the very infrastructure floods destroy.
Why now
Traction
17 PDFs · ~60 MB
OCHA, UNGRD, ACAPS, CERF, FAO, IASC, UNICEF, IOM, IISD, SIDA, UNAL — all mirrored as searchable markdown. Plus ~80 curated external sources by theme.
471 Q&A pairs
Bilingual EN/ES, role × phase × region tagged, source-cited. Built by 6 parallel agents in ~16 minutes wall time. Local sub-second retrieval via Nomic + DuckDB.
LFM2.5-VL-450M
Full Modal H100 fine-tuning pipeline in Deno/TypeScript. Anthropic auto-labeller. 14 anchor locations, 9 events, ~115 paired samples. Sentinel-1 SAR re-attempt is a 2-week swap.
55-row geo-database
Static Mapbox viewer joining locations × events for La Mojana and Putumayo. JSON / CSV / GeoJSON export.
The ask
Funding, partnerships, and first-pilot access to take humaid from working prototype to deployed system before the next ENSO cycle.
If any of this resonates — start with a 30-minute working call. We'll show the components live and walk through the architecture.
hello@humaid.app