In Kyrgyzstan, a third of Asia’s water originates in mountains that most global weather models treat as a calibration gap. The forecasts that reach river basins like Naryn are built on grids designed for somewhere else, and the people downstream — communities, river managers, the Ministry of Emergency Situations — get warnings that arrive hours after a glacial lake has already moved. AI for Climate Action is KG Labs’ research direction for closing that distance: testing what machine learning can see at the resolution the terrain actually requires.
The work is organised along three reinforcing pillars.
An open forecasting model for Third Pole disasters. Built with the Vector Institute (Canada) and BE Node / Başlangıç Noktası — Turkish Philanthropy Funds (Türkiye), with sensor data contributed by the Internet Society Kyrgyz Chapter. The model began with heatwave prediction and is now expanding toward glacial lake outburst flood (GLOF) forecasting in the Tien Shan. The 2018 Catalogue of Glaciers of Kyrgyzstan — Landsat 8 imagery from 2013–2016 at 15-metre resolution, across 22 major river basins — anchors the training data on the Kyrgyz side. Public AI weather models currently operate at a 4-kilometre resolution; on terrain like the Pamir Alai, that is too coarse to see what the river already knows.
Geospatial localization of open foundation models. A parallel direction adapts already-developed, openly available foundation models — Microsoft Aurora, Google DeepMind’s GraphCast, and the European Centre for Medium-Range Weather Forecasts’ Destination Earth (DestinE) — to Third Pole conditions. The same cross-border team leads this pillar. The point is not to train a new global model, but to make strong general-purpose models legible on terrain that under-represents in their training data.
Edge AI and affordable sensing. Climate monitoring on Raspberry Pi, Arduino, NVIDIA Jetson, and similar accessible hardware that local institutions can deploy without long procurement cycles or vendor lock-in. Inference happens on-device. Intermittent connectivity is the norm. Firmware and model artefacts stay open. This pillar is co-organised with the joint ICTP–IAEA workshop on Edge Artificial Intelligence for Environmental and Radiation Sensing (Trieste, 14–18 September 2026).
A field reading from the Naryn valley sits underneath the technical brief: when the water turns from blue-grey to brown, something has moved up where the glacier sits behind its moraine wall. The reading is informal and accurate. It is also late. Heatwaves follow a seasonal logic — they are anticipated, even when they arrive ahead of schedule. Glacial lake outburst floods do not. When one fails, it fails fast, and the warning travels in hours, not days. The forecasting work exists to put a signal in the system before the river has to carry it.
Partners. Ministry of Emergency Situations of the Kyrgyz Republic; Ministry of Natural Resources of the Kyrgyz Republic; Central Asian Institute of Applied Geosciences (CAIAG); Institute of Water Problems and Hydrology, Academy of Sciences of the Kyrgyz Republic; Vector Institute (Canada); BE Node / Başlangıç Noktası — Turkish Philanthropy Funds (Türkiye); GFZ — German Research Centre for Geosciences; Tien Shan High Mountain Research Center; Internet Society Kyrgyz Chapter (sensor data).
Status. Active. Heatwave prediction architecture tested; GLOF forecasting in development. Sensing pillar being scoped. Companion field write-up: What the Mountain Already Knows. Full programme page: AI for Climate Action.