KG Labs leads climate AI research in the Third Pole — testing machine learning architectures on terrain that most global weather models treat as a calibration gap. The work began with heatwave prediction and has since evolved toward glacial lake outburst flood forecasting.
The climate direction works along three pillars that reinforce each other: an open AI forecasting model built with international partners, a geospatial layer that localizes open foundation models for the Third Pole, and an accessible sensing layer (Edge AI / TinyML) that can run in places where global infrastructure does not reach.
The Third Pole is not a calibration gap. It is the place where a third of Asia's water originates, and the place where the people downstream are owed forecasts that match the terrain they actually live on.
KG Labs is developing an open AI model for forecasting natural disasters in mountainous Third Pole countries — beginning with Kyrgyzstan. The work is built in partnership with the Vector Institute (Canada) and Turkish Philanthropy Funds — BE Node / Başlangıç Noktası (Türkiye). KG Labs provides the research framework and field grounding; Vector contributes model architectures; the Turkish partners support the cross-border collaboration that makes the project viable. The Internet Society Kyrgyz Chapter contributes datasets from its sensor network, supplying ground-truth observations from terrain that public meteorological grids do not cover at usable resolution.
The model began with heatwave prediction and is expanding toward glacial lake outburst flood (GLOF) forecasting in the Tien Shan. Both problems share a common difficulty: the Third Pole is under-instrumented, its terrain is steep and mixed, and most publicly available training data was assembled for somewhere else.
The forecasting work is augmented by a research direction in geospatial AI: taking already-developed, openly available foundation models and localizing them for Kyrgyzstan and the wider Third Pole. This is where the power of AI becomes legible to people on the ground — not by training a new global model, but by adapting strong general-purpose models to terrain that under-represents in their original training data.
Models we are exploring as starting points include Microsoft Aurora, Google DeepMind's GraphCast, and the European Centre for Medium-Range Weather Forecasts' Destination Earth (DestinE). The same cross-border team — KG Labs, the Vector Institute, and BE Node / Turkish Philanthropy Funds — leads this direction.
The third pillar is hardware: sensor devices for climate monitoring in places where commercial weather and hydro-meteorological infrastructure does not reach. The brief is deliberately practical — Edge AI and TinyML running on Raspberry Pi, Arduino, NVIDIA Jetson, and similar accessible and affordable platforms that local institutions can deploy without long procurement cycles or vendor lock-in. Inference happens on-device, intermittent connectivity is the norm, and the firmware and model artefacts stay open.
This pillar is being scoped. Team members have not been defined yet, and intentionally so — the device, deployment, and partnership decisions will determine the right people to bring on, not the other way around.
Edge Artificial Intelligence for Environmental and Radiation Sensing · Trieste, 14–18 September 2026 (smr 4190).
KG Labs is co-organizing this workshop with ICTP and the IAEA. The programme is directly aligned with this pillar: deploying Edge AI on accessible hardware for environmental sensing in under-instrumented contexts. Application deadline: 15 May 2026. Research abstracts (poster or talk) are encouraged in the application form.
Event page · indico.ictp.it/event/11116 · KG Labs co-organizer credit will be added to the ICTP page.
Three datasets that anchor the programme. The full repository of seven datasets (1967–2018, 534 extracted tables) lives at /climate-datasets/.
Comprehensive re-survey of all glaciers in the Kyrgyz Republic, comparing 2013–2016 state against the Soviet-era catalogue baseline across 22 river basins. Landsat 8 imagery at 15 m/pixel; full Markdown conversion plus 464 extracted data tables. Shabunin, A.G. (Ed. Moldobekov B.D.), CAIAG, Bishkek.
Machine-readable edition of the Soviet-era glacier catalogue (Каталог ледников СССР, Том 14, Выпуск 2), compiled from Parts 1–11. Eleven basin JSON records plus two manifest files and a README. Source: Hydrometeoizdat 1967–1986; compiled by KG Labs, 2024.
Web platform for the Central Asian sensor network jointly operated by CAIAG and the German Research Centre for Geosciences (GFZ). Hydro-meteorological in-situ station data across the region, with API access. DOI: 10.5194/essd-13-1289-2021.
The climate programme is built with a mix of Kyrgyz government bodies, regional research institutions, and international partners. The mix is deliberate: forecasts and sensing devices are only useful if the agencies that respond to disasters and manage water are part of the work from the start.
The same cross-border team works across the forecasting and geospatial localization pillars — research framework and field grounding from KG Labs in Kyrgyzstan, model architectures from the Vector Institute in Canada, and partnership with Turkish Philanthropy Funds (BE Node / Başlangıç Noktası) in Türkiye. The sensing pillar (Pillar 3) is being scoped and does not yet have an assigned team.
KG Labs actively seeks research partners, institutional collaborators, and development support for Phase 2 of the Third Pole forecasting project — expanding from heatwave models to glacial lake outburst flood prediction.
Support This Work See All Research