Existing AI weather models look brilliant on global averages. Their forecasts fail in the regions that matter most.
Pravāh is an AI lab building foundation models for the electric grid. Working with utility partners across India, the US, and Europe, we kept hitting the same wall: the weather forecasts our grid models rely on are dramatically less accurate over the Global South than over the regions the field was built around. When the forecast misses a heatwave, a squall line, or a monsoon onset, the dispatch, reserve, and storage are all wrong, often at the moments the grid can least afford it. Indus is our scientific initiative to fix the layer underneath.
In under three years, AI weather models like GraphCast, GenCast, Pangu-Weather, and Aurora have matched the forecast skill that traditional numerical models took fifty years to reach. The leap is remarkable. But the models assume that the world observationally looks like the Global North. Which is incorrect.
These AI models train on ERA5, a record of historical weather built from a century of dense observation across Europe, North America, and East Asia. Where measurements are dense, ERA5 closely tracks reality. Over India, sub-Saharan Africa, and interior South America, observations thin out and ERA5 fills the gap with its own physics. An AI model trained on this signal learns a dataset only partially informed by reality, and errors compound at every forecast step.
The regions most exposed to climate risk are served by the models least equipped to see them: the Indo-Gangetic plain, the West African Monsoon, the Pacific-Andean rainfall coupling. The places where priors are weakest are the places that matter most.
A scientific program, executed by an AI lab.
Indus is a scientific initiative of Pravāh. Pravāh is an AI lab from Stanford training and deploying foundation models for the electric grid, enabling utilities to make real-time decisions on grid planning and operations to reduce blackout risks and the volatility caused by extreme weather. Pravāh is working with utilities across the US, EU, and India, and is funded by Khosla Ventures, Pear VC, and Conviction.
Pravāh operates at material scale. Across five Indian states, its utility partners serve several hundred million consumers and manage tens of gigawatts of peak demand, spanning monsoon-driven hydro, summer thermal peaks, and distribution networks absorbing rapid solar and wind growth into architectures designed for thermal baseload.
The work splits into two layers, and Indus is built around the split.
The operational layer is a 3 km, 15-minute system that powers Pravāh's commercial products for utilities. Energy is a high-value sector with serious budgets for forecast accuracy. Utility revenue sustains the engineering, observation infrastructure, and field deployment capability that everything else depends on.
The public layer is a 9 km, hourly-resolution model released free for any government, researcher, or development institution across the Global South. Zambia's national met agency, smallholder farmers in Bihar, and disaster response teams along the Brazilian coast are exactly the users where forecast skill changes outcomes most, and exactly where commercial monetisation is limited. A free public model is the right instrument.
Both layers share the same scientific core. Keeping them under one roof is deliberate: the science improves faster when grounded in operational reality, and the operational system improves faster when built on an open scientific base. And the best time to build this is now. Electrification is reshaping electricity demand worldwide. The energy crisis has made every percentage point of forecast skill economically tangible. Extreme weather is intensifying year on year. And AI architectures can finally turn sparse, heterogeneous data into useful forecasts. NOAA's NEXRAD and MRMS took four decades to build; the AI-era equivalent can be built in two.
Peer-reviewed research, deployed at national scale.
The team has published peer-reviewed research on foundation weather models, graph neural networks for power systems, and geospatial AI, and is deploying models at national scale across India.
Former World Bank economist; runs one of India's most popular YouTube channels, with 4.4 million subscribers.
Completed his Stanford PhD in Energy Science in two years and three months.
Previously led AI initiatives at Zoom, spanning graph ML and generative AI products used at global scale.
Co-developed Prithvi WxC with NASA, the world's first foundational AI weather model.
The first undergrad in IIT Patna's history to publish a paper, across 2,000 students.
