anil

Talk at the FP Launchpad kickoff event, April 13 2026, IIT Madras.

Anil Madhavapeddy on TESSERA, a foundation model for satellite Earth observation data that makes global geospatial analysis accessible through OCaml, Python and TypeScript interfaces.

https://anil.recoil.org/

I stayed at the Bose Einstein guest house at IIT-Madras, which has a rather phenomenal setting on campus with lots of wildlife around.

This visualisation projects TESSERA's 128-dimensional embeddings for over 210,000 hexagonal cells from England's CROME 2024 crop map into three dimensions using UMAP. Each point is coloured by its crop class.

Crops with similar seasonal growth patterns cluster together in the embedding space — winter cereals group tightly, as do permanent grasslands — demonstrating that TESSERA captures fine-grained phenological variation from the raw spectral-temporal signal without any labelled training data.

Browse it at https://geotessera.org/crop-umap.

Credit: Sadiq Jaffer, https://toao.com

This is the Tessera Zarr Explorer (https;//tze.geotessera.org) illustrating how to explore the TESSERA geospatial embeddings using just a webbrowser, HTTPS streaming and WebGL/wasm to do analyses such as classification and segmentation. See https://geotessera.org for more information.

A little screencast of a fully browser based streaming interface to manipulate TESSERA embeddings. All the classification and UMAPs run directly in a browser, with no server required aside from static HTTP serving of the embeddings!

The original codename for Docker Desktop was "Pinãta", so this was our launch party for the beta circa 2015!

Kanishka Narayan and Amanda Brock at the British High Commission party firing up the crowd

Wandering around the summit on Day 2

Mirror of https://www.youtube.com/watch?v=goDYPS6Wuhw

In this keynote, Professor Anil Madhavapeddy (University of Cambridge) explores how modern AI systems can help make sense of complex, fragmented environmental data — without losing transparency, accountability, or human judgement.

Drawing on cutting-edge research and real applications, the talk examines how AI can:

The keynote also addresses the risks of poorly designed AI systems, including model collapse, loss of traceability, and misplaced confidence — and outlines what responsible, verifiable AI should look like in nature-based applications.

This session is essential viewing for anyone working at the intersection of AI, ecology, conservation, or environmental policy.

Demonstrating how Bill uses his bell!