αlphabell.
α · People · Felicity Anjali Sandirasegaram
Stylised avatar of Felicity Anjali Sandirasegaram

Felicity Anjali Sandirasegaram

Long-tenured contributor

Based in Colombo
Node-cell voronoi-19
ORCID 0000-8090-4524-2886

Research

Felicity is one of the senior contributors in the world-models axis. Her thesis at Cambridge was on operator learning in classical dynamical systems; her work at DeepMind extended that to learned environment models for robotics. At alphabell she co-led the compositional latent dynamics result that has become the canonical reference for operator-based dynamics models at scale.

Outside the headline result, she is one of the architects of the cross-cell simulation pool — the 70k-environment shared training pool maintained jointly by voronoi-19, bessel-04, and fermat-31. She is unusually willing to litigate the question of what counts as out-of-distribution evaluation; the lab's internal world-models benchmark suite owes most of its discipline to her insistence.

Felicity is based in Colombo and works closely with the Colombo university cluster's robotics group on sim-to-real transfer for their cable-manufacturing partner. She speaks Tamil, Sinhala, and English in roughly that order of fluency.

Background

Ph.D. physics, University of Cambridge, 2016. Undergraduate at Colombo (physics + mathematics, double major).

Prior to alphabell: University of Cambridge; DeepMind robotics team (2017-2020); Wayfarer Institute.

Selected publications

Full publications index →

Recent talks

  • Compositional dynamics — why operators matter, NeurIPS 2024
  • The 70k-environment study, ICML 2025
  • Counterfactual confidence in production, RLDM 2024
Working with

Felicity is currently part of node-cell voronoi-19, working under the World models research axis. The cell is open to substantive correspondence from researchers working on adjacent problems; route requests through voronoi-19@alphabell.com or directly to Felicity at felicity-anjali-sandirasegaram@alphabell.com.