I build machine learning systems. Over the past decade, I’ve been the founding ML engineer for three products that went from zero to revenue, spanning computer vision, generative models, search, and document intelligence. Since 2025, I’ve run an independent engineering practice building search, retrieval and agent systems for clients in logistics, legal, and finance.
Earlier, I studied computational neuroscience at Brandeis.
Projects
Latent World Physics
Trained PPO controllers on a lunar lander with adjustable physics, with and without explicit physics parameters. Physics-aware agents learned smoother control, but often generalized worse under distribution shift.
World Model Ladder
Built and compared world models from linear predictors to recurrent state-space models, on prediction accuracy, long-horizon rollout, and choice of training objective.
World Model Audits
Designed diagnostic tests for latent decodability, predictive value beyond identity baselines, and action-conditioned physical consistency in JEPA and RSSM models.
orglens / scad
Persistent work records for coding agents: session histories, decisions, and handoffs across repositories and agent runs.
The Wooster Papers
A fan site for the Jeeves and Wooster books. Every scene is indexed by who is in it, where it happens, and what goes wrong, so you can find the one you half-remember. The indexing and the search are model-assisted.
Reach me at vignesh.saptarishi@gmail.com.