A small lab for inspectable AI.

Why this lab exists

Useful AI should not require sending every private thought, document, or interaction to a remote service. Local models create another option, but only when the whole system—hardware, runtime, interface, and fallback—is usable.

Likewise, an agent is more than a model. Its prompts, tools, memory, permissions, and control flow determine what it can do. PumaAI researches the harness as a first-class system.

How we work

We prefer small reproducible experiments to broad predictions. A useful research artifact explains its setup, records representative failures, distinguishes measurements from interpretation, and leaves enough detail for someone else to challenge it.

Memory store design is part of the method. Durable, inspectable context makes long-horizon agent behavior easier to understand without hiding complexity behind a single prompt window.

Founder

Yuriy Dybskiy

Yuriy is the founder of Puma, a private-by-design mobile browser, and leads PumaAI. His background spans engineering and developer relations at Cloudant, Meteor, Parse, and Lyft.

Personal site · LinkedIn

Editorial policy

Authorship and updates

Research notes are published by PumaAI Research unless an individual author is named. Pages include publication and modification dates. Material updates should describe what changed; unmeasured proposals remain labeled as frameworks or research agendas.

Corrections and research conversations are welcome at [email protected].