· Research framework
A practical framework for private local AI inference
Evaluate privacy boundaries, memory pressure, latency, quality, and recovery instead of relying on a single tokens-per-second number.
Read the frameworkAI research lab
Puma AI Private local inference, agent harnesses, and gaming experiments.
We study how useful AI can run closer to the user — practical methods for private local models, observable agent workflows, and game-like testbeds.
Research notes
Methods for practical AI research
Three working frameworks for experiments we want to make measurable and reproducible.
· Evaluation method
How to evaluate an AI agent harness
Measure the system around the model: tools, state, permissions, control flow, traces, failures, and human intervention.
Read the method· Testbed design
Playable environments as agent testbeds
Games offer visible state, bounded actions, repeatable scenarios, and failures that are easier for humans to inspect.
Read the proposalActive programs
What we are working on
Research tracks organized around privacy, control, and legible system behavior.
Private inference
Finding model, runtime, and hardware combinations that keep sensitive context close to the user.
Agent harnesses
Building evaluation loops that make tool use, state transitions, decisions, and recovery inspectable.
Playable evaluations
Using games as bounded environments for testing planning, memory, adaptation, and human oversight.
Long-term vision
AI that is private, steerable, and enjoyable
Private compute
More intelligence should run locally, with cloud services used deliberately and visibly.
Inspectable workflows
Agent systems should expose decisions, failures, permissions, and tradeoffs to their users.
Playable testbeds
Complex AI behavior becomes easier to understand when people can observe and interact with it.
Backed by builders
Angel and pre-seed support from leaders in crypto and AI.
Stefan Thomas (Interledger, BitcoinJS), Chris Larsen (Ripple), Anatoly Yakovenko (Solana), Illia Polosukhin (NEAR, Attention paper), Sridhar Ramaswamy (Snowflake, Neeva), Don Ho (Quantstamp, Orange DAO), Kartik Talwar (ETHGlobal, A.Capital), Jason Warner (poolside.ai, GitHub), Oleksandr Maksymets (Meta Superintelligence Labs), Evil Rabbit (Vercel), and many others.
Work with us
Help make local AI practical
We are interested in engineers and researchers who care about private inference, reliable agent systems, and unusually good test environments.