Roland Gavrilescu

From model libraries to agent infrastructure

Gavrilescu’s flaxvision repository ports selected computer-vision models from torchvision to JAX and Flax. Its initial work began through Google Summer of Code under Avital Oliver’s mentorship. He later worked on AI agents at Superhuman and on agent infrastructure at xAI before founding Introspection with Julian Bright. His current company focuses on making the experience accumulated by deployed agents reusable.[1][3]

Agents governed by the firm

Introspection’s current platform describes agents tested on a firm’s past work, with expert corrections becoming checks for later versions. It offers deployment in the customer’s cloud or on its servers and emphasizes ownership of agents, evaluations and models. These are the company’s product capabilities and deployment commitments, rather than claims of individual authorship or independently measured outcomes.[5]

In The Loop Is the Product, Gavrilescu separates an agent’s task execution from a surrounding process that studies traces, diagnoses mistakes and tests revisions. Failure patterns can become evaluations; recurring successful behavior can become instructions or skills. He calls for portable, versioned recipes that preserve these decisions across changes in models or providers. The point is to keep what a system has learned inspectable and under its owner’s control.[2]

His recruiting example shows an agent finding prominent candidates when its owner wants less obvious people with strong work. A human defines good sourcing; an agent can implement that preference as a test. Production experiments then ask whether users value the behavior.[2]