Catalyst

Mission

Catalyst is building AI systems where corrections, approvals and outcomes become inspectable improvements in future agent behaviour.

We're interested in systems that improve through use, without hiding how they changed.

Here is the part that matters: a correction is evidence, not permission to rewrite a person into a permanent profile.

A useful system should keep the source. It should propose a narrow change. It should apply that change only to matching work. It should compare the next result with the old behavior. It should let the user limit or roll it back.

We think this layer should survive a change of model or tool. Your working history should not belong to one chat window.

We don't know if the full idea works yet. Meraki Core has tests for the local correction, guidance, evaluation and rollback loop. The next job is proving it with invited users and a live database without losing the controls.

The sentence we keep returning to is simple:

Every correction should improve the next result.

If we cannot show the evidence, the change, the evaluation and the rollback path, we do not call it learning.