I’m working on a deterministic execution layer for AI systems.

The idea is to treat LLMs as constrained components inside explicitly defined workflows: strict input/output schemas, validated DAGs, clear failure modes, and replayable execution. Most “AI unreliability” I’ve seen isn’t model-related — it comes from ambiguous structure and hidden assumptions around the model.

We’re exploring this through a project called FACET, focused on making AI behavior testable, debuggable, and reproducible in the same way we expect from other parts of a system.

Still early, but the goal is simple: less magic, more contracts.