An embeddable Incremental View Maintenance engine. Declare a SQL view once, mutate the base tables, and the view stays correct by applying only the deltas, never recomputing. It implements the DBSP model, where every row carries a signed weight (insert +1, delete −1), so what streaming databases run as a separate service becomes an importable library.

A from-scratch replication of the TRINITY coordinator. A frozen 0.6B encoder maps each question to a 1024-dim vector, and a ~10K-parameter head picks which of three open-source LLMs answers and whether it acts as Thinker, Worker or Verifier, over up to 5 turns. The choice is discrete, so the head is trained with separable CMA-ES, an evolution strategy, instead of gradients.

One deployment serving many independent businesses. Tenant isolation is enforced centrally rather than left to each view: tenant IDs are signed into the JWT and checked on every tenant-scoped route, and the X-Tenant-ID header is stripped on cross-origin requests to block forgery.