Six systems, three kinds of hard
What shipped, and the repo that proves it.
Kitchen Kontrol
AI that disappears into the work — staff finish logs with wet hands, bilingual, mid-service, because the kitchen can't pause for software.
Phase-based ops surface + HACCP-mapped forms + voice where hands are full. Evidence, not theater.
Canon Forge
Creative velocity without drift — identity, set, and shot locks that hold canon across providers, models, and cuts.
Express proxy with identity-lock prompts + canon stills in front of Gemini, xAI, Bedrock, Venice. The model renders inside the contract.
The Hermes Librarian
Vector recall and a knowledge graph in one Postgres — the model can find what you said, and you can see the walk.
Two stores, one semantic topology. Apache AGE keeps the session flower; Postgres holds the shared noun manifold with ordered mentions. One bounded walker scores both prefetch and /search.
NeuroNote
AI that can invent software, but never owns the runtime — proposals stay proposals until a state machine proves them safe.
Dual-kernel: Guest proposes, Gatekeeper verifies (structure → semantics → honesty), Host executes in WASM. Authority stays with the host.
AI/ML Ops Factory
The opinionated factory so teams stop reinventing scaffolding and start where the value is — manifest in, governed Azure ML repo out.
Terraform, OIDC, digest-bound apply, live taxi reference that trained and served in Dev/Prod. The factory is the product.
Tokyo Eye
AI as lab partner, not oracle — hyperbolic GNN + provenance that makes every inference retraceable, not plausible.
Poincaré embeddings, MoE routing, pgvector control plane, and a 20-tool coordinator bound to the open structure and MLflow gates.
Llora Workbench
Standardized, auditable fine-tuning — declarative playbook, gated MLflow registry, full provenance on a 4GB T1000 — so the small model you ship is the model you can prove.
corpus25_playbook.yaml declares datasets, thresholds, and stages; trainer.py stays untouched. CE→BCE without code change, r=4 beats r=8.