Trellis

What it does

Trellis is a sketch of what an AI control plane for a mission-driven education org actually looks like when you treat it as infrastructure. Every skill, permission, persona, PII rule, and data connector is a file in platform/. A change is a PR. A rollback is a git revert. The deployed app is the rendered state of the repo.

Five skills are live: parent-comms-attendance (streaming agent with handbook retrieval), lesson-plan-qa (answer grounded in citations), policy-handbook-summary, parent-contact-translate, and ops-ticket-triage (classifier). Four personas — Teacher, Principal, Ops Analyst, Admin — each see a catalog filtered by what they’re allowed to run. Three DLP rules run on both input and output. Every invocation is logged.

The interesting design choice

The manifests are the documentation. There is no separate “what does this skill do” wiki. If you want to know which fields the parent-comms-attendance skill is allowed to pull from the student-information connector, you read platform/connectors/*.yaml — which names scoped fields and, just as importantly, explicit forbidden fields. If you want to know what counts as PII, you read platform/dlp/*.yaml. The Zod-validated manifests are loaded at startup; a malformed YAML file fails CI, not production.

Every skill has a deterministic stub mode. Without GOOGLE_GENERATIVE_AI_API_KEY set, the full pipeline — policy → input DLP → retrieval → generation → output DLP → audit — still streams end-to-end, just with canned responses. A reviewer can clone the repo, run npm run dev, and see the shape without paying for a single token. A “stub mode” badge appears on the results panel so you always know which you’re looking at.

Explicit deferrals

No SSO. No database. No multi-tenant isolation. No cost tracking. In-memory state. This is a sketch of a control plane, not a production platform, and the README is clear about that. The point was to see whether “IaC for AI skills” felt good to live inside — writing a new skill as a YAML PR, rather than as a ticket for a platform team. It does.

Credits

Stands on shoulders: Vercel AI SDK (v6 streaming + Gemini gateway), Zod, js-yaml. The policy-as-code framing owes a debt to Open Policy Agent — though Trellis intentionally keeps the rule language dumb YAML, not Rego, because the target author is a PM, not a platform engineer.

Live · Source · PLATFORM.md