Target length: 3 to 5 minutes.
Most AI demos stop at chat. This repository shows three enterprise AI systems that handle the controls real deployments need: permissions, citations, approval gates, release gates, traces, audit logs, and evals.
Show http://127.0.0.1:8765.
Narration:
This is a permission-aware knowledge copilot. Alice is an employee. She can ask about remote work policy and gets a cited answer.
Ask:
How many days per week can employees work remotely?
Then:
Now Alice asks about a confidential finance plan. The system detects that relevant evidence exists but is not accessible to Alice, so it abstains instead of leaking or hallucinating.
Ask:
What is the finance retention plan?
Then switch to Morgan:
Morgan is a manager, so the same question returns a finance citation.
Show http://127.0.0.1:8770.
Narration:
This agent handles a regulated product-recall workflow. It can investigate, create internal records, draft a seller notice, and schedule follow-up. It cannot send the notice directly.
Run:
Check whether Market Blue still has an active listing for the recalled RX-900 product.
Point out:
- tool calls
- approval request
- blocked direct side effect
- trace ID copy button for connecting the UI result to audit evidence
Approve as supervisor:
The notice is only sent after supervisor approval.
Show terminal:
python -B scripts/dev.py verifyNarration:
The repo includes health checks, evals, smoke tests, and a public-release quality gate. This is what keeps the demo from silently regressing.
The important design choice is that the model is not the security boundary. Permissions, approval gates, audit, traces, and evals live in the application layer.