Sora

Role Assistant Foundry gives every role in a small company (about ten people) its own AI assistant that works inside the company's rules. The problem: owners want AI but fear leaked client data and AI deciding money or approvals on its own; employees fear AI the company can watch. So work gets pasted into outside chatbots. How it works: the model drafts, code checks, a person approves. - Draft: the owner answers 10 questions. Muse Wonder designs the org chart, and Muse Spark writes each role's assistant and the handoff form passed between teams. - Check: code rejects a chart with cycles, orphans or more people than the company has, and a form with any field about a person. Prices and margins are computed by code, the thresholds are the owner's, never the model's, and a handoff priced below cost stays blocked whatever its note says. A subsidiary can tighten its group's rules, never loosen them. - Approve: each step is approved on that role's own invite link, and the server signs a receipt for every approval. Privacy by design: chat stays in the employee's browser; emails, phone numbers, card numbers, keys and listed client names are masked in the browser before a chat message or a handoff's notes reach a model; nothing about how a person works is stored. Finance runs in Private mode on open-weights Muse Glimmer 30B and never switches model on its own. Measured on small author-written sets: rule questions answered 9 of 10 with the role playbook and company rules in the prompt vs 2 of 10 with a plain role prompt; 10 of 10 contradictions the rules cannot see flagged, with 1 false alarm in 10 clean packages; the live link walked end to end with Finance answered by Glimmer (endpoint awake). 790 Python test cases and 39 JS tests pass without calling a model. Try it (no sign-up, sample data, 24-hour sandbox): https://role-assistants-662084666675.us-central1.run.app Video: https://youtu.be/n30I40GJSMQ Code: https://github.com/hoachauphuoc/MetaAIHackathon