ilmVault for Institutions

The problem

In an AI-rich classroom a polished assignment says little. Schools need students to show real understanding of what they read, and teachers need structured, reviewable evidence rather than a general chatbot.

What I built

• A tenancy and role layer (institution memberships, admin/teacher/student checks) that sits outside the existing coaching engine instead of forking it.
• Roughly 100 Postgres tables across 10 Drizzle schema modules: institutions, subject vaults, assignments, guided learning, reading coach, reader, invitations, and AI.
• A Thinking Assignments system: 11 techniques with per-technique Zod schemas, AI question generation with validation and repair, an AI job queue, and secure fullscreen checks.
• A model registry across four AI providers, RAG retrieval grounded in the book text, answer guardrails, and speech-to-text for spoken answers.

Impact

Live at institute.ilmvault.ai for a 10-institution pilot, giving teachers rubric-scored evidence of how each student reasons about assigned reading.