work / vault-ai flagship2025 → presentshipping
Vault AI
An AI workspace built for trust, not demos. Multi-model workspaces, personas, knowledge bases, document processing, a realtime gateway, tenant isolation, audit verification and production guardrails. Built for teams whose AI use has to hold up to review.
Problem, build, break, lesson
problem
Most AI workspace demos collapse once a second tenant, a compliance reviewer, or a confidently wrong model shows up. Teams needed one place to use several models without losing control of their data or their audit trail.
what i built
A tenant-isolated workspace with per-tenant checks on every data path, human approval gates on consequential actions, verifiable audit records, and a realtime gateway that keeps voice and streaming inside the same guardrails.
what broke
Early guardrails lived in the UI. The first integration that bypassed the UI proved they had to move into the API layer, where they now sit next to the tenant checks.
what i'd do differently
Design the audit model before the first feature, not after the first customer asked for it. Every later system was easier because the trail already existed.
Run the audit simulator
try an export without approval, or a tenant leak