The problem
HR teams wanted an assistant that does the work, not one that only answers questions. That means an AI with real write access, and it has to be safe.
What I built
Sana, an agent on the Claude API with 350+ typed actions. Every action has a schema, a permission check against the user's role and a risk level. Anything sensitive pauses for a person to approve.
Deterministic evals in the test suite, so a prompt or model change that breaks an action fails CI.
A multi-tenant platform on PostgreSQL with row-level security, so one company's data can't leak into another's.
Hardened sign-in: httpOnly cookie sessions, CSRF protection and TOTP two-factor.
CI/CD with parallel image builds and lockfile-keyed caching, which cut deploy builds from about 7 minutes to about 2.
Stack: TypeScript, Node.js, Next.js, PostgreSQL, Redis, Claude API, Docker
0 likes · 2 views · Oct 30, 2026
Multi-tenant HR and payroll SaaS. Its AI agent Sana runs 350+ actions via the Claude API with permission checks and human approval.
