Talent Track 360

Seven roles: management, team lead and recruiter in a hierarchy, plus standalone accountant, analyst and office-admin roles, each seeing a different slice of the product. Access is enforced in three independent places: client-side route guards, API-route authorisation, and Postgres row-level security. The client guard is never trusted on its own, so a forged request still gets nothing back from the database. On top of that: single active session per user with forced logout of superseded devices, 30-minute idle timeout, a password policy with history and 90-day expiry, audited "view as user" impersonation for managers and TLs, and a full audit trail of every action.

~50 tables in Postgres, RLS enabled on all of them, with the trickier invariants pushed down to database triggers rather than the UI: ownership and month are pinned once a record is created, accountants can only touch invoice columns, and a candidate record that's already in the selections pipeline or carries invoice data cannot be deleted at all. Deleted entries go to a recycle bin that only management can restore from. All file storage is private, server-write-only, and read through short-lived signed URLs.

Screens CVs against configurable eligibility rules using GPT-5.6 Luna. Built for volume: token-efficient prompting, structured outputs, and a cached organisation-lookup table so the same employer or college is never researched (or billed) twice. Lookups are gated on whether the answer could still change the verdict, which cut the search bill substantially.

An in-house chat assistant on the same model with web search and inline citations, scoped to the work this team actually does instead of being a general chatbot. Folders, saved prompts, and a weekly background pass that distils personal preferences. Every request runs with retention disabled, so candidate CVs never sit in a vendor's dashboard.

Management can switch the whole organisation between OpenAI and OpenRouter from the UI. Model capabilities are verified against the provider's published catalogue rather than assumed, and per-answer cost is tracked in a usage dashboard split by provider and model.