NAZAR

THE PROBLEM. In Uzbekistan, owners pay crews by the day — and a foreman can report 25 workers when 15 showed up. The prosecutor general's office found 31,087 non-existent construction jobs reported in a single year (anhor.uz, 2024). We build NAZAR — face-recognition attendance with geofencing — so the raw truth is recorded. But data is only useful if someone reads it, and the owner of a company with six sites and 120 workers has no time to page through check-ins. WHAT NAZAR AGENT DOES. A team of AI agents reads NAZAR the way an experienced administrator would and answers any question an owner actually asks — "who really comes to the Riverside site?", "what should I check before payday?" — in Russian, Uzbek or English, by text or voice. It finds workers checked in from the foreman's phone with photos of a screen, and crews marked "late" every day because the site schedule is wrong. Each finding is a card: the facts with links to the records, the innocent explanations, and an action with NAZAR's own before → after. Nothing changes until a human presses Apply; every applied action is auditable and undoable. The agent never accuses — a skeptic model defends every person a card is about. HOW IT'S BUILT. Models propose, code verifies, humans decide. A dispatcher reads the request; analysts investigate with read-only tools over the NAZAR API; a fact checker written in plain code looks up every id and number the model wrote and sends the report back until it is honest; a validator checks each proposed action against NAZAR's own pay preview; the executor — the only component with write access — runs only after a human confirms, atomically, with undo and an append-only audit trail. HONEST NUMBERS. 70/72 on a 72-question owner-question bank in three languages, scored by plain code — no model grades another model. After development froze, an outside reviewer wrote 15 held-out questions; first run: 13/15, published exactly as it came out, both failures described in the README. 39/39 against live injection attacks on the employee cabinet; 556 tests; every eval re-runs with one command — a judge can check our numbers, not trust them. PRODUCTION. NAZAR is live with paying companies, and the agent ran end-to-end against it: proposed counting a shift from the worker's real arrival, NAZAR previewed late 74 → 0 min, applied, audited, undone. A judge can try everything in 3 minutes with zero setup: https://demo-agent.nazarhr.uz