ZeroHR

Overview

Most AI screeners hand you a ranked list and ask you to trust it. ZeroHR does the opposite: it shows you how it understood the role, waits for your approval, and then shows its evidence for every score it gives.

The whole product is built around one idea — an AI you can check is worth more than an AI you have to trust. Built with a UK-based partner: scoring stays locked until a human approves the hiring brief and the rubric, and every score carries the evidence behind it.

The mechanic

Scoring is locked until the hiring brief is approved, and locked again until the rubric is approved. That isn't a setting or a best practice; it's enforced in the software. There is no path around it — not for the user, not for me.

Four steps, and two of them belong to the human:

1. It reads the role. Drop in a job description. The agent works out seniority, industry, team shape, and what the minimum requirements really are — then writes it down as a brief.
2. You approve the brief. Read what it understood, fix what it got wrong. This is a gate, not a notification.
3. You approve the rubric. It proposes weighted criteria for the role. Edit them, reweight them, add your own, or have it rewrite one. The rubric is yours before a single resume is opened.
4. Then it scores. Each resume is scored against your rubric, with the evidence for every criterion and a flag when that evidence is thin.

Why it's defensible

When a candidate asks why they were rejected — or a regulator does — "the AI decided" is not an answer.

• Evidence per criterion. Every criterion gets a score and the specific evidence behind it, quoted from the resume rather than paraphrased into a vibe.
• Thin evidence is flagged, not counted. When evidence is absent or tangential, the score is marked for human review. The model admits what it doesn't know.
• The maths isn't the model's. Weighted totals and the final recommendation are computed in code at temperature zero. Same resume, same rubric, same answer — every time.
• Blind by default. Contact details are redacted before a resume is sent for scoring. The model judges the experience, not the name, the address, or the phone number.

Honest economics

Every AI call is metered — which step, which model, how many tokens, what it cost. It's visible inside the product, per job, down to the individual call. Not a line item on an invoice you can't reconstruct.

Sovereign deployment

Some teams can't send candidate data to anyone else's cloud, and most vendors' answer to that is simply no. ZeroHR is built to deploy on your own infrastructure running local models, so candidate data never leaves your walls. It's scoped as a bespoke deployment rather than a checkbox on a pricing page.

Status

Early, and deliberately so — I'm working closely with the first teams using it.