The ATS stays. The model never leaves.

The problem we were given

The ATS already held the records. Screening AI was the leak: CVs left the agency into a public or bring-your-own-cloud model. One desk handles many employer clients. A mix-up of candidates across clients is a reportable incident. Auto-reject without a trail is a legal problem.

What our engineers built

Parse, match and pack-draft on Foundation, inside the agency perimeter.

CV and job description parse for PDF, DOCX and scan, with per-field citation back to the page. Match and rank with an explanation, not a silent score. Draft outreach and interview packs. A recruiter sends them. RBAC so client A candidates cannot retrieve into client B desks. Bias and audit log on the same install.

The ATS remains the system of record. No CV, phone or national ID in a prompt that can leave the building.

What it is not

It is not a new ATS. It does not auto-reject. It does not scrape job boards. Offer and reject stay with a named recruiter.

Why the cloud answer fails here

A staffing firm is a processor for thousands of employer clients. One CV in a public model is a reportable incident. Ranking without an audit trail is a legal problem, not a UX one.

Controls in this workflow

These controls run on Private AI Foundation. Every ask goes through the same path.

• Work stays inside the customer perimeter. Inference does not require a network path out.
• You hold the encryption keys in every tier that touches your data.
• Every ask goes through the Model Gateway. There is no side door to a public model.
• Policy and DLP evaluate the request before weights run.
• RBAC names who may invoke which model class on which data class.
• Each step writes confidence and a citation back to the source. Originals are preserved.
• Consequential actions wait for a named approver. Nothing silent-posts to your core systems.
• Immutable audit is written on the same install: who, what, which model, which policy, outcome.
• Apache 2.0 and MIT weights only. Licence text and an SBOM ship on disk per release.

The paid proof of value

Thirty to forty-five days. Fixed fee. In the customer environment. One desk, one role family, agreed volume. Success criteria and conversion price agreed in writing before starting.

Success is time-to-shortlist, parse quality against a human, zero CV egress, and human-owned reject and offer. If it misses the threshold, the customer keeps the report and owes nothing further.