ADVIS Placement Engine — Model Card

The placement engine helps candidates discover open roles that fit their experience and helps recruiters surface plausible candidates from the ADVIS pool. It produces an explainable shortlist with matched and missing skills — it does not make hiring decisions.

• Inputs: CV text the candidate uploaded, the open + approved role list (title, organisation, location, required skills, description), and an optional preferred location.
• Outputs: structured evidence (skills, years, education), a seniority band, recommended placements with a 0–100 score, plain-language rationale, matched skills, gap skills, and suggested next actions.
• Never used: gender, race, religion, ethnicity, age, disability, marital status, nationality, or any proxy for these.

1. The CV is converted into a numerical embedding (Google text-embedding-004, 768 dimensions) and compared against embeddings of every open, approved role using cosine similarity in Postgres pgvector.
2. The top ~20 retrieved roles go to a large language model (Gemini 2.5 Flash, with Flash-Lite as automatic fallback) which scores and explains each match.
3. The score is clamped 0–100. Only matches scoring ≥ 40 are shown. The candidate sees the same matched-and-gap skill view a recruiter would.

• Every recruiter decision is recorded in the audit log against a verified workspace identity. A named human always makes the hire decision.
• Candidates can request a Subject Access Report and full erasure from the Rights Portal. Consent is captured per application.
• The system prompt forbids the model from using or inferring protected characteristics. This is an instruction, not a guarantee — see Limits.
• The model is called through the Lovable AI Gateway; a fallback provider takes over automatically on outage, and short-lived caches reduce duplicate calls.

• Large language model outputs can vary run-to-run. Embedding retrieval pins the shortlist, but rationale wording will not be identical.
• The engine cannot read scanned image-only PDFs without OCR; the candidate is told and asked to paste text.
• Bias can leak through proxies (school names, employment gaps, location). ADVIS runs periodic adverse-impact reviews; methodology is being formalised.
• Multi-language CVs are supported but not yet evaluated for parity of quality.

ADVIS treats recruitment AI as high-risk in line with the EU AI Act (effective 2 August 2026). This card, the audit log, the rights portal and the published bias review together form the documentation an EU Article 13 transparency request would expect. A Fundamental Rights Impact Assessment and DPIA are maintained internally and available on request to enterprise customers.

• v1.0 — 14 June 2026: First public model card. Vector retrieval (pgvector) added. Fallback model wired. CV extraction now cached by file hash.