AI at RheinTalent

1. What AI is used for

RheinTalent uses server-side language models where they provide structured suggestions or explanations:

• CV import: structured profile fields are extracted from uploaded CV files (src/app/api/candidate/cv/route.ts, feature cv).
• LinkedIn import: profile details are taken from a LinkedIn PDF export (src/app/api/candidate/linkedin/route.ts, feature linkedin).
• Company profile from website and documents: the employer assistant suggests fields for company profile, Golden Circle and roles from uploaded or scraped text (src/lib/employer-agent/).
• Purpose statement: a statement is drafted from Ikigai or Golden Circle answers (src/lib/purpose.ts, API src/app/api/candidate/purpose/route.ts).
• Career coaching: analysis, interview preparation and market pulse based on the profile (src/app/api/candidate/coach/, features coach / market-pulse).
• Salary estimate: the career coach returns a salary band; output is aligned with FL/CH reference data (src/lib/coach-prompt.ts, src/lib/salary-data-fl.ts).
• Purpose comparison in matching: two released purpose statements are scored for fit (src/lib/purpose.ts, scorePurposePair).

2. What AI is not used for

• No automatic rejection of candidates or contact requests.
• No rating of people as a suitability or success verdict — match values sort suggestions; they do not decide.
• No surveillance of users beyond logging needed for operation and security.
• No decision without a human: the agent path does not write releases, consent or moderation.

In code: employer assistant suggestions stay pending in the UI until clicked (src/components/employer/AgentIngestPanel.tsx). Columns such as profilePublished, consent and moderation are blocked for the agent (src/lib/employer-agent/persist.ts, tested in persist.test.ts). confirmedAt stays null on the agent path until an explicit human confirmation.

3. Evidence and suggestions

Employer assistant suggestions can carry a text passage (evidence), a source URL (evidenceUrl) and a confidence level (src/lib/employer-agent/types.ts). Evaluation lives in src/lib/employer-agent/evidence.ts:

• Weakly supported suggestions are marked evidenceWeak and shown separately — they need a deliberate click.
• Unsupported suggestions are not applied but remain visible as «unverified» (discardedPatches).
• For the Golden Circle: without a purpose formulation in the source, no patch is created — even a literally supported suggestion is never applied automatically.

4. Models and operation

Configured provider and models are in src/lib/llm.ts. For Mistral the default endpoint is https://api.eu.mistral.ai. There is a quality and a fast model (MISTRAL_MODEL_QUALITY / MISTRAL_MODEL_FAST via environment variables). API keys are used server-side only.

Website, application and database run on our own infrastructure in Europe (Hetzner EU), as described in the privacy policy.

5. Purpose comparison in matching

scorePurposePair returns a numeric value (0–100) and a short rationale (explanation). The overall score combines purpose and skills (src/lib/matching.ts, weighting 60 % / 40 %). Job suggestions are sorted by this score (src/app/api/candidate/job-matches/route.ts) — there is no automatic selection or rejection.

6. Anonymity

Candidate profiles stay anonymous to employers until the candidate releases contact themselves. What data is visible when is set out in the privacy policy — we do not repeat it here.

7. Objection and correction

Questions or concerns about AI use: admin@rheintalent.li. Every AI-assisted entry in a profile can be corrected, rejected or overwritten by hand before it becomes visible to others.

8. Regulatory context

The EU AI Act can serve as a framework for how AI systems are handled. RheinTalent does not claim on this page any audit or specific classification under that law.