Recruiting AI touches some of the most sensitive personal data your business will ever process: full CVs, salary expectations, interview recordings, protected characteristics. Sending that to a SaaS endpoint usually ends with a conversation with your DPO. A self-hosted stack on the RTX 5060 Ti 16GB at our UK dedicated GPU hosting keeps every CV and every interview transcript inside your GDPR perimeter while delivering the throughput to process thousands of candidates per day on one card.
Contents
- Recruiting tasks the card covers
- Stack on one 5060 Ti
- Processing times
- ATS integrations
- Fairness, bias and audit
Tasks
- CV parsing into structured JSON (name, experience, education, skills, locations)
- Candidate-job semantic matching via embeddings
- Shortlist ranking with reasoned justification
- Interview transcription and summarisation
- Interview question generation tailored to the role and candidate
- Personalised outreach email drafts
- Duplicate candidate detection across sources
Stack
| Component | Tool | Role |
|---|---|---|
| OCR | PaddleOCR or Azure Form Recognizer (self-host) | PDF and image CVs |
| Structured extraction | Llama 3.1 8B FP8 with guided JSON | CV -> schema |
| Embedder | BGE-M3 | Candidate and job vectors |
| Vector store | Qdrant | Semantic match index |
| Transcription | Whisper Turbo | Interview recordings |
| Diarisation | pyannote.audio | Speaker-labelled transcripts |
| Summary | Llama 3.1 8B FP8 | Interview summary + flags |
All GPU components coexist on one 16 GB card with headroom.
Processing times
| Task | Per-item time | Per hour on one 5060 Ti |
|---|---|---|
| CV parse (2-page PDF) | ~2.5 s | ~1,400 CVs |
| Candidate-job match (1 job vs 10k candidates) | ~0.5 s | Thousands of matches |
| Shortlist justification (5 candidates) | ~3 s | ~1,200 shortlists |
| 1-hour interview transcription | ~60 s | ~60 hours of audio |
| Interview summary + flags | ~10 s | ~360 summaries |
| Outreach email draft | ~1.5 s | ~2,400 emails |
ATS integrations
- Greenhouse: Harvest API for candidate pull, Ingestion API for score write-back
- Lever: REST API, webhook on stage change
- Workday Recruiting: SOAP Get_Candidate + REST wrappers
- BambooHR: REST API for applications
- Teamtailor, Recruitee, Ashby: webhook + REST
- Generic pattern: subscribe to “candidate created” webhook, pull PDF, run pipeline, write parsed fields and score back as custom fields
Fairness, bias and audit
AI in recruiting is explicitly in scope for the EU AI Act (high-risk) and the UK ICO’s employment guidance. Keep decisions human-owned: the model produces summaries, shortlists and question suggestions, never hire/reject verdicts. Strip names, photos and demographic fields from prompts when not needed. Log every model input, output and human override for audit. Test for disparate impact on protected groups quarterly. A UK-hosted self-hosted deployment makes the “we don’t transfer candidate data outside the UK or EU” conversation straightforward.
GDPR-friendly recruiting AI
CV screening and interview summaries on Blackwell 16GB. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: classification, Whisper benchmark, document Q&A, internal tooling, embedding throughput.
