Skip to content
AwakastLet's talk

Confidential HRTech client, blind CV generator

Anonymizing a CV now takes under 2 minutes instead of 25

We built a private pipeline that reads CVs in any format, removes protected personal details and keeps the layout, before candidates are scored. It runs on EU-hosted infrastructure.

Growing a product

less time spent on manual anonymization
93%
per CV, end to end
<2 min
European languages supported
15+
protected fields removed
18

Where the client started

Before any CV could be scored, someone on the client's team had to anonymize it by hand, editing out the same protected details on every profile. CVs arrived in mixed formats and languages, some scanned and some as text. The new step had to read all of them reliably, find personal details accurately, keep the original formatting and record what was removed from each document. Personal data could not leave EU-controlled infrastructure or pass through external AI services.

What we built

We built an anonymization module that plugs into the client's existing HR platform. OCR reads scanned and text-based CVs. A privately hosted model for named entity recognition (NER) finds the protected details, and the redaction step removes them while keeping the layout intact. For every document, the pipeline produces the redacted CV and a structured log of what was removed. Redaction and logging run as separate stages, so each one can be checked on its own. It all runs in containers on EU-hosted infrastructure.

Deliverables

  • OCR intake for scanned and text-based CVs

  • Private NER model for protected personal details

  • Structured JSON redaction log for every document

  • Containerized module integrated into the existing HR platform

How it works

  1. Read any CV.

    OCR turns scans and text files into text the pipeline can work with, in 15+ European languages.

  2. Find the protected details.

    A privately hosted NER model spots the 18 protected fields, with no external API involved.

  3. Remove them, keep the layout.

    The CV comes out redacted but formatted as before, ready for scoring.

  4. Log every change.

    A machine-readable log records what was removed from each document.

Results

less processing time
93%
end to end per CV
<2 min
European languages supported
15+
protected fields redacted
18
  • Per candidate CV25 before, 2 after

  • Far less manual work for recruiters

  • The same anonymization rules across languages and formats

  • A record of what was removed, for every processed CV

The client's name and some internal policy thresholds are not shown.

8 weeks to production

  1. Week 1

    Fields and rules

    Agreed which fields are protected and how they are redacted.

  2. Weeks 2–5

    Pipeline

    Built OCR, NER and redaction that keeps the formatting.

  3. Weeks 6–7

    Integration and testing

    Connected the module to the client's workflow and checked accuracy across languages.

  4. Week 8

    Handover

    Stabilized operations and delivered the runbook.

Technology

  • Python
  • Mistral (privately hosted)
  • OCR
  • FastAPI
  • Docker
  • PostgreSQL
  • Kubernetes

Show us the manual step your team repeats on every document.

Tell us what you're building