A model that reads CVs like a recruiter
A domain-specific model that extracts roles, skills, seniority and availability from CVs into structured, searchable data, served on private European infrastructure.
- Client
- Recruitment technology company
- Sector
- Recruitment
- Type
- Custom model
- Services
- Dataset design, Fine-tuning, Evaluation pipeline, Private inference
Outcomes
Challenge
General models read CVs well enough for a demo and badly enough for production: inconsistent fields, invented dates, and every document sent to a third party. The company needed structured output it could search and trust, without CVs leaving its own infrastructure.
Approach
We started with the evaluation harness: a set of CVs labelled by recruiters, scored on every field. Then an open-weight model was fine-tuned on anonymised, labelled documents to produce a strict schema, and compared against the baseline on that set after every change.
The model is served with vLLM on European infrastructure, exposed through an API, and feeds the matching system directly. A demo environment lets the sales team show it on real documents.
Outcome
CVs become structured data the moment they arrive, inside the company's boundary, with a measured quality bar instead of an impression.