What happened

Recrewty, a small HR technology company from Montenegro, worked with DigitalSmart to develop a language encoder for Bosnian, Croatian, Montenegrin and Serbian. The project ran under the FFplus programme and used the Leonardo supercomputer. Its practical goal was narrower than the fashionable label “generative AI”: match CVs to job requirements, extract structure from documents and prepare an explainable shortlist for a recruiter.

In the production case published by FFplus, the system processed 1,318 applications and helped reduce them to three interviews. The project participants report an 85% reduction in initial screening workload and 55.7 hours saved. People retained the final decision over candidates.

Those figures are useful as a reference, not as a universal guarantee. They come from a single project described by its participants; no public independent audit of the time calculation or comparison method is available. The more transferable lesson for an executive is that the result did not come from an autonomous “AI recruiter.” It came from a focused retrieval, extraction and ranking pipeline under professional supervision.

Why an encoder fits better than a generative model

A generative model is useful for writing explanations and conducting a dialogue, but first-pass retrieval and document comparison are different tasks. The system must turn a vacancy and CVs into numerical representations, retrieve close documents quickly and order candidates against defined criteria. Encoders and ranking models are normally a better fit for that work.

ModernBERTić-large, released by the project team with open weights under Apache 2.0, has 395 million parameters and accepts sequences of up to 8,192 tokens. Its model card says it was trained on 66 billion tokens across four closely related South Slavic languages, using 64 NVIDIA A100 64 GB accelerators for roughly ten hours of the main training run. Safetensors weights and Transformers support are available.

There is an important caveat for Russian businesses: ModernBERTić is not a ready-made Russian-language model. Its vocabulary and training target BCMS languages and Latin script; the developers recommend transliteration before processing Cyrillic input. Copying this exact model into a Russian environment would therefore make little sense. The reusable part is the architecture: a compact local model for retrieval and extraction, domain validation on the company’s own vacancies, and a human at the decision point.

A practical pipeline

The production design can be split into independent stages.

  • An intake gateway receives CVs from an ATS, corporate email or application form, checks file type and malicious content, and assigns versions and retention periods.
  • OCR reads scans, while a parser extracts roles, employment periods, skills, education and contact details. Each extracted field should retain a reference to the original fragment, not just a value.
  • An encoder creates vectors for the job and the CVs. Retrieval returns a deliberately broad pool, perhaps 30–50 candidates, so a suitable person is not removed too early.
  • A reranker or transparent rules compare mandatory and preferred requirements. Missing information must not automatically be treated as a missing skill.
  • The interface shows the recruiter the ranking, matched requirements, uncertain fields and supporting CV fragments. The recruiter confirms or changes the shortlist, or sends a document for manual review.
  • An audit log records the versions of the vacancy, model and rules, the data source, the result and the human action. This supports both incident analysis and quality re-evaluation after an update.

A generative model can be an optional layer that writes a neutral experience summary or prepares interview questions. It should not make an irreversible rejection. It should also receive no more personal data than required for the specific operation.

Required data and integrations

A pilot does not need the whole HR archive. One high-volume or recurring vacancy, 200–500 anonymised or lawfully processed CVs, and a reference set prepared by two experienced recruiters are enough. The reviewers should independently mark candidates who must not be missed, then adjudicate disagreements together.

The minimum integration set covers the vacancy source, document storage, a retrieval API and the recruiter’s work interface. Corporate identity, role-based access, encryption, backups and scheduled deletion belong in the pilot whenever real personal data is used. Results should be exported to the ATS only after human review, rather than directly from the model ranking.

Historical decisions must not be accepted automatically as perfect ground truth. Old shortlists reflect an earlier labour market, individual recruiter habits and possible bias. The reference set should be built around job requirements and verifiable experience while excluding fields that are unnecessary for a business assessment.

Legal and governance considerations in Russia

CVs contain personal data. Before launch, the company must establish a lawful processing basis, purpose, data scope, retention periods, access roles and deletion process. When collecting data on Russian citizens, it should separately verify the requirements of Federal Law No. 152-FZ concerning the initial use of databases located in Russia. The design should be reviewed by the company’s privacy owner and legal counsel.

Article 16 of 152-FZ is especially relevant. It restricts decisions made solely through automated processing when they create legal consequences or otherwise affect a person’s rights and legitimate interests, except where legislation or written consent permits them. The operator must also be able to explain the decision process and consequences and allow an objection.

The safest starting point is shadow mode: the model rejects nobody and does not change a candidate’s status. It produces a recommendation alongside the existing workflow and the team compares the outputs. After validation, the company can automate card preparation and queue ordering, while the final shortlist and rejection remain actions of an authorised employee.

Evaluation must go beyond average accuracy. Age, gender, location, photograph, university name and career gaps may become direct or indirect ranking signals. Removing those fields does not guarantee that the information cannot leak through wording. The team should compare miss rates for qualified candidates across groups where lawful and feasible, investigate disparities and document acceptable thresholds.

What the economics really mean

FFplus reports the project participants’ estimate that one search fell from roughly €90 using commercial LLM services to €0.05 with the team’s own model. Without a published methodology, request composition, infrastructure pricing and depreciation assumptions, those numbers cannot be transferred into another company’s budget. Training on dozens of A100 GPUs was a research phase, not a required purchase for every employer.

For an SME, the sensible starting point is an existing Russian-language encoder, a local server or protected cloud environment, and a small reranker. The main pilot costs are often not tokens but document cleanup, integration, reference labelling, risk review and ongoing support.

Consider a model calculation. If the initial review of 500 CVs takes two minutes per document, the queue requires about 16.7 hours. Cutting review time by 60% would release around ten hours. The monetary benefit is those hours multiplied by the fully loaded cost of a recruiter’s time, plus the value of filling the role sooner. Integration, operation, quality control and relabelling when job definitions change must be subtracted. This is a pilot assumption, not a forecast.

A four-week pilot

In week one, select one vacancy, freeze the criteria and legal design, and prepare the reference set. In week two, connect OCR, extraction and retrieval in shadow mode. In week three, review false negatives, OCR errors and group differences; tune the threshold so the system expands the recruiter’s attention rather than narrowing it without explanation. In week four, measure the outcome and make a go or no-go decision.

Acceptance criteria should include:

  • recall among candidates the experts consider suitable;
  • precision at the top of the list and the share of manual reorderings;
  • time from application arrival to human review;
  • extraction errors for dates, roles and skills;
  • performance differences across auditable groups;
  • processing cost, latency and service availability;
  • an explanation and supporting source fragment for every material criterion.

The pilot succeeds not when the model guesses the “best” candidate, but when recruiters consistently spend less time, do not lose strong applicants and can explain every system action.

Executive takeaway

The Recrewty case illustrates a mature use of local AI in hiring: automate reading and retrieval, not accountability. Do not begin by training a large proprietary model or promising headcount reduction. Begin with one vacancy, shadow ranking, a reference set and measurement of missed candidates. If quality is demonstrated, data is protected and a human can change the outcome, the pipeline can deliver measurable savings without turning an algorithm into a hiring committee.