CV screening software that explains every score

The demo shows eleven because eleven fits in a short video. The pipeline runs in batches and has no practical upper limit: five hundred applications is a normal batch, and the per-application cost stays the same. Large batches run in parallel, so throughput scales with how many requests run at once rather than with the size of the pile.

Eye-tracking research by Ladders puts a recruiter's first pass at about 7.4 seconds per CV — that is a glance, not an assessment. A considered review against an actual job spec takes minutes. At three minutes per CV, five hundred applications is roughly twenty-five hours of reading before anyone has a shortlist.

About 1.8 seconds per application on a measured run, and that includes scoring every criterion and writing the evidence for each one. It is not a keyword filter: each application gets a full assessment.

It never produces a score. It marks each criterion 0–5 and quotes the line in the CV it based that on. Weighting and totals are arithmetic in code. If the model returns anything that is not valid structured output, the run retries once and then fails loudly rather than inventing a number.

That is a real attack and it is handled explicitly. The prompt instructs the model to ignore instructions found inside candidate documents and to flag them in the summary, so the attempt shows up in the report instead of changing the score.

For the first conversation, nowhere: a plan can be built from a job advert, which contains no personal data. In production, CV text is sent to the model provider, and that is written into a data processing agreement together with the provider, the retention period and the deletion process. If EU-only processing is a requirement, say so early — it changes the architecture, not the approach.

AI used to screen job applications falls under the employment category in Annex III of the EU AI Act, so it is treated as high-risk. The design reflects that: the system ranks and explains but does not decide, every score carries evidence, decisions are logged, and human oversight is the default rather than an option. Compliance is shared between the provider and the deploying organisation, so your legal counsel should be part of the conversation — this is not legal advice.

Any screening process can be, including a purely human one. Two things reduce it here: the criteria are written down and agreed in advance instead of living in someone's head, and every score is traceable to a line in the CV, so a biased criterion becomes visible and can be removed. Protected characteristics are never criteria.

Not the software. It ranks and explains. Every rejection carries a written reason precisely so a person can overrule it and see what they are overruling.

Usually yes. The pipeline is separate from the ATS, so it can read from an export or an API and write the shortlist back, or simply deliver a daily digest. Tell us which system you use and it will be part of the plan.

The assessment is not tied to one language: applications in English, German and Hungarian are handled, and mixed batches are normal. Before a rollout, a sample is tested in the actual languages your applicants use rather than assumed.

Empty exports, photos renamed to .txt, broken encodings: filtered out before any model call, logged by name, and the run continues and exits with a warning. This is tested with several kinds of malformed input, because a screening run that silently drops a candidate is worse than one that stops.

A single-role pilot is typically live within one to two weeks of the rubric being agreed. Most of that time is the rubric conversation, not the code.

It depends on volume, integrations and whether real candidate data is involved on day one, so a number without those answers would be a guess. Answer the four questions above and our team comes back with a scoped plan within 24 hours.