IT talent qualification
Nova runs the interview. You read the report.
Nova, our AI recruiter, runs a fifteen-minute technical interview, in writing or over video. You receive a score out of 100 and timestamped extracts of the interview. Your teams step in only when the report is read.
Hosted in France · Every company is validated before any file is opened · Human review on request
| Python | Validated | 00:03:10“… a billing service in FastAPI, 40,000 lines …” |
|---|---|---|
| Spark | To confirm | 00:07:42“… I did not design the pipeline from scratch.” |
| Kubernetes | Not validated | 00:11:08“… I deploy, I don’t administer.” |
Not verified: designing a full pipeline.
The journey
Four steps, from the invitation to the report.
You invite, Nova runs the interview, you read the report. Nothing to prepare on your side: Nova starts from the CV.
You invite your candidates
You send a link to each candidate, or you pick a profile that is already qualified. An invited candidate consents by name to passing the report on to you before starting.
The candidate takes the interview
- Fifteen minutes, in writing or over video, in French or in English
- The background, then two to three targeted technical questions
- A situational question
- Cross-checking against the CV
Nova calculates the score
Four axes, weighted on a universal scale, applied identically to every candidate. Neither the face, nor the emotions, nor the appearance are scored.
You receive the report
A score out of 100, a verdict per skill, extracts timestamped to the minute. You decide whom you meet.
Nova speaks French in this video; English subtitles are provided.
What Nova says in the video
Hello, I am Nova, TalentIA's recruiter. I am an artificial intelligence, and I am telling you up front.
I run the first technical interview with your candidates. You invite them, or you find them among the profiles already qualified. Each one takes the same interview: fifteen minutes, in writing or over video.
I start from the CV, ask technical questions, and check what is claimed. You receive a report: a score out of 100 and timestamped extracts of the interview.
The candidate you invite consented to pass it on to you before starting. On an already qualified profile, you read the score and my comments. The extracts and the name arrive when the person accepts the contact request.
Neither the face, nor the emotions, nor the appearance are scored. Your candidates' data does not leave Europe. The candidate can request a human review. A person handles it, and it is not me.
What you see
The report, depending on how the candidate reaches you.
The full report
Score, verdict per skill, timestamped extracts, comments. The candidate receives it first. They consented by name to passing it on to you before starting.
The report, without the extracts
Before the contact request is accepted, you read the score, the verdicts and the comments, without surname or employers. The extracts and the identity arrive when the person accepts your request.
What does not enter the rating
The face, the emotions, the appearance. A break in the career. A skill that was never asked about stays “not covered” and weighs on nothing.
Your tools
What your company space gives you.
One role, one list of profiles
You describe the position or the assignment. For each role, you get the list of qualified profiles that match. You can open it to candidates; your budget is never shown to them.
One table
Invitations, contact requests and interview dates, in a single follow-up, up to the interview with you.
Nova prepares your candidates
Before the interview with you, Nova prepares the candidate from the brief you wrote about the position.
Your colleagues, your tools
A team code for your colleagues. API keys to receive your invitations and your assessments in your own tools.
Checking claimed skills
The skills you claim are tested.
When an answer is right, Nova asks a harder question. Two verification questions are prepared in advance for every technical question. You read each case in the report, with the extract.
What the report keeps from it
A skill that was never asked about stays “not covered”, never “not validated”: nobody is held to account for a question they were not asked. An “I don't know” does not touch the technical rating and earns a consistency point, capped at two. A false statement costs five, capped at fifteen.
Explainability
The score is explained.
On the technical, consistency and language axes, the report quotes the interview once or twice, word for word and timestamped, with what the extract demonstrates. On the marketplace, until the candidate has accepted the contact request, you read the comment behind the rating, never the extract.
The fourth axis is the exception. Communication is not scored on the candidate's sentences, but on a rewrite of their answers into standard French, stripped of the marks of a French learnt elsewhere: a marker must not confuse “expresses themselves badly” with “is not from here”. Quoting that rewritten text would credit them with words they never spoke. So this axis carries no extract of the rewrite. If the candidate contests the rating, the human review decides.
“The site went down every Monday morning, when the newsletter went out. We found that the mailings and the orders went through the same database: we separated them, and in six months there has not been a single Monday outage.”
Where the data goes
Your candidates' CVs, answers and voice do not leave Europe.
If any processing were to leave the Union, the interview does not start and the service does not deploy. This is not a setting: the code checks it at every start.
The European endpoint is checked at start-up
The application and the voice agent build their own connection to the model, check the address actually used and refuse to start if it is not the European endpoint. The deployment fails if the voice stack is not the European stack, or if Nova's face were to be computed anywhere but on our own machines.
The score is computed in-house
The servers are in France; your roles and your reports stay there too. The Trust Score is computed by our own code, from the ratings returned by the model and the published scale. Models whose terms allow training on the data sent are refused by name.
One exception: identity verification is entrusted to a provider whose processing is declared in the Union, but whose publisher is American. The list of sub-processors and the state of our contracts with them are written in the privacy policy.
What we measure about ourselves
We publish our limits too.
A system that assesses people must itself be assessed. We measure ours and publish the method.
The language bias is measured
The same answer, given in native French and then by a non-native speaker, must receive the same rating. We measure this every month on the model in service, with an alert if the gap widens, and we publish the result with its date.
Human review, and what it can change
The score is calculated with no human involvement. The candidate can request a human review of it: a person issues a reasoned decision, which can invalidate the report, unpublish the profile and reopen an interview. An invalidated report is no longer served.
The method, in full
The measurement protocol, the gaps axis by axis, what has changed since and what we do not yet know: it is all written on How we assess you, freely accessible — in French for now.
Companies
Test us on your own candidates.
Have your own candidates qualified and judge for yourself. We validate each company before any file is opened.
No commitment. Every company is validated before any file is opened.