Screening applications with an algorithm only makes sense if every score can be explained to the person concerned, and if you can show it does not penalise anyone for the wrong reasons. This is the "Business to Human" principle applied to hiring.
Context
Screening applications is slow and prone to subjectivity. CVs arrive in French, English and Arabic, in widely varying formats. Regulation also requires that any algorithm-assisted decision can be explained. The project builds on the job offers and applications module already running in the B2H back-office.
What you will do
- Extract structured content from multilingual CVs (FR/EN/AR) using a vision-language model
- Map the extracted skills onto a skills graph based on the ESCO ontology
- Train a candidate–job ranking model with contrastive learning
- Explain every score skill by skill (SHAP), in a form a recruiter can actually read
- Audit bias (gender, region, school) and implement measurable corrections
- Offer an anonymised screening mode
- Build the evaluation set and the measurement protocol
Technical challenges
- Little labelled data: you will need to be inventive with weak supervision and evaluation
- Algorithmic fairness: choosing the right metrics and owning their trade-offs
- Protecting personal data at every processing step
- Robustness to poorly structured CVs or CVs mixing several languages
Deliverables
- A prototype integrated into the job offers and applications module of the B2H back-office
- A bias audit report, before and after correction
- An annotated evaluation set and ranking results
- Technical documentation and a demo
- The final-year thesis
Success criteria
- NDCG@10 measured on the evaluation set, compared with a simple baseline
- Fairness gap reduced after correction, quantified per group
- Screening time saved, measured on a realistic scenario
- Every score delivered with a per-skill explanation
Tech stack
- Vision-language models, OCR
- Sentence embeddings, ESCO ontology
- SHAP, Fairlearn
- Python, Next.js, Prisma
Who we are looking for
- A final-year engineering or master's student with a data science or AI background
- Solid foundations in NLP and machine learning
- An interest in AI ethics and data protection
- Able to present quantitative results clearly
- French and technical English; reading Arabic is a real plus
Practical details
- Duration: 4 to 6 months (February – June 2027, indicative)
- Supervised by the B2H engineering team, with a weekly check-in
- Integrated into the internal B2H platform (single identity, event bus, shared data)
- How to apply: CV + cover letter, at most 2 subjects per candidate
- Deadline: 15/12/2026