An assistant that answers everything is easy to build. One that answers correctly, cites its sources and never reveals a document to someone who is not allowed to see it is much harder. That second assistant is the one we want.
Context
A company's information is scattered across documents, e-mails, code and meeting notes, in several languages. And not everyone is allowed to see everything. An internal search assistant is only useful if it respects those permissions to the letter and resists attempts to subvert it.
What you will do
- Build hybrid search (BM25 + embeddings) with reranking
- Implement GraphRAG for questions that require combining several documents
- Enforce access control at retrieval time, document by document — never after generation
- Defend the system against prompt injection and data leakage: filtering, content isolation, automated red teaming
- Produce answers in French, Arabic and Darja, with verified citations
- Build the relevance benchmark and the attack suite
Technical challenges
- Security of an LLM-based system: a document can itself carry an attack
- Multilingual, multi-script retrieval, including Darja
- Guaranteeing that a citation really matches the passage it refers to
- Preserving relevance despite permission-based filtering
Deliverables
- A working prototype integrated into the B2H platform, using the single identity for permissions
- An internal relevance benchmark and its results
- An automated attack suite and a red-teaming report
- Technical documentation and a demo
- The final-year report
Success criteria
- Relevance measured on the internal benchmark
- Zero leakage: no document served to an unauthorised user across all tests
- Attack success rate measured and reduced over the course of the project
- Verified citations for every answer
Tech stack
- GraphRAG
- Hybrid search (BM25 + embeddings), reranker
- PostgreSQL + pgvector
- Guardrails
- Python, TypeScript
Who we are looking for
- A final-year engineering or master's student with an AI or security background
- Foundations in NLP and information retrieval
- A strong interest in cybersecurity and an attacker's mindset
- Rigour in evaluation and documentation
- French, Arabic and technical English
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