Beetween
HRTech | ATS
AI-powered CV parsing, job parsing, and matching at tens of millions of CVs scale. with prompt injection défense built in.
Beetween operates one of France's leading Applicant Tracking Systems. With a strong engineering team already in place, they needed an AI partner who could move faster and bring specialized expertise in LLM-powered document processing and matching at massive scale.
Rakam delivers dedicated monthly embedded AI development. building CV parsing, job parsing with nonce-based prompt injection défense, and multi-dimensional CV-job matching designed to operate across Beetween's tens of millions of CVs database.
« At Beetween, we have a very strong engineering team. Yet, we wanted to go faster, to beat the competition and take an AI leadership rôle in the HRTech sector. »
CEO
Beetween
Millions
CVs at scale
Monthly
Embedded development
3
AI services built
Nonce
Prompt injection défense
Business
Business Impact
AI-powered matching across tens of millions of CVs gives Beetween a significant competitive advantage in the French HRTech market. Recruiters can now find the right candidates faster and with higher relevance than manual search or keyword-based systems.
The embedded monthly model means Beetween gets dedicated AI engineering capacity without the overhead of hiring specialized ML engineers. Prompt injection défense ensures the system is secure against adversarial CV manipulation.
Product
What We Built
CV Parser (LLM + OCR)
Hybrid parsing pipeline combining LLM understanding with OCR for scanned documents. Extracts structured candidate profiles from any CV format. PDF, Word, images. with high accuracy across diverse layouts and languages.
Job Parser with Nonce Défense
Structured extraction of job requirements from postings, with a nonce-based prompt injection défense system. Prevents adversarial content in job descriptions from manipulating the LLM's extraction behavior.
CV-Job Matching (Multi-Dimensional)
Multi-dimensional matching engine that scores candidates across skills, experience, location, and cultural fit. Built to operate at Beetween's tens of millions of CVs scale using vector search and intelligent pre-filtering.
Technical
Technical Architecture
The system uses Gemini 2.5 Flash for cost-efficient document parsing and matching, with Qdrant as the vector database for semantic search across millions of CVs. Pytesseract handles OCR for scanned documents.
LangChain orchestrates the parsing and matching pipelines, while MLflow tracks experiments and model performance. The nonce défense system adds a cryptographic vérification layer to prevent prompt injection in parsed content.
# Stack
Gemini 2.5 Flash (inference)
Qdrant (vector DB, massive scale)
Pytesseract (OCR)
LangChain (orchestration)
MLflow (experiment tracking)
Kubernetes OVH (orchestration)
# Security
Nonce-based prompt injection défense
Adversarial CV détection
# Scale
Millions of CVs indexed
Multi-dimensional scoring
Move from AI hype to the real thing
Leave us a message, our team will get back to you within 24 hours.