AI classification of slides at scale for due diligence

Problem

Capitalise on Vendor Due Diligence reports by automatically classifying slides into a business taxonomy, at the scale of 250,000+ slides (around 1,500 missions).

Constraint

Strict GDPR: the client refuses any data transfer outside the EU. LLM call costs to be kept under control over a massive volume.

Approach

• Cost-aware cascade pipeline: regex → embeddings → LLM arbiter called only on ambiguous cases.
• BGE-M3 embeddings run locally and systematic anonymisation (spaCy + GLiNER) before any call to the vision LLM (Mistral Pixtral, EU).
• Hexagonal architecture; parallelisation via a bounded pool with retry/backoff.

Result

• Automatic classification of slides into the business taxonomy, with the LLM arbiter used only on ambiguous cases to keep cost under control.
• End-to-end GDPR compliance: embeddings computed locally and anonymisation before any LLM call — no client data leaves the EU.