Three Modalities of Production AI in Financial Services: Retrieval, Generation, and Detection

This thesis is about three AI systems built and put into production at Lombard Odier, a Swiss private bank. Each one solves a different problem with a different kind of model, and all three run under FINMA regulation, used every day by bankers, fund screeners, and tax specialists. One lesson kept coming back: the integration work takes more effort than building the models, and a system matched to its task beats a more general one. The work was done at EPFL’s Visual Intelligence and Learning Laboratory (VILAB), in collaboration with Lombard Odier.

Retrieval

A retrieval-augmented generation (RAG) pipeline that answers questions over the bank’s research documents. Every chunk keeps the page it came from, so each answer cites the exact page. It indexes a 52-page prospectus in 44 seconds instead of about 16 minutes, and points to the right page 95% of the time on a 20-question French benchmark. A sweep over 30 GGUF quantization variants shrinks the embedding model by 70% with almost no quality loss.

Generation

ALBA is an on-premise advisory pipeline. It combines a client’s portfolio with curated summaries of the bank’s macro views and equity research, then drafts structured buy and sell recommendations. A 3B-active-parameter model running locally scores 96% on advisory quality, against 47% for GPT-5.1 with web search. The takeaway: for this kind of structured financial reasoning, good context beats a bigger model.

Detection

A regression model (ElasticNet) learns the normal relationships between the columns of a tax statement, then flags new statements that break those patterns. It needs no labeled examples of errors. Domain-specific preprocessing cut false positives from 90% to 5% while keeping 80% precision on the cases it escalates, and it runs across the bank’s Swiss client base.

Artifacts

• Full thesis (PDF)
• GGUF quantized embedding models on Hugging Face
• PrivilEdge RAG benchmark dataset on Hugging Face