Three Modalities of Production AI in Financial Services: Retrieval, Generation, and Detection
Master's thesis, EPFL, 2026. Advised by Prof. Amir Zamir (VILAB). Hosted by Bank Lombard Odier, Oct 2025 – Apr 2026. · DOI → · PDF → · GGUF models →
Production
Deployed three AI systems under FINMA rules, used daily by bankers, fund screeners and tax specialists. Managed their releases, CI/CD, unit tests and Kafka integration.
Retrieval
Sped up indexing of a 52-page prospectus from 17 minutes to 44 seconds at 95% page-citation accuracy in the production RAG system by optimizing embedding inference (ONNX Runtime, llama.cpp, Azure). Measured as 1,012 s to 44.3 s, mean of 10 runs (thesis Table 3.3). Page-citation accuracy is on a 20-question French-language benchmark.
Cut vector search latency from about 40 to under 10 seconds for 300-chunk queries by moving the index from in-memory LlamaIndex storage to Elasticsearch HNSW.
Designed a callback architecture in which Python services post embeddings to a Spring Boot endpoint that writes to Elasticsearch, with idempotent IDs and exponential backoff. This avoided months of database access review.
Shrank the on-premises embedding model 70% at a 0.62 perplexity cost by benchmarking 30 GGUF quantization variants, and published them on Hugging Face.
Generation
Built an on-premises LLM advisory pipeline (Qwen3-30B-A3B) whose buy and sell recommendations passed the bank’s research, CIO and portfolio checks 96% of the time, on par with GPT-5.1 with RAG (93%). Pooled over three test portfolios: 24 of 25 for the pipeline, 28 of 30 for GPT-5.1 with RAG, and 14 of 30 for GPT-5.1 with web search. The thesis abstract quotes only the web-search comparison.
Detection
Cut the false positive rate of unsupervised anomaly detection on 21,679 tax statements from 90% to 5% by filtering out non-cash payments and merging same-day rows. Of 10 escalations, 8 were confirmed as errors.
Learned tax-field relations with ElasticNet at out-of-fold R² of 0.93 to 0.97 in nested 5-fold cross-validation over 4,267 clients (160,461 rows), then calibrated alert thresholds on held-out residuals.