Vincent D.

Assisted the BPCE Group in designing and deploying custom RAG (Retrieval-Augmented Generation) architectures at scale to meet the specific needs of various business units (Legal, Compliance, IT). The main challenge was to move from simple "PoC" to industrialized, secure, and continuously evaluated enterprise solutions.

🎯 Achievements & Missions:

Strategic Scoping & Agile Delivery: Led scoping phases with project directors and Business Analysts (BAs) to define use cases, budgets, and time-to-market (PoC ➔ MVP ➔ Prod). Worked in iterations (Sprints) in close collaboration with developers.

Multimodal Ingestion & Data Processing: Created complex ingestion pipelines from heterogeneous sources (SharePoint, Confluence, internal APIs). Used Layout Analyzers to structure documents (PDF, Word, Excel) via polygon extraction, combined with dynamic chunking strategies enriched with metadata.

Advanced Search Engine: Implemented a two-level retrieval architecture for internal chatbots: vector search (Bi-encoders) combined with fine-grained reranking (Cross-encoders). Integrated business filters via NLU (stopwords, synonyms).

Evaluation & Reliability (Anti-hallucination): Ensured model accuracy by deploying the RAGAS framework. Created Golden Datasets to measure document coherence, evaluate generated responses ("LLM-as-a-judge"), and ensure absolute reliability for compliance departments.

Cloud Architecture: Deployed the entire solution on the bank's secure infrastructure (Azure AI).

🛠️ Technical Stack: Azure AI, Python, RAGAS, LLMs, Vector DBs, Bi/Cross-Encoders, Layout Analyzers, NLU, Agile/Scrum Methodology.

Designed and built a serverless multi-agent LLM platform from scratch, orchestrating 9 specialized agents (Claude 4.6) for parallel analysis of market structure, order flow, and derivatives risk.

• Engineered a hierarchical synthesis engine with a strict 12-step financial reasoning framework, deterministic conflict-resolution matrices, and kill-switch guardrails.

• Built the end-to-end GCP pipeline (Cloud Run, Cloud Scheduler) to ingest market data, generate analytics, and autonomously publish institutional-grade PDF research and real-time alerts; implemented a unified memory and portfolio-tracking layer enabling narrative continuity.

• Designed and built a LangGraph multi-tool agent (Claude Haiku 4.5) orchestrating a hybrid RAG pipeline (ChromaDB + BGE cross-encoder reranker) over SEC filings and earnings transcripts, complemented by structured SQL retrieval and real-time news ingestion.

• Engineered systematic source citation with reliability tiers, deterministic token-budget control via round cap, and full execution traceability through LangSmith.