Alexandre Poupeau

● Context: As part of EDF's generative AI adoption strategy, supporting technical and business teams in identifying use cases, tools, and best practices to improve employee productivity and efficiency.

● Actions:

- Analysis and mapping of generative AI tools and projects deployed or under development.

- Writing a reference guide presenting the uses, benefits, and best practices of generative AI for technical and non-technical users (Developers, Product Owners, business teams).

- Recommendation of technologies and frameworks suitable for different use cases (LangChain, LangGraph, ADK, agentic architectures).

- Development of methodological support for the use of generative AI in software development activities.

- Creation of interview questions to evaluate future developers joining EDF on their mastery of generative AI tools.

● Results:

- Provision of a consolidated vision of EDF's generative AI ecosystem.

- Acculturation of teams to the uses and best practices of generative AI.

- Contribution to structuring future recruitment around AI skills and new AI-assisted development practices.

1️⃣ Technical Audit — Azure Data Pipeline

● Context: A pipeline processing millions of documents (images, PDFs) on Azure had never been audited, with an architecture not scaled for expected growth.

● Action: Comprehensive pipeline profiling study, production of a 50-slide PowerPoint (bottlenecks, estimated performance/cost/quality gains) and proposal of a scalable architecture divided into 5 prioritized phases.

● Result: Clear technical roadmap allowing the client to plan its scaling up with confidence.

2️⃣ Optimization of Critical Points

● Context: The audit revealed several critical points: slow keyword search, inefficient MongoDB calls, outdated OCR, under-optimization of GenAI & LLM models, lack of document similarity.

● Action: Replacement of the search algorithm with Aho-Corasick, MongoDB refactoring, migration to Azure Document Intelligence, rebalancing of generic LLM tasks with specialized models (LID, NER), optimization of LLM usage via LangChain, and development of a Text Embeddings feature for document matching.

● Result:

- ⭐ Keyword search time reduced from 3h45 to 5ms for a 100-page document. Performance gain ×2,000,000 representing ~ $65,000 savings per 100,000 documents. The client can now sign contracts with clients of a scale they considered out of reach, in 2 weeks of development. Furthermore, it can support the search of hundreds of thousands of keywords + support for brands of less than 4 letters + avoids detection of false positives.

- MongoDB post-processing reduced from several minutes to less than a second, with storage time reduced by a factor of 100.

- Faster, cheaper, and more reliable GenAI / LLM pipeline thanks to specialized models.

- Optimization of document matching via Text Embeddings.