Jérôme Diaz

Code quality & robustness

• ▸ Introduced strong schema validation (@sinclair/TypeBox) and stricter lint rules
• ▸ Improved test coverage and relevance (unit & integration), reduced reliance on mocks
• ▸ Added pre-commit hooks and lint stages to the CI/CD pipeline
• ▸ Integrated Sentry and BetterStack heartbeats to monitor AI models

Architecture & performance

• ▸ Implemented Server-Sent Events (SSE) to eliminate client polling
• ▸ Significantly reduced the number of network calls
• ▸ Refactored React code to limit duplication and improve stability
• ▸ Built a complete admin interface (React / Redux Toolkit) for user management

Organization & knowledge sharing

• ▸ Trained and coached the team on testability and security practices
• ▸ Set up proactive and corrective measures; steered production releases
• ▸ Wrote a script detecting untranslated text; fixed missing translations
• ▸ Regularly supported the AI team on model debugging and stability

Architect phase (2020–2025)

• ▸ Designed AWS architectures for high-traffic APIs (Club Med B2C mobile app)
• ▸ Defined a React Native architecture easing collaboration between data/business and UI teams
• ▸ Wrote technical designs, performed code audits, mentored apprentices
• ▸ Led the “Technical Improvement” initiative: best practices, tooling, starter kits
• ▸ Researched and shipped generative-AI solutions: chatbots, RAG, knowledge bases (LlamaIndex, LangChain, LangGraph)

Lead developer phase (2016–2020)

• ▸ R&D: comparative analysis of text engines, decision engines, search engines and graph generation
• ▸ Maintenance and development of mobile apps (satellite modem configuration)
• ▸ Strong quality focus: Sonar, MyPy, Ruff and Flake8 integrated into workflows

• ▸ Designed and built the Pinceau tool: ontology-driven applications, validity-period management, graph generation via Graphviz/DOT
• ▸ Studied knowledge management frameworks (MGKME, GAMETH); modeled a CMDB
• ▸ Co-authored a paper presented at KMIS 2012 on reducing knowledge loss during vendor transitions

• ▸ Enhancements and fixes for the Planett software suite (HR management & airport planning)
• ▸ Developed PDF/CSV report-generation modules (PHP, MySQL, JavaScript)

Technical architecture

• ▸ Python back end (FastAPI / Pydantic)
• ▸ Task distribution to remote workers via RabbitMQ + Celery
• ▸ Real time: WebSocket between the API and the front end, Pika/RabbitMQ between workers and the API
• ▸ Web interface built on React-Admin
• ▸ MVP stage, demo available

Why this project

The initial goal was to build a task automation and orchestration system, with the underlying desire to revisit knowledge management systems (KMS) — a topic I have been interested in for 15 years. I then realized the system lends itself perfectly to operationalizing AI: feeding models with training data, or exploiting their results, often requires upstream task orchestration. The platform is not meant to do AI itself, but makes it possible to quickly build any business process that needs orchestration, with or without an associated database.

A laboratory for topics documented on Medium

• ▸ Building a RAG with a dynamic knowledge domain, able to extend (learn) its scope through conversation
• ▸ RAG/GenAI based on LlamaIndex for ingestion, with a custom hybrid indexer (two Elasticsearch indexes)
• ▸ Training and evaluating a text-classification model (~150 categories)