Portfolio

Context. Flow Line, historically an ERP integrator now repositioning as an IT services company, has launched an AI platform that plugs into its clients’ ERP (reference connector: Sage X3, architecture designed to be ERP-agnostic) and extracts their supply chain data. On that foundation it delivers several decision services: sales forecasting, production scheduling and planning, anomaly detection and client prospecting, and replenishment optimisation up to raising purchase order proposals. The core relies on classical machine learning that is explainable, frugal and retrainable per client, and every capability is designed to be driven by AI agents. It is the group’s flagship product.

My role. Platform owner: I hold its architecture, and every technical and product decision goes through me. Within a four-person AI team, I industrialise the data science approaches validated upstream, turning them into reliable, secure, production-grade services.

Key deliverables.

• Architecture: a modular core per domain (forecasting, scheduling, anomalies, replenishment) on top of a shared ERP ingestion and normalisation layer, with a multi-tenant design and strict isolation of each client’s data

• Industrialisation: orchestration with Prefect (ingestion, retraining, scheduled computation), model versioning and tracking with MLflow, artifacts and datasets on RustFS, testing, continuous integration, monitoring and failure recovery

• Agentic layer (ADK + MCP servers built with FastMCP): each business capability exposed as a documented, agent-callable tool, with guardrails, human validation before any side-effecting action, and full traceability

• Security and compliance: access control, secret management, isolation of client environments, auditable processing

• Technical scoping of new use cases with the data science team, code review, and pre-sales support (demos, prospects’ technical and security requirements)

Context. APERAM commissioned Flow Line to build a bespoke planning and optimisation tool for production orders. The challenge: give planners a way to build, compare and adjust optimised production plans without going through a hard-to-use expert tool.

My role. The optimisation engine was built by the data science team. I joined the project to make it agentic: a tool driven in natural language, fully traceable and safe.

Key deliverables.

• Agentic layer (ADK) exposed through a chatbot, letting planners talk to the solver in natural language

• Agent tooling: moving and rescheduling production orders, running the optimisation engine (OR-Tools), comparing scenarios, and summarising what each iteration changed

• End-to-end traceability: logging of every agent action, change history, reversible decisions

• Guardrails: business constraint checks and human validation before a plan is applied

Context. An internal platform for centralising and intelligently orchestrating knowledge, unifying access to documents, databases and enterprise resources through a conversational AI agent that answers in natural language (RAG, text-to-SQL, SQL-to-visualisation).

My role. Designed, developed and deployed the entire platform, except the initial API development, which I migrated from Flask to FastAPI.

Key deliverables.

• Built an MCP (Model Context Protocol) server that dynamically supplies contextual tools to the AI agent based on user queries

• Deployed on Azure Kubernetes Service (AKS) with Terraform, with CD through Flux CD (GitOps)

• CI pipelines in Azure DevOps (build, test, packaging, validation)

• Integrated the AI modules (RAG, text-to-SQL, SQL-to-viz) into a scalable Kubernetes architecture, plus the user interface, conversational context and access rights

Context. Université Jean Moulin Lyon 3 wanted to improve access to information on its Moodle platform, where students and teaching staff regularly need help navigating courses, finding documents or obtaining administrative information.

My role. Designed and deployed an on-premise RAG chatbot on the university’s own infrastructure, guaranteeing the security and confidentiality of internal data.

Key deliverables.

• On-premise RAG chatbot answering in real time from the university’s internal document base

• Moodle integration: automated lookup of teaching and administrative documents, guided navigation through courses

• Structured the document base with the IT and teaching teams and iterated on answer relevance

• Performance tracking and monitoring to guarantee the tool’s reliability and security

Context. SNCF wanted to pass on the know-how of its experienced staff to new joiners. A proof of concept had shown the full chain (interviews, transcription through Azure AI Speech, summarisation through Azure OpenAI, delivery through a RAG chatbot backed by Azure AI Search), but it was not usable as it stood: no tests, no reproducible deployment, no monitoring, and uncontrolled costs.

My role. Industrialised the solution: a complete refactor of the proof of concept into a deployable, maintainable, production-grade application.

Key deliverables.

• Reworked a monolithic prototype into a modular application: capture, transcription, summarisation, indexing and retrieval became isolated, testable, replaceable components

• Production deployment on Azure Container Apps: containerisation, CI/CD, secret and managed-identity handling, separated environments, monitoring

• Cost control: rationalised cognitive-service calls, caching and batch processing

• Strict access partitioning for sensitive data, test coverage and documentation so SNCF’s own teams could take ownership

Context. Crédit Agricole Technologies et Services wanted to speed up the analysis and reporting of internal survey results through a platform that automatically produces PowerPoint reports with charts, combining RAG, automated visualisation and report generation.

My role. Infrastructure optimisation, deployment automation and technical delivery to the client.

Key deliverables.

• Reorganised the AWS ECS clusters and migrated selected functions to AWS Lambda, cutting cost and complexity

• CI/CD pipelines with AWS CodePipeline: fast, reliable, reproducible releases with no manual steps

• Handed over the codebase with documentation and supported the client in integrating it into their internal architecture (environments, security, interconnections)

Context. Run in partnership with Dell, this project delivered a complete on-premise offering for putting AI models into production in regulated or sensitive sectors: optimised hardware infrastructure, a tailored MLOps stack, plus maintenance and governance services.

My role. Contributed to the design of the Sovereign AI platform, in particular its MLOps layer, from architecture choices through to technical implementation.

Key deliverables.

• Full MLOps infrastructure orchestrated with Kubernetes: data ingestion, training, deployment and monitoring of AI models

• ML pipelines with Python and Apache Airflow, continuous model deployment with Argo CD, versioning and tracking with MLflow

• Selected and integrated tooling suited to on-premise use: scalability, data security, maintainability

Context. Scottish Water, Scotland’s main drinking water provider, wanted to move from reactive to proactive, predictive management of its wastewater network using real-time data from 4,000 IoT sensors, reducing pollution incidents and flooding and meeting SEPA’s regulatory requirements.

My role. Owned the MLOps side, within a multidisciplinary team (hydrologists, data scientists, data engineers, a data architect).

Key deliverables.

• End-to-end ML pipelines on Azure Databricks: ingestion, PySpark processing, model training and deployment, monitoring

• Experiment tracking with MLflow

• Continuous integration and delivery with Azure DevOps

Context. Help DevOps teams without cybersecurity expertise automate the detection and remediation of code vulnerabilities. The solution combines SAST/DAST tooling (KICS) with an on-premise LLM using RAG that automatically raises corrective pull requests.

My role. Designed and developed the platform’s technical infrastructure as a whole.

Key deliverables.

• On-premise infrastructure: Docker, Kubernetes, Airflow and Argo CD to orchestrate the analysis and fix-generation workflows

• Ray Serve for distributed LLM calls, RAG with ChromaDB for contextual fixes

• Reinforcement learning loop: prompts are refined automatically based on whether pull requests are accepted

• Monitoring with Prometheus and Grafana, artifact storage with MinIO

• Platform fully configurable through YAML (IaC), agnostic to projects, LLMs and Git repositories

Context. The Aix-Marseille-Provence metropolitan authority commissioned Atos to deliver an AI solution for handling citizen reports submitted through the “Ma Métropole Dans Ma Poche” mobile app. The aim: automate the categorisation of reported incidents (cleanliness, roads, signage) through image analysis, shortening response times for technical services.

My role. Set up the network and security infrastructure for industrialised AI model deployment on two virtual machines.

Key deliverables.

• Reverse proxies with Traefik and Kong: secure traffic management, request routing and controlled exposure of the AI services

• Configured certificates, HTTPS redirection, and authentication and access restriction mechanisms

Context. Design of an MLOps platform for IoT data processing geared towards anomaly detection, whose main innovation is using reinforcement learning to dynamically select training datasets according to usage context and expected performance.

My role. Led the MLOps architecture design, developed the full data processing chain, and contributed to drafting the patent alongside Atos’ expert team.

Key deliverables.

• Full chain from IoT ingestion through to deployment of anomaly detection models

• Reinforcement learning module driving dataset selection in production

• Co-inventor on two patents: US2024412073A1 (United States) and EP4475044B1 (Europe, granted January 2026)

• Winner of the 2023 Atos Inventor Trophy

Context. An internal Atos R&D assignment to build an MLSecOps platform securing the full chain of data processing and AI model deployment, by recording integrity hashes on the Ethereum blockchain through smart contracts.

My role. Developed and integrated the whole platform (smart contracts excepted).

Key deliverables.

• Secured MLOps chains from data ingestion through to model deployment

• Integration with the Ethereum blockchain to verify the integrity of critical steps (datasets, models, artifacts)

• Integrity proof mechanism through hashes published in smart contracts

• Secured on-premise Docker infrastructure and automated verification workflows

Context. Atout Majeur Concept, a software publisher for hospitals, wanted to improve the management of patient portering flows. The objective: model porter assignment so as to minimise patient waiting times and optimise routes within the facility.

My role. Operations research engineer, responsible for designing and implementing the optimisation models.

Key deliverables.

• Mathematical modelling of the assignment problem with time, availability and priority constraints

• Solved as mixed-integer linear programming (MILP) with CPLEX Studio

• Proposed concrete improvement scenarios for the operational management of porters

Context. The emergency department at Castelnaudary Hospital needed a tool to track working time and manage the emergency physicians’ rota. The project involved the emergency department, medical affairs and the IT department.

My role. Sole developer, from writing the requirements through to delivery.

Key deliverables.

• Gathered requirements from the managers of each department involved

• Built a rota management tool in Excel VBA, delivered in 2 weeks

• Trained the head of department to carry on developing the tool

Context. Institut Medicoach wanted to make its database of dietary supplements and medicines available online, with detailed information on each product: composition, therapeutic indications, interactions and contraindications.

My role. Sole web developer, from design to production.

Key deliverables.

• Built a search interface in PHP and JavaScript for querying the database

• Product pages showing composition, properties, uses, compatibilities and incompatibilities

• Integration and deployment on the institute’s infrastructure