// About — Mohamed Saadi
Closing the Gap Between
Research and Production.
Most AI initiatives stall not because the research is wrong — but because the gap between theoretical capability and a production-grade, observable, maintainable system is enormous. That gap is where I work.
My approach has never changed: identify which research advances will matter in practice, understand them at the mathematical level, then build the architecture and engineering culture that makes them work under real constraints — delivery timelines, legacy systems, regulated environments, and teams that weren't hired to run AI platforms.
"The rarest skill in AI is not knowing the research — it's knowing which research matters in production, and how to make it work for people who can't afford experiments."
01 / Proven Enterprise Impact
EDF Group AI/ML Platform · AWS + OpenShift · 100+ daily training jobs · 10+ data science teams
From ~6-week manual deployment cycles to 5-day automated pipelines. Achieved through standardized CI/CD, containerized model serving, and self-service tooling for data scientists.
On a multi-million euro AWS bill across a large industrial group. FinOps framework: right-sizing spot instances for training, reserved capacity for inference, automated idle resource cleanup.
Across 100+ daily training jobs and production inference services. Achieved with multi-AZ deployments, automated failover, Prometheus alerting, and zero-downtime deployment patterns.
From median 4h to under 90 minutes. Systematic observability (ELK, Evidently, Grafana), runbook automation, and on-call rotation across 3 timezone-aligned teams.
02 / On Environments
Depth Was the Point. Not the Limitation.
A legitimate question about a career with significant depth in one environment: does this translate elsewhere?
The answer is in the constraint profile. EDF is not one environment — it's three simultaneously: a regulated utility (compliance, auditability, no fast experiments), a large-scale legacy system integrator (Oracle, SAP, unmovable infrastructure), and an internal startup (the data division being built as a new capability inside a 150,000-person organization). Most difficult AI deployment scenarios are a subset of this combination.
For raw greenfield chaos: SeriesMind. Zero budget. No team. No infra. Solo architecture, product, and engineering decisions in parallel with a full-time mission. That context is as different from EDF as environments get.
And Sanofi: a GxP-adjacent pharmaceutical environment with its own validation requirements, documentation standards, and organizational culture. The ML engineering principles transferred. The execution approach had to be rebuilt from scratch.
03 / A Pattern, Repeated Three Times
Arriving Before the Market Has a Name for It.
Research signal: Distributed computing, columnar storage, stream processing
Real-time 360° Customer View at EDF. Constraint: surrounding Oracle systems without replacing them.
Research signal: Reproducible pipelines, feature stores, model governance
Full MLOps stack built from scratch on AWS & OpenShift. Lesson: governance without team autonomy kills adoption.
Research signal: Multi-agent negotiation, neuro-symbolic reasoning, LLM orchestration
LangGraph, RAG, MCP at enterprise scale. The 2016 coordination theory — now stress-tested by production constraints theory never anticipated.
04 / On the 2016 Research
A Framework, Not a Trophy.
The LIP6 thesis on multi-agent negotiation is not mentioned as a credential. It's mentioned because it provides a precise vocabulary for evaluating current tools that most practitioners lack.
When I evaluate a LangGraph orchestration architecture, I'm asking the same questions I asked about JADE agents in 2016: what is the communication protocol? How is partial state handled? What are the failure modes when an agent receives contradictory information? How do you recover a multi-step workflow from a mid-execution fault?
The compute changed. The questions are identical. That theoretical grounding is why I can distinguish a framework that will hold in production from one that demos well.
2016 · École Nationale Polytechnique × LIP6 / CNRS Paris
A Multi-Agent Negotiation Approach for Supply Chain Management
188 pages · JADE · Nash Bargaining · AUML
05 / Experience Timeline
-
2015–16
Research — Multi-Agent Systems
LIP6 / CNRS Paris. Autonomous agents negotiating under incomplete information. Bargaining protocols in JADE. 188-page thesis. The coordination problems I formalized then are what LangGraph solves today.
-
2017
ML Engineering — Pharma
Sanofi. Demand forecasting for pharmaceutical distribution centers. 8% stock reduction. First lesson: ML in a regulated, politically sensitive environment requires as much change management as engineering.
-
2017–19
Wave I — Big Data
EDF. Real-time 360° Customer View on Spark/HBase. Enterprise-scale data engineering before it was standard at French utilities. The constraint: legacy Oracle systems that couldn't be replaced, only surrounded.
-
2019–21
Wave II — Cloud Native & MLOps
AI/ML platform on AWS & OpenShift: 100+ daily training jobs, feature store, CI/CD, model registry — built from scratch when MLOps was still a conference topic. Learned: governance without autonomy kills adoption.
-
2021–Now
Wave III — GenAI & Agentic Architecture
EDF Group (Senior). LangGraph agentic workflows, RAG, MCP, LLM governance across 10+ teams. The 2016 theoretical framework now deployed at enterprise scale — with all the production constraints theory never anticipated.
06 / Architecture Trade-offs
Opinions Grounded in Production.
A tool list proves nothing. What follows are architectural decisions where the answer wasn't obvious — and the reasoning that drove the choice.
LangGraph vs raw orchestration
Explicit state machines win in production.
LangChain's abstraction hides the state machine. In a 4-agent workflow with conditional branching and human-in-the-loop checkpoints, implicit state management becomes a debugging nightmare at scale. We moved to explicit LangGraph graphs — more verbose, every state transition observable, testable, and restartable. Trade: developer ergonomics for operational resilience. In production, you always choose resilience.
RAG vs fine-tuning
Wrong question. Right answer depends on the knowledge lifecycle.
RAG wins when the knowledge base changes frequently and you need audit trails. Fine-tuning wins when output format consistency matters and the domain vocabulary is stable. At EDF, the regulatory and contractual corpus updates quarterly — RAG was the only defensible architecture. Fine-tuning was answering the wrong question.
Feature stores: when they create more problems than they solve
Feature stores earn their place in ~40% of use cases.
Feature stores solve a real problem — training-serving skew — but introduce operational overhead that exceeds the value for most teams. At 100+ daily training jobs, we found the break-even point: feature stores are justified for sub-hour refresh cycles and cross-team feature reuse. For everything else, a versioned feature pipeline with documented contracts is simpler, more debuggable, and cheaper to operate.
MLflow vs custom model registry
Build custom only after you've hit the OSS ceiling.
MLflow's registry is sufficient for most production use cases. We customized exactly three things: upstream metadata schema enforcement, downstream deployment contracts, and rollback trigger logic. The instinct to build bespoke tooling is almost always wrong until you've exhausted the OSS tool at scale. We didn't exhaust MLflow for two years.
Coming
Architecture decision record: AutoCons-Radar RAG pipeline
Full ADR for the LLM-powered eligibility engine at EDF — model selection, chunking strategy, retrieval design, evaluation framework, failure modes. Publishing Q3 2025.
07 / Technical Reference
Tools used in production. Depth varies — ask me to draw the boundary between "comfortable in production" and "evaluated and rejected" for any of these.
Agentic AI & LLMs
MLOps & LLMOps
Cloud Native & Infrastructure
API & Architecture Patterns
Governance & Engineering Leadership
Foundations
08 / What I'm Building
Independent Consulting Practice
Senior Solution Architecture for enterprise AI platforms. Current: EDF DSIN Direction Commerce — GenAI, Agentic AI, MLOps at scale.
SeriesMind — early stage
Multi-Agent Time Series Platform
Six specialized autonomous agents, human-in-the-loop checkpoints, conformal prediction for distribution-free confidence intervals. The proof-of-work on the architecture decisions described above.
09 / Beyond the Screen
The Athlete
Former professional handball player. Today, cycling as commitment: each year I ride the Tour de France routes across Europe, stage by stage. Hyrox races fill the winters.
The Mind
Iyengar Yoga — precision and structural alignment as practice. Sustained interests in neuroscience, cognitive science, and contemplative traditions.
The Listener
Cosmic jazz: Ibrahim Maalouf, Dhafer Youssef, Avishai Cohen. Structured improvisation — the beauty of constraint.
The Aesthetic
Brutalism and neo-brutalist architecture. Structures that expose their raw logic and carry their weight honestly, without ornament.
Available for senior missions
Solution Architect AI · Agentic AI · MLOps · LLMOps · GenAI Platform
Paris region & remote · Minimum 6 months ·
Get in touch