Availability: Remote & Hybrid
From MCP and A2A protocol design to production text-to-SQL, vehicle-diagnostics, and fraud-detection systems, pairing experimentation with governance so analysts, technicians, and legal teams trust every output.
Questions are logged anonymously to improve the answers.
0+
Full-stack AI initiatives across finance, energy, mobility, and legaltech teams.
0+
Analysts, technicians, and compliance leads now shipping insights without waiting on IT queues.
0+
Medium deep-dives on operationalising AI systems with 10k+ cumulative reads.
Île-de-France, France
Calendly →“I managed Jaafar for several years and greatly appreciated the quality of his work across various projects, his ability to share his knowledge with colleagues, and his strong listening skills and willingness to find common ground.”
Hedi Aloui · Partner, in charge of Data and AI activities in France
Read all recommendations →
Each partnership blends data science, AI integration, and production-grade craftsmanship. I help teams move from proof of concept to trusted production systems with responsible AI practices embedded throughout.
Designing repeatable pipelines with observability, automated evaluations, and compliance built in from day zero.
Embedding retrieval-augmented generation, semantic parsers, and copilots that accelerate expert workflows without sacrificing trust.
Designing MCP and A2A protocol patterns that let specialised agents coordinate safely, with tool permissions, audit trails, and the freedom to swap LLM backends.
Testimonials from colleagues, managers, and clients who have worked with me on AI projects across various industries.
“I managed Jaafar for several years and greatly appreciated the quality of his work across various projects, his ability to share his knowledge with colleagues, and his strong listening skills and willingness to find common ground.”
“I had the pleasure of working with Jafar on several architecture topics around generative AI, notably design patterns, security, authentication and MCP. I greatly appreciated the quality of our exchanges, his technical expertise and his pragmatic approach to the topics. Jafar is a reliable and committed professional whom I gladly recommend.”
“I had the chance to work with Jaafar recently on a project testing Denodo's MCP and security capabilities, and was impressed by how quickly he was fully operational on the Denodo platform and built complete, well-organized and well-documented implementations. I appreciated both his technical and his communication skills.”
“I had the pleasure of working with Jaafar on two key AI projects at Stellantis Financial Services: text-to-SQL automation and IoT anomaly detection using LLMs. He consistently brought deep technical expertise, especially in NLP and large language models, paired with a strong understanding of business needs. Jaafar stands out for his ability to simplify complex ideas, build scalable solutions, and collaborate effectively across teams. I highly recommend him for any AI or data science role. He's a rare blend of technical depth and practical insight.”
“I had the pleasure of working with Jaafar on an interactive POC exploring a RAG-based architecture for SQL generation from natural language (TXT2SQL). Jaafar demonstrated great autonomy, strong technical insight, and quickly adapted to the complex and highly technical banking environment. His communication skills, the quality of the testing, results, and documentation were all highly appreciated. A valuable contributor I would gladly recommend.”
“I had the opportunity to work with Jaafar at GRTgaz, where he contributed to several projects. Jaafar is a serious and involved person, with a real will to learn and progress. He integrates well into a team and has shown that he can be an asset in a well-structured and organized environment.”
“I worked with Mr. BENABDERRAZAK on a POC for fraud prediction at TER for all trains. He was proactive, persevering, driving in choosing external sources, and very professional. He provided us with very methodical documentation describing his hypotheses and choices, as well as numerous explanatory diagrams. We discussed various algorithms, errors, sources, methods, and I appreciated challenging each other to achieve a better result. The cherry on top is his relationship with others, his good communication, and very good exchanges with clients, teams, and experts. In summary, Mr. BENABDERRAZAK is a good professional in his field and a good communicator whom I highly recommend.”
“I worked with Jaafar for a year. His expertise in data science and AI engineering is remarkable. He is professional, passionate about his work, and has in-depth knowledge of emerging technologies. He knows how to adopt an innovative approach to solve complex problems. He is genuinely interested in this field and therefore always knows the latest innovations, which he sometimes explores through personal projects.”
Four disciplines and the connections between them. Hover a skill to read every mission where it shipped, or filter by discipline to isolate one cluster.
Tap a skill to read where it shipped.
Agent protocols, retrieval, and the LLM systems built on them — from MCP design patterns to a Text-to-SQL copilot in production.
The AWS and CI/CD platform layer: managed training, serverless and foundation-model inference, infrastructure as code, versioned pipelines.
Predictive modelling, deep learning, and NLP — fraud scoring, time-series forecasting, document understanding. The oldest craft here.
Distributed pipelines, graph queries, and the search indexes every retrieval system upstream of here actually runs on.
Joining Google as a Senior Forward Deployed Engineer, embedded with customers to take generative and agentic AI systems from first prototype to production.
Start
5 Oct 2026
Org
Role
Forward Deployed
Crédit Agricole Technologies and Services (Reply)
Led architectural initiatives for next-generation AI integration protocols and multi-agent coordination patterns within enterprise banking infrastructure.
Focus
MCP + A2A
Env.
Enterprise banking
Scope
Architecture
Stellantis Financial Services (Reply)
Architected a production text-to-SQL assistant empowering 40+ financial analysts to interrogate complex datasets without SQL expertise.
Query accuracy
85.6%
Analysts empowered
40+
Decision cycle
−27%
Stellantis R&D (Reply)
Brought LLM reasoning to vehicle telemetry, diagnosing anomalies for connected cars within the R&D division.
Anomaly accuracy
72.8%
Diagnostic time
−24.5%
Warranty savings
€1.1M/yr
SNCF (Talan)
Partnered with the TER division to curb multimillion-euro ticket fraud through predictive controls.
Fraud rate
7.2 → 5.9%
AUC-ROC
>0.87
Volume RMSE
14.7%
GRTgaz (Talan)
Advanced operational safety initiatives for France's gas transport network within a 6-person innovation program.
Support time
−38.5%
Program
R&D innovation
Practice
TDD adopted
LexisNexis (Devoteam)
Scaled AI products for thousands of legal professionals within a dedicated AI Operations squad.
Deployment cycles
−41%
Squad
MLOps · 3
Scope
Global scale
Generali (CGI)
Strengthened fraud prevention and AML/CTF compliance inside the Risk & Compliance department.
Manual workload
−24%
Scale
Millions of txns
Domain
AML/CTF
BPI France · AG2R La Mondiale (CGI)
Automated document analysis for innovation labs by combining OCR and NLP workflows.
Processing time
−34%
F1-score
+11.8%
Focus
OCR + NLP
DGFiP (CGI)
Supported modernization of citizen-facing digital services with NLP automation.
Ministry
Finance
Scope
Citizen services
Focus
NLP automation
Each card shows a 30-second proof: the real before/after number underneath the headline.
Tier-1 European bank
A governed natural-language query layer for finance directors. Every answer cites the underlying SQL and schema.
Validated query accuracy
Connected mobility operator
Live LLM reasoning over telemetry. Agents escalate with a compact root-cause brief.
Time-to-diagnosis
National rail operator
Hybrid gradient + graph model scoring anomalies, built for a court-admissible audit trail.
Ticketing fraud rate
Open-source · reproducible
150-run controlled study comparing 8 agent harnesses on a frozen open-source LLM. Default ReAct lost to a 15-line single-shot baseline. Source + data are reproducible end-to-end.
Best harness vs. default ReAct (success/15)
Professional certifications demonstrating expertise in machine learning, cloud architecture, and AI technologies from industry-leading organizations.
Amazon Web Services (AWS)
Hugging Face
Credential ID: jbenabde
Amazon Web Services (AWS)
Amazon Web Services (AWS)
Amazon Web Services (AWS)
Dataiku
Credential ID: xeizska962nr
National Research University, Higher School of Economics
TensorFlow Certificate Program
Whether you're validating an MVP, hardening an LLM copilot for production, or modernising analytics infrastructure, I bring a blend of experimentation, governance, and delivery discipline.
I read every message myself — tell me what you're building.
Reserve a slot that fits your schedule. The invite will include call details instantly.