Gabin Maxime N.-AI/ML Engineer · Agentic AI
Check rate
Experience
Multi-Agent R&D Pipeline (3 Custom Agents)
Independent Project
Claude Code subagents, MCP, Pydantic V2, pytest, bandit
Designed and shipped 3 specialized agents that hand work down a line: a research agent writes a cited implementation spec, a coding agent builds the modular code and its tests, a review agent ranks findings by severity and applies the fixes. Each handoff is a structured document, so no stage depends on another agent's context window.
Connected the research agent to an academic-research MCP server (Semantic Scholar, ArXiv, Hugging Face Hub, citation snowballing) so every reference traces to a tool result rather than the model. Gated commits behind ruff, mypy, pytest and bandit, required human sign-off before installs and commits, and persisted session state on disk so long runs survive a context reset.
Agentic ERP Supply-Chain Copilot
Independent Project
LangGraph, MCP, OR-Tools, CVXPY, CRAG, DSPy, AKS
Split the work between the language model and code: a LangGraph orchestrator sorts each question into 1 of 10 types, the model handles the language, and 7 OR/CVXPY/SciPy solvers run the math behind typed Pydantic contracts. Code with provable guarantees executes the plan the model writes.
Built the safety net: a live 100-question test on the real classifier (90% accuracy gate), a retrieval pipeline (BGE-large, pyvector and BM25) that rewrites its own query when the documents returned are off-target (Recall@5 above 0.80), and a human sign-off on decisions above $10,000.
LLM Alignment Pipeline
Independent Project
PyTorch, Hugging Face, LoRA, QLoRA, DPO, MLflow
Fine-tuned Llama 3 8B on 6,000 FinQA examples via LoRA, raising token accuracy from 49.9% to 74.8%. Labeled 1,800 preference pairs with Zephyr 7B-β (RLAIF), then applied DPO to push reward accuracy to 98.1%. Graded with Prometheus 2, kept separate from Zephyr so no model marked its own work, tracked every run in MLflow, and published the adapter to Hugging Face (G-Maxime-N/llama3-8b-finqa-dpo).
Production RAG Chatbot
Independent Project
FastAPI, Chainlit, FAISS, LangChain, Docker, GitHub Actions
Built an async FastAPI service that streams answers token by token, remembers the last 12 turns of a conversation, and reports its own health. Fixed a thread-safety fault in an asyncio generator and removed a 230 MB cold-start delay behind Docker health checks.
Automated releases through GitHub Actions to GHCR so every deployed image traces back to the commit that produced it, and reached 0.8373 citation precision (ALCE) on questions spanning multiple scientific documents.
AI Research Consultant: LLM Reasoning and Agentic AI Evaluation
Outlier AI and Mercor
- Identify where a language model's step-by-step reasoning goes wrong on hard STEM problems and write the correction, which becomes the training signal (RLHF) that makes the next version of the model more accurate.
- Evaluate AI agents on whether they pick the right tool and stay on track across long tasks (ReAct-style reasoning, function calling, long-context management), and settle the ambiguous cases against a written rubric.
- Design the scoring rubrics and lead peer review to keep prompt engineering consistent across a distributed group, turning hard mathematical content into training data the team can trust.
Doctoral Research and Teaching Assistant
Ludwig Maximilian University of Munich
- Established the conditions under which stochastic gradient descent (SGD), the method behind almost all model training, is guaranteed to converge, showing engineers how aggressively they can train before a run falls apart.
- Delivered an explainable-AI (XAI) project that flags failures early in industrial power plants and gives the reason behind each alert. Reached F1 0.99 with 6 ensemble learners and SHAP, and held fairness across sites (Disparate Impact Ratio 0.95).
- Taught postgraduate optimization, machine learning and data science, and defended a disputed proof step before the Springer Nature editorial board through to acceptance.
Doctoral Research and Teaching Assistant
RWTH Aachen University
- Derived practical learning-rate conditions for gradient descent on deep networks that do not shrink exponentially as the network gets deeper, removing a limit that kept the earlier theory out of practical use.
- Backed the theory with large-scale PyTorch training runs on high-performance computing infrastructure, spreading the workload across nodes to cut the runtime of each experiment.
- Taught continuous optimization and mathematics of data science, and built course materials that made advanced theory concrete for engineering students, within a research group of 9 nationalities.
Machine Learning Intern
Group One Holding Company
- Analyzed telecom fuel-consumption data to pinpoint the root cause of fuel loss and compared 4 machine learning models (Gradient Boosting led at 98% Nash efficiency). Deployed the winning model as a Flask web application with a monitoring dashboard, securing 84,617 liters of fuel.
- Automated log ingestion, cutting reporting time from days to seconds, then presented the findings and the case for rollout to company managers and the operations director.
Industry experience
See where this freelancer has spent most of their professional time.
Experienced in Education, Information Technology, Energy, Telecommunication, and Transportation.
Business area experience
See which departments and functions this freelancer has contributed to most.
Experienced in Research and Development, Information Technology, Operations, Quality Assurance, Business Intelligence, and Logistics.
Summary
I am an AI/ML engineer and applied scientist who builds AI systems. My work sits at the intersection of optimization and production AI: agentic AI, multi-agent systems, RAG, and LLM alignment.
Deep learning theory guarantees convergence under conditions real neural networks don't satisfy. Practitioners tune learning rates by trial and error. My PhD narrowed that gap: convergence established for gradient descent and stochastic gradient descent under conditions that actually hold, with learning rates you can compute before training starts. Springer Nature (2024)
What I build: Agentic AI copilots that support ERP activities while handing the arithmetic to code. My supply-chain copilot routes each business question to 1 of 7 deterministic optimization solvers, so the language model plans the work and never computes the number itself (LangGraph, FastMCP, MCP, OR-Tools, DSPy, Pydantic V2). Any spend above $10,000 pauses for human approval. Deployed on Azure AKS behind a 168-test CI/CD pipeline, with 98 percent of prompt-injection probes rejected under a promptfoo red-team suite (Docker, GitHub Actions, LangSmith).
RAG systems that cite sources you can check. My scientific question-answering platform reaches 0.84 ALCE citation precision on the QASPER benchmark and catches the case where a model confidently references a paper that does not exist (CRAG, ColBERT v2, SPECTER2, pgvector, BM25, FAISS, Neo4j, FastAPI). LLM alignment pipelines where no model marks its own work. I trained Llama-3-8B through a 3-stage pipeline of SFT, RLAIF, and DPO for multi-step financial reasoning over SEC filings. DPO reward accuracy reached 97.7 percent on 1,800 preference pairs, with separate models generating and judging the preferences (LoRA, QLoRA, Hugging Face Transformers and TRL, vLLM, MLflow, Prometheus 2).
Predictive models that hold up in the field. My fuel-prediction system for telecom base stations flagged consumption discrepancies and helped secure 84,617 liters of fuel, running live at NSE 0.986 (Gradient Boosting, anomaly detection, explainable AI with SHAP, Flask, Render). I adapted automotive fuel-prediction methods to power generation plants in 2018, among the first ML applications in that domain at the time.
I taught the mathematical foundations of machine learning, deep learning, data science, and optimization as a teaching assistant at RWTH Aachen University and LMU Munich.
Skills
Programming & Data
- Python (Modular, Pandas, Numpy, Scipy, Scikit-Learn)
- Pytorch
- Sql And Vector Databases (Postgresql, Pgvector, Faiss)
- Fastapi
- Pydantic V2
- Flask
- Rest Apis
- Bash
Agentic Ai, Llms & Rag
- Ai Agents
- Agentic Systems
- Multi-Agent Orchestration
- Chain-Of-Thought
- React
- Long-Context Management
- Tool-Augmented Generation
- Subagent Orchestration
- Mcp Server Integration
- Context Engineering
- Human-In-The-Loop Checkpoints
- Langgraph
- Langchain
- Mcp
- Dspy
- Rag
- Crag
- Prompt Engineering
- Hugging Face (Transformers, Trl, Peft)
- Fine-Tuning (Lora, Qlora)
- Sft
- Rlaif
- Dpo
- Rlhf
Evaluation & Experiment Tracking
- Evaluation Framework Design
- Live Evals
- Failure-Mode Analysis
- Root-Cause Analysis
- Red-Teaming
- Rubric-Based Evaluation
- Preference Labeling
- Quality Gates
- Recall@K
- Alce
- Ragas
- Prometheus 2
- Mlflow
- Hallucination Mitigation
Machine Learning & Research
- Deep Learning
- Statistics
- Non-Convex Optimization
- Convergence Analysis
- Bayesian Optimization
- Experimental Design
- Gradient Boosting
- Random Forest
- Svm
- Anomaly Detection
- Explainable Ai (Shap)
- Predictive Modeling
- Exploratory Data Analysis (Eda)
- Nlp
- Operations Research Solvers
Mlops, Devops & Cloud
- Docker
- Github Actions (Ci/Cd)
- Ghcr
- Kubernetes
- Opentelemetry
- Langsmith
- Git
- Pytest
- Distributed Training
- Hpc
- Aws
- Google Cloud Vertex Ai
- Azure Kubernetes Service (Aks)
- Agile (Scrum)
Languages
Education
LMU Munich and RWTH Aachen University
PhD · Applied Mathematics: Deep Neural Networks Optimization · Germany
African Institute for Mathematical Sciences (AIMS)
MSc · Industrial Mathematics: Machine Learning · Cameroon
University of Yaoundé I
MSc · Applied Mathematics: Dynamical Systems and Modeling · Cameroon
Certifications & licenses
Advanced Agent Coding
Outlier AI
Model Parallelism: Building and Deploying Large Neural Networks
NVIDIA
Generative AI with Large Language Models
DeepLearning.AI and AWS
Machine Learning Engineering for Production (MLOps) Specialization
DeepLearning.AI
Structuring Machine Learning Projects
DeepLearning.AI
Statistics
Experience
Global experience
Expertise
Qualifications
Profile
Frequently asked questions
Have questions? Find more information here.
Gabin Maxime is based in Freising, Germany and can operate in on-site, hybrid, and remote work models.
Gabin Maxime speaks the following languages: French (Native), English (Advanced), German (Intermediate).
Gabin Maxime has at least 8 years of experience. During this time, Gabin Maxime has worked in at least 7 different roles and for 5 different companies. The average length of individual experience is 1 year. Note that Gabin Maxime may not have shared all experience and actually has more experience.
Based on recent experience, Gabin Maxime would be well-suited for roles such as: Multi-Agent R&D Pipeline (3 Custom Agents), Agentic ERP Supply-Chain Copilot, LLM Alignment Pipeline.
Gabin Maxime's most recent position is Multi-Agent R&D Pipeline (3 Custom Agents) at Independent Project.
In recent years, Gabin Maxime has worked for Independent Project, Outlier AI and Mercor, Ludwig Maximilian University of Munich, and RWTH Aachen University.
Gabin Maxime is most experienced in industries like Education, Information Technology, and Energy. Gabin Maxime also has some experience in Telecommunication and Transportation.
Gabin Maxime is most experienced in business areas like Research and Development, Information Technology, and Operations. Gabin Maxime also has some experience in Quality Assurance, Business Intelligence, and Product Development.
Gabin Maxime has recently worked in industries like Education, Information Technology, and Energy.
Gabin Maxime has recently worked in business areas like Research and Development, Information Technology, and Quality Assurance.
Gabin Maxime holds a Doctorate in Applied Mathematics: Deep Neural Networks Optimization from LMU Munich and RWTH Aachen University, a Master in Industrial Mathematics: Machine Learning from African Institute for Mathematical Sciences (AIMS), a Master in Applied Mathematics: Dynamical Systems and Modeling from University of Yaoundé I and a Bachelor in Applied Mathematics from University of Douala.
Gabin Maxime has 5 certificates. Among them, these include: Advanced Agent Coding, Model Parallelism: Building and Deploying Large Neural Networks, and Generative AI with Large Language Models.
Gabin Maxime is immediately available full-time for suitable projects.
Daily rate distribution
The rates shown represent the typical market range for freelancers in this position based on recent contracts on our platform.
Average rates for similar positions
Rates are based on recent contracts and do not include FRATCH margin.
The average daily rate is the mean of all daily rates from recent contracts of comparable freelancers on our platform.
The median daily rate is the middle value of all daily rates — half of comparable freelancers charge less, half charge more. Unlike the average, it is barely affected by outliers.
Calculated based on our freelancers’ daily rates as of 30 Sep 2026. Actual rates may vary depending on seniority level, experience, skill specialization, project complexity, and engagement length.
Similar freelancers
Discover other experts with similar qualifications and experience
Stylianos I.
Thomas H.
Anton K.
Tim R.
Chaima D.
Aziz A.
André G.
Tobias R.
Filipp T.
Mirza K.
Experts recently working on similar projects
Freelancers with hands-on experience in comparable project as a Multi-Agent R&D Pipeline (3 Custom Agents)
Marcel M.
Tomas S.
Patrick W.
Felix S.
Stephan F.
Manuel P.
Serge K.
Anthony M.
Matthias I.
Kai S.
Nearby freelancers
Professionals working in or nearby Freising, Germany
Karlheinz W.
Florian S.
Rallis T.
Christoph H.
Andreas L.
Roland C.
Johannes H.
Marco S.
Marc E.
Peer W.
