Agentic AI
Multi-agent orchestration with MCP tool boundaries and CRAG retrieval. The agent reads the request; typed contracts and exact solvers do the rest.
PhD in applied mathematics · shipping since 2018
AI Research Scientist & AI/ML Engineer
Math Agents Production
I build agentic AI systems and put them in front of real users: multi-agent orchestration, RAG, and MCP, deployed on Azure behind typed contracts, a test suite, and a human-approval gate. The PhD in applied mathematics is where I learned to prove a system works before it ships.
Consulting on agent engineering, evals, RAG and context engineering, and production deployment. Near Munich, on-site, hybrid, or remote.
What I do
Where I focus
Production ML
Pipelines that survive the handoff: containers, CI/CD, monitoring, and a rollback path. Deployment is part of the build, not a step afterward.
Deep learning systems
Distributed training and model parallelism on HPC clusters. Learning rates chosen from the theory, not from a search that costs GPU hours.
Research and evaluation
Convergence analysis published in Springer Nature, and the evaluation frameworks that tell you whether a model actually improved.
Selected work
Featured projects
All AI engineering projectsA deployed multi-agent system, an alignment pipeline, a retrieval stack, and a forecasting model that operators still run. Every number below opens onto the case study behind it.
Multi-agent · deployed
Agentic ERP supply-chain copilot
A planner asks a supply question in plain English. The system does the arithmetic exactly rather than guessing at it, and stops for a manager on anything above $10,000.
A LangGraph agent routes the question to 1 of 7 exact solvers (OR-Tools, CVXPY, SciPy) or a pgvector CRAG pipeline. Typed Pydantic inputs stand between the model and the solvers, and a human gate pauses any decision above $10,000 until a manager approves it.
- LangGraph
- FastMCP
- Neo4j
- pgvector
- OR-Tools
- Azure AKS
- Bicep
Frontier model alignment · active consulting
Agentic AI and post-training strategies
Making frontier models better at hard science problems, and designing the scoring so the improvement can be measured rather than asserted.
RLVR where the answer is objective, RLHF where it is not, and rubric criteria in between for graded scoring. Plus agent trajectory analysis and the preference datasets that feed SFT and DPO.
- RLHF
- RLVR
- SFT
- DPO
- Chain-of-Thought
- Rubric evaluation
Retrieval research · ongoing
Hybrid RAG for cited answers over research papers
QASPER benchmarkA pipeline that answers questions about research papers with a citation on every sentence, plus an audit of my own evaluation code.
SPECTER2 and BM25 retrieval, ColBERT v2 ranking, and Llama-3.1-8B generation. Table bodies recovered from arXiv LaTeX for 809 of 888 papers.
- ColBERT v2
- SPECTER2
- BM25
- vLLM
- Prometheus 2
- ALCE
Forecasting · deployed
Intelligent fuel consumption and warning system
LiveTelecom sites in Cameroon were losing diesel and nobody could see where. This predicts what each site should burn and flags the clusters and sites whose predicted use stands out. Operators still run it.
A gradient-boosting regressor behind a Flask web app, with a deviation threshold at the mean plus 2 standard deviations, a monitoring dashboard, and exports for audits.
- Python
- Flask
- Scikit-Learn
- Pygal
- Render.com
- NVIDIA
- DeepLearning.AI
- Coursera
- Munich Center for Machine Learning
- Outlier AI
Certified in advanced agent coding, advanced prompt engineering, generative AI with large language models, model parallelism for building and deploying large neural networks, and MLOps for production.
About
A bit about me
I started in Cameroon, building a fuel-forecasting app that base-station technicians ran themselves. Then, in my PhD, I established that gradient descent and SGD converge on deep linear networks, with learning-rate bounds you compute before training instead of tuning by trial and error.
Both halves show up in what I ship now: agentic systems that route decisions to exact solvers, retrieval that cites its sources, and a measured result behind every claim. I take each system into production myself instead of handing it off at the prototype.