Ahmed M.Private profile: Upwork shows its earnings only to signed-in members, so we check it every few weeks, not daily. MRR here is a rough estimate (Upwork's six-month earnings ÷ months active). Last read Nov 5, 2026.

I build RAG systems, LLM agents and machine learning models that are measured before they ship, not just demoed. PhD in fluid mechanics (Ecole Centrale de Lyon) and 7 years of numerical simulation; I now build Python ML systems end to end, from raw data to evaluated, tested code. A common way AI prototypes fail: retrieval quietly returns the wrong context, the model answers confidently, and nobody has a test set to catch it. My background is numerical validation, so every system I build comes with an evaluation set, measured results, and rules that catch bad outputs before your users do. WHAT I CAN BUILD FOR YOU - RAG / document Q&A over your manuals, specs, reports or papers, with source citations and a scored test set - LLM agents and structured-output pipelines (LangGraph) with validation logic around the model - Audits and fixes for an existing RAG or agent: retrieval quality, hallucinations, latency, cost - Machine learning models for tabular, time-series and sensor data (scikit-learn, PyTorch, TensorFlow) - Physics-informed neural networks and surrogate models that replace or speed up expensive simulations PUBLIC WORK YOU CAN CHECK BEFORE WE TALK 1) CFD Simulation Copilot: a LangGraph agent with RAG (Qdrant, cross-encoder reranking) that turns a plain-English engineering request into validated OpenFOAM configuration files. 20-scenario evaluation suite, deterministic checks on every output, FastAPI + Streamlit, CI on GitHub Actions. I cut end-to-end latency from 310 s to 80 s by bounding structured outputs, compressing prompts and removing an LLM grading call. 2) Physics-informed neural network (PyTorch) that solves the 2D Navier-Stokes equations with no training data: 2.7% L2 error against the Ghia et al. benchmark, cross-checked with my own OpenFOAM solve. Extended to a parametric surrogate over a Reynolds-number range, with checkpoint/resume across GPU platforms and documented negative results. HOW I WORK - A fixed-scope first milestone, so you see working output before committing further - Written progress updates, tests, and a README the next engineer can pick up - I will tell you when a simpler method beats ML - English, French (fluent), Arabic (native). Based in France (CET): full overlap with UK/EU hours and US East Coast mornings New to Upwork, not to technical work. Send me a short description of your data or your current system and I will reply with how I would approach it.