( Software & ML engineer · MSc Statistics & Data Science )
I build AI systems and production software end to end, writing the code and training the models myself. Then I multiply that with agent systems that keep your software shipping after launch, and I help companies do the same.
( Models & agents I work with )
( 01 · The agent org )
1 teams online
PM project manager · FE frontend · BE backend · QA gates · REV reviewer
01 It starts with one team. A project-manager agent turns my brief into tickets for its frontend and backend engineers.
02 Every team runs the same loop: QA gates, an adversarial reviewer, and findings sent back with proof.
03 Need more capacity? I spin up another team. After launch, the same teams keep maintaining and extending your product, and every one reports to me.
( 02 · The shift )
Agents only make you faster if you can read every line they write. I've spent years writing software and training models by hand, so I know when the agents are wrong, and why. They multiply my engineering. They don't replace it.
( 03 · Under the hood )
Before any agent touches a codebase, I can build it myself. Five years of production code, an MSc in Statistics & Data Science, and models I've trained and deployed on real hardware. That's the difference between engineering with AI and vibe coding.
01 / 04
TypeScript · JavaScript · Python · C/C++ · SQL · R
Next.js · React · Node.js · FastAPI · Django
proof → A sandboxed indicator-script runtime with capability gating, execution contracts and security fixes on a production trading terminal.
( 04 · The workflow )
Every ticket starts from the contract, not a vibe. I scope the work, and agents read the API spec at the exact branch they're working on.
Claude Code and Codex agents each take a ticket in their own git worktree. Several streams move at once and never collide.
Lint, unit tests, production build and end-to-end tests. Every push has to pass every gate, with no exceptions and no "works on my machine".
A reviewer agent attacks every change on five fronts and has to prove each bug with a failing test. Findings get fixed or rejected with evidence until none are left. Then I sign off.
( 05 · The AI toolbelt )
My daily driver. I build custom subagents with written rulebooks: implementers, reviewers and specialists that each own one job.
OpenAI's coding agent, for parallel workstreams so more tickets move at the same time.
Implementers never grade their own work. A separate agent reviews every commit and needs proof for every finding.
Lint, unit tests, builds and Playwright e2e wired into the loop, so agents can't skip them.
The Anthropic API and other LLM APIs inside the apps I ship: content generation, AI reply flows, strategy copilots.
Open-weight LLMs running on your own hardware: private by default, no per-token bill, and tuned to the job.
Multimodal pipelines that generate and edit images and video, locally or through APIs, wired into real products and workflows.
Models designed and trained for your problem, on your data: CNNs, Vision Transformers, sequence and tabular models. I handle everything from data labelling and architecture to training, evaluation and a deployed endpoint.
Open models adapted to your domain with LoRA and QLoRA, plus synthetic data where real data runs short. Every model gets an evaluation set, so you know it's actually better, and a pipeline that retrains it as your data grows.
Retrieval over your own documents with embeddings and vector search, so LLM answers come from your data and cite their sources, not the model's guesses.
Token-efficient by design: prompt caching, routing easy work to small models and hard work to big ones, batching, and local models for private or high-volume jobs. AI spend is measured per feature, so it stays predictable as you scale.
( 06 · What I do )
Full-stack products, AI features and ML systems, built to run under real users, not demos. Next.js and TypeScript, Python, PostgreSQL, and LLMs wired into the product.
Launch isn't the finish line. I set up agent teams (PM, engineers, reviewer, quality gates) that keep maintaining and extending your software: bug fixes, dependency and security updates, new features. Every change is reviewed and tested before it merges, so the product keeps improving without a full dev team on payroll.
For companies that want to automate or adopt agents safely: workflow audits, deciding what to automate first, agentic system design, cutting AI and token costs, and getting your own team productive with AI coding agents.
( Specialties )
( 07 · Selected work )
( 08 · Experience )
Education is the main branch, and every role branches off it. Blue branches are still running.
Mohammed VI Polytechnic University · Ben Guerir
OCP SA · Safi, Morocco
Power BI dashboards for real-time production KPIs, and a predictive-maintenance model forecasting time between failures for a key turbine.
Hasselt University · in progress · graduating January 2027
AutoMates Robotics · Antwerp · freelance
Responsible for the software in the robotics team: the bin-picking vision pipeline and end-to-end ML on 2D and 3D depth data, from labelling and synthetic data to training, inference and deployment.
Syvera · AI trading-bot platform
UX research and a full UX/UI revamp for complex trading tools like risk management and strategy building.
Throughput · development agency · Trading Terminal
One of the engineers on the team building a production trading terminal. My work covers frontend and backend: charting, indicator runtime, broker and trade workflows, execution contracts, security. I ship it through my own multi-agent workflow.
MSc graduation next: January 2027 · Arabic · English (C1) · French
( 10 · Contact )
New products, agent systems that keep your existing software moving, or consulting on automation and AI adoption. Open to freelance, consulting and contract-to-hire.
( Elsewhere )