Open to remote AI, ML and full-stack opportunities.
I'm an AI Engineer focused on LLM fine-tuning, RAG and AI agents, building full-stack products from data and models to user interfaces: OMNIA, AtlasKick AI, a Darija LLM and Real or Clone?. I hold a Bachelor in Data Analytics & BI from EST Oujda.
- AI. At LEADZ Tech Services I fine-tuned a conversational LLM for Moroccan Darija, then built a RAG pipeline with OCR for intelligent document processing.
- Machine learning. At CHU Mohammed VI I built a deep-learning model to help detect breast anomalies on mammograms, from image preprocessing to evaluation.
- Full-stack. React 19, Next.js and TypeScript on the front. Django REST Framework or NestJS with Prisma and PostgreSQL behind it. Docker, Playwright, deployed.
- WEBZI1 Studio. My studio designs and ships multilingual (FR / EN / AR) websites for cafés, restaurants and shops in Morocco and Brussels.
I work in Arabic (native), French and English, and I ship interfaces in all three, right-to-left included.
An AI interview practice room. Paste a job post and your CV, answer out loud, and get a score out of 5, what worked, what to fix and a stronger answer built only from your real experience: a number that isn't in your CV gets flagged, not invented. A Live mode suggests answers during calls where notes are allowed. It runs on your own AI key across five providers with automatic fallback, and your key, CV and answers go straight from the browser to the provider, never through my server. The landing page is a scroll-scrubbed film I made from five AI video clips, retimed with optical flow and upscaled locally with Real-ESRGAN.
Next.js 16 React 19 TypeScript GSAP Groq · Gemini · OpenAI · Claude · OpenRouter Vitest
→ Live · Project page · 60-s video
Scammers can clone a voice from 3 seconds of audio to send fake requests for money. I fine-tuned Meta's XLS-R 300M on ~50,000 English and French clips turned into WhatsApp-style voice notes, including 1,619 fresh voice clones I generated on an NVIDIA L40S. It reaches 96.1% accuracy overall and 2.4% EER on unseen cloning engines (ElevenLabs, OpenAI, Gemini). My own tests caught a shortcut in the first model (it flagged 96% of real voices) and I fixed it, bringing false alarms down to ~7%. The app provides a verdict, suspicious seconds and AR/FR/EN safety tips, with Clerk login and private history. Built in one day at the GOMYCODE × NVIDIA hackathon.
Evaluation context: These results are from a held-out English/French test set of 10,122 clips; Arabic is supported in the interface, but detection on Arabic/Darija voices has not yet been validated.
PyTorch XLS-R 300M Chatterbox NVIDIA Brev FastAPI Clerk Groq
→ Code & results · 90-s video
Every match probability breaks down into named, quantified factors: an Elo + Poisson + ML ensemble with SHAP explanations, a 10,000-run Monte Carlo tournament simulator, and a grounded AI analyst. Includes a dedicated Morocco mode and runs entirely in the browser, with no backend.
React 19 TypeScript Vite Tailwind v4 Framer Motion Groq · Llama 3.3
→ Live · Code
An agentic AI concierge that plans and books stays, food and rides in Morocco through plain language. Four full-stack apps share one Clerk login: typed React frontends on NestJS + Prisma APIs over Neon Postgres, with rate limiting, server-to-server auth and a Playwright E2E suite.
|
An end-to-end fitness tracker and social network for athletes. I designed the relational schema, a Django REST Framework API with SimpleJWT, filtering, pagination and OpenAPI docs, and a React 19 app on TanStack Query for workouts, analytics, feed and follows. Runs under Docker Compose.
|
Kami, a specialty coffee bar in Saint-Gilles, Brussels. Bilingual Next.js site with a full menu, reviews, a scroll-through "walk inside" and complete local SEO. |
MiTo Brugmann, an Italian trattoria in Ixelles, Brussels. Bilingual FR/EN rebuild with reservations, a priced menu, motion and structured data. |
More from the lab: private research, data work and smaller builds
| Project | What it is | Stack |
|---|---|---|
| Gold Agent System (private) | A simulation-only, multi-agent research platform for XAUUSD. Data, features, strategies, an LLM macro arbiter, sizing, risk, and backtest / walk-forward / stress testing. The honest result so far: technical-only strategies showed no durable edge on M15, H1 or H4, so it stays simulated while I add fundamental data. | Python · LLMs |
| Instagram Engagement Analytics | An ETL pipeline into a data warehouse, Random Forest classification and KMeans segmentation, feeding KPI dashboards. | Python · scikit-learn · Power BI |
| Mammography anomaly detection | Hospital internship project: deep-learning assistance for breast-anomaly detection, covering preprocessing and model evaluation. | Python · deep learning |
| Sunglasses S2L | An eyewear storefront concept with 3D product visuals. | Vite · Three.js |
| Dynamo Fit | A landing page for a gym in Salé, Morocco. | HTML · CSS · JS |
| AI & data | PyTorch · TensorFlow · LLM fine-tuning · RAG · OCR · deep learning · scikit-learn · pandas · NumPy · ETL · data warehousing · Power BI |
| Backend | FastAPI · Django REST Framework · SimpleJWT · NestJS · Prisma · REST & OpenAPI · PostgreSQL / Neon · Supabase |
| Frontend | React 19 · Next.js App Router · TypeScript · TanStack Query · Zustand · Framer Motion · GSAP · Three.js |
| Shipping | Git & PR workflow · Docker Compose · Playwright E2E · Vercel · SEO & structured data · i18n with RTL |
I'm open to remote AI, ML, data and full-stack opportunities, and to building a website for your business. The fastest way to reach me is LinkedIn or idaoud361@gmail.com.