📍 Morocco 🇲🇦 · Open to PFE Internships & AI/ML Roles
I build AI systems end-to-end — from model training to production deployment. I specialize in LLM fine-tuning, RAG pipelines, and generative models, and I can ship the full stack around them (FastAPI, Docker, React).
Three things I've built that actually work:
- Fine-tuned Mistral-Nemo 12B on a custom dataset → 91% quality vs Gemini at 7x lower cost
- Built a CVAE + Latent DDPM for medical image synthesis → 22% better FID than CVAE-only
- Deployed a distributed MAS with 20+ agents and online ML → <200ms consensus, survives 30% node failure
Fine-tuned Mistral-Nemo 12B + multi-agent RAG pipeline
Built a synthetic training dataset via knowledge distillation from Gemini 2.5 Pro (1,096 samples, published on Zenodo). Fine-tuned with QLoRA, achieving 0.92 training loss. RAG pipeline uses LangChain + ChromaDB with Tavily web search fallback for low-confidence retrievals. Ships with 3 production-ready n8n workflows (ingestion, retrieval, fact-checking).
| Gemini 2.5 Pro | Fine-Tuned SLM | |
|---|---|---|
| Overall quality | 97.4% | 91.0% |
| NLI faithfulness | 0.778 | 0.789 |
| Cost / 1k docs | $3.49 | $0.47 |
| Win rate vs base model | — | 88% |
PyTorch QLoRA Unsloth LangChain ChromaDB Ollama n8n Supabase Tavily
Two-stage liver MRI synthesis across 5 cancer classes
CVAE learns a class-conditioned latent space (val PSNR 24.49 dB, SSIM 0.742). Latent DDPM then refines samples in that space — skipping pixel-space diffusion cost while improving perceptual quality. β-VAE warmup prevents posterior collapse on the small medical dataset.
| CVAE-only | CVAE+DDPM | |
|---|---|---|
| Avg FID ↓ | 1043.2 | 814.1 (−22%) |
| Avg SSIM ↑ | 0.257 | 0.264 |
| Val PSNR | — | 24.49 dB |
PyTorch Diffusers LPIPS BERTScore Kaggle T4
20+ agent distributed system with online ML and chaos engineering
Each agent runs HalfSpaceTrees (River) for incremental anomaly detection on streaming sensor data. Agents communicate over XMPP with a trust-based reputation protocol — dead nodes are pruned via heartbeat, low-trust nodes auto-isolated. Chaos mode simulates random failures. Streamlit War Room dashboard for live topology monitoring.
- <200ms consensus latency · 30% node loss with full operational integrity
Python River SPADE XMPP Prosody Docker Streamlit
Comparative study on 24,169 MRI images: classical ML (LBP+SVM, XGBoost) vs deep learning (ResNet50, VGG16, U-Net). ResNet50 hit 92.1% accuracy vs 82.3% best classical baseline. U-Net segmentation DSC 0.88 vs 0.41 for XGBoost.
PyTorch TensorFlow Transfer Learning U-Net Computer Vision
EsteQuiz — University exam platform with student and admin panels · React Express MongoDB Zustand ChakraUI
Magic UI — Open source contributor · components, linting system, docs · React Next.js Shadcn TailwindCSS
express-api-initializer — npm CLI tool for scaffolding Express.js apps · TypeScript Node.js Commander
AI/ML · PyTorch TensorFlow HuggingFace QLoRA LangChain ChromaDB Ollama River Scikit-learn Diffusers Unsloth
Generative AI · Mistral Gemini API OpenAI API n8n Tavily Supabase
Deployment · FastAPI Docker Streamlit Vercel XMPP
Full-Stack · Python TypeScript React Next.js Node.js MongoDB PostgreSQL
🎓 Master's in Data Science & Analytics — Cadi Ayyad University (2024–2026)
💼 Front-End Developer — Integral Progress Technology (Apr–Aug 2024) · React · TypeScript · Redux Toolkit
🌍 Arabic (Native) · French (Professional) · English (Professional)