Ayoub ZOUHRI

– Participated in redesigning PiChat into a modular multi-agent architecture using MCP: sandboxed execution, skill routing, on-premise open source LLM deployment for data-sensitive clients. – Built MCP servers for Office automation (Excel, PowerPoint) and a document summarization tool with Map-Reduce chunking, shipped to production. – Built a Graph RAG POC with LightRAG for multi-hop enterprise document retrieval. – Built structured OCR pipelines for complex scanned document extraction.

– Modeled the Hadamard conjecture using Geometric Deep Learning with optimization on the Stiefel manifold. – Implemented MCTS-based combinatorial search in high-dimensional discrete spaces.

– Adapted DRUSvar, a conditional diffusion model, for ultrasound image denoising. – Trained a Swin Transformer segmentation model, achieving +20 mAP over baseline on the reference dataset. – Validated results against standardized industrial metrics (PSNR, SSIM) in a regulated medical research environment.

– Provisioned reproducible training and deployment environments using Terraform (IaC). – Monitored data drift in production with Evidently AI on the Telco Customer Churn dataset. – Integrated SHAP for model explainability, addressing GDPR and AI Act compliance requirements.

– Built a reinforcement learning agent from scratch based on the Gumbel MuZero/AlphaZero framework. – Implemented the Gumbel-Max trick for root node action selection, ensuring policy improvement without replacement. – Optimized sample efficiency compared to standard UCB-based MCTS for solving specific chess endgames.