AI & Data Science Engineer · Researcher in Explainable AI, Multi-Agent Systems & Autonomous Perception
I am a final-year AI & Data Science Engineering student at ENSAM Meknès, Université Moulay Ismaïl (Morocco), graduating June 2027. My work sits at the intersection of explainable AI, multi-agent LLM systems, and autonomous perception — with a common thread of making complex, multi-stage AI pipelines transparent and auditable rather than treating them as black boxes.
Alongside research, I have spent the last two years building and shipping production AI systems across fintech, healthcare imaging, and applied machine learning tooling, giving me hands-on experience with the practical failure modes that motivate my research questions.
- Currently completing a vehicle trajectory prediction study (PointNet + BiLSTM with attention) on nuScenes
- Background spans applied ML (computer vision, time-series, NLP) and production AI systems engineering across three internships and one independent AI agent platform
- Open to research collaborations and co-authorship
|
Explainable AI (XAI) Causal and structural explanation methods for opaque, multi-stage AI pipelines. Multi-Agent Systems Coordination, failure modes, and interpretability of hierarchical LLM-based agent architectures. |
Autonomous Systems & Perception Trajectory prediction and spatiotemporal modeling for autonomous vehicles. Deep Learning Architectures combining geometric representations (point clouds) with sequential modeling for structured prediction tasks. |
Hierarchical Explanation Graph (HEG): Causal Explainability for Hierarchical LLM Multi-Agent PipelinesAbstract / Contribution: HEG introduces a causal explainability framework for hierarchical, multi-agent LLM pipelines — systems where multiple orchestrators, planners, and executors interact across layers, making failure attribution notoriously difficult. A dual-orchestrator architecture executes the pipeline normally while a parallel orchestrator perturbs each layer's intermediate state and measures the resulting shift in the final output via cosine divergence, producing an experimentally grounded causal weight per layer. Across 7 tasks and 42 perturbation experiments, the framework detects systematic confabulation — self-explanations that are causally disconnected from the layer's actual influence on the output — most notably in the final evaluator layer, which explains its output with full confidence despite having near-zero measured causal contribution. This work is positioned at the intersection of interpretability research and practical multi-agent system reliability, providing a diagnostic tool for auditing complex agentic AI systems rather than trusting their self-reports. DOI: 10.5281/zenodo.21009192 Affiliation: ENSAM Meknès Status: Published on Zenodo; arXiv submission in preparation |
In progress: Vehicle Trajectory Prediction (PointNet + BiLSTM with Attention, nuScenes)
A trajectory prediction model combining point-cloud feature extraction (PointNet) with a BiLSTM-Attention sequence model, benchmarked on the nuScenes dataset. Builds on prior CARLA-simulation-based autonomous vehicle work. Full research pipeline set up (EDA, baseline benchmarking, model implementation, ablations, writeup). Manuscript in preparation.
|
Causal explainability framework for hierarchical multi-agent LLM pipelines, detecting confabulated self-explanations via controlled perturbation experiments.
|
AI agent SaaS platform designed and built solo (Corarea). Multi-layer planning architecture (Orchestrator → Planner → Mid-Level Planner → Analyser → Executor → Evaluator), 100+ tools across 15 categories, semantic memory via pgvector.
|
|
React + FastAPI back-office platform for corporate action processing and treasury funding allocation. Built an LLM monitoring agent on top of the existing multi-agent orchestration system for financial model drift detection (PPCA-based) and governed alerting.
|
Decentralized collateral/asset management platform for financial institutions. Hyperledger Fabric chaincode for asset state transitions and audit trails, Next.js frontend, Fabric network hosted on Amazon EC2.
|
|
AI-powered chest X-ray analysis system detecting 14 thoracic abnormality classes with YOLOv8, fine-tuned on VinBigData (18,000+ images), with Gemini-generated natural-language explanations of findings. Research/educational tool, not for clinical use.
|
Offline-first desktop research platform unifying scikit-learn, PyTorch, TensorFlow, MediaPipe, and Stable-Baselines3/Gymnasium behind one GUI — classical ML, deep learning, computer vision, and RL, all local, no cloud dependency.
|
|
Pure Python, from-scratch reimplementation of the original CART algorithm (Breiman, Friedman, Olshen & Stone, 1984), with cost-complexity pruning, native categorical feature handling, and full tree inspection.
|
COLFI — London-based fintech · DApp Developer Intern, Summer 2025 Built the Hyperledger Fabric-based collateral management DApp described above.
COLFI — London-based fintech · Second internship Contributed the LLM-based financial model monitoring agent described above, built on the platform's existing multi-agent orchestration system.
OCP MEA — Integrated Mine, MEA washing facility · Internship Proposed a machine-learning-based soft sensor for pulp density estimation, scoped and delivered over a three-week engagement.
- 1st Prize, Innov'am 2024 — Autonomous TGV bogie inspection robot
- 3rd Prize, Innovathon 2024
Languages: English, French, Arabic, German (B2, self-studied)
Expand
- Advanced causal inference methods for AI interpretability
- Geometric deep learning for 3D perception
- Formal approaches to multi-agent system verification
"The purpose of research is not to eliminate uncertainty, but to make it legible enough to act on."