AI Engineer | Master’s in AI & Digital Computing | AI Systems & Optimisation Theory Enthusiast | AI Solutions & Full-Stack Integration
01 — About
AI Engineer and Master’s student in Artificial Intelligence & Digital Computing at FST Beni Mellal. I build document AI pipelines, quantization research, and full-stack systems that connect models to real products — from OCR and layout detection to RAG chatbots and interactive demos.
I was always into math and computer science and I couldn’t leave one for another — but I was lucky to find out AI is the perfect mix between both, so here we are :)
Master’s — Artificial Intelligence & Digital Computing
Faculté des Sciences et Techniques de Beni Mellal
Oct 2024 – July 2026
Bachelor’s — Computer Software Engineering
FST Errachidia, Morocco
2021 – 2024
President — Code Crafters
FST Beni Mellal CS/AI student club · Jan 2024 – 2025
Member — Great Debaters
FST Errachidia, Morocco · 2022 – 2024
02 — Projects
Filter by Deep Learning, Machine Learning, NLP, xAI, Tools, or Fun.
Compress long industrial maintenance histories into episodic memory, then reconstruct what matters for failure understanding. On NASA C-MAPSS FD001, ~20% memory keeps event recall near full history with stronger temporal structure than retrieval-only.
Uncertainty-aware active perception so a VLM robot can observe again instead of making unsafe crop interventions — scores information gain and chooses PICK / OBSERVE / ABSTAIN.
Does JEPA-style latent prediction encode naive physics better than pixel reconstruction? Trains JEPA vs pixel baselines on procedural 2D pymunk worlds with linear probes and violation-of-expectation surprise curves.
Language-conditioned visuomotor policy with BC → RL fine-tuning, active perception, and latent world-model planning on LIBERO. Dashboard exports 5 LIBERO Spatial rollouts — research in progress.
Decision-support dermoscopy classifier on HAM10000 with measured benchmarks, Grad-CAM++ interpretability, and an interactive referral-threshold showcase. EfficientNet-B0 matches ResNet-50 at 6× lower CPU latency; melanoma AUC 0.888.
Hamiltonian NN vs MLP vs Neural ODE on a 1D bouncing ball where contact breaks smooth-Hamiltonian assumptions. Continuous-time structure beats the MLP by ~300× rollout MSE; HNN wins only in the elastic case.
Large suite of RL-MPQ checkpoints across LLMs and VLMs — Llama, Mistral, Gemma, Qwen, LLaVA, and more — plus VLMEvalKit evaluation artifacts.
Frequency-decomposed Chebyshev KAN for long-term forecasting — RevIN, per-band M-KAN, cross-band attention; benchmarks vs PatchTST, iTransformer, and more. Part of the KAN Portfolio showcases.
PyTorch custom EML activation with stabilized autograd; benchmarks vs ReLU MLPs on 19 2D classification scenarios.
YOLO detectors for invoice page layout regions with annotated bbox dataset and live Gradio demo.
Finetuned SLANeXt wired table-structure model with 61K+ training samples and Gradio demo.
Simulate, search, and learn policies for classic 2048 — Monte Carlo bakeoffs, expectimax, and Double DQN ablations, plus a playable Angular companion. Heuristic and shallow expectimax reach ~4970 mean score vs ~894 random; short CPU Double DQN demos beat random and approach heuristic with longer training. Charts load from committed metrics via a Render API. Streamlit lab: https://twenty48-lab.onrender.com
Causal ML demo measuring how advanced weather information changes solar forecasting error across regimes. Estimates ATEs with S/T/X-Learner and Causal Forest — headline ATE ≈ −10.3 pp absolute error, strongest under extreme conditions.
Function-approximation arena — fight models, run tournaments, and play Graphwar in the browser. Client-side play works offline; fights and tournaments hit a FastAPI backend. Also published as the Python package graphwar on PyPI.
Provably secure route planning for high-value transport — cargo + decoy routes that preserve Type-B opacity at checkpoints. LangGraph pipeline on OSM/NetworkX graphs with live God/Observer map playback.
DINOv2 embeddings → QUBO max-diversity lunar site selection with greedy, simulated annealing, and QAOA solvers. Greedy scores 8.460 diversity, beating Neal SA and Aer-simulated QAOA on a 12-qubit pool.
Emotion-conditioned agentic RAG that picks a retrieval/generation strategy and explains why. A classifier tags affect; an agent maps emotion → concise / scaffolded / standard strategy with a rationale trace.
Nine RAG architectures built in order — each motivated by a measured failure of the one before it. v8 Agentic hits 0.876; v9 Modular trades accuracy for perfect adversarial resistance; v6 RAPTOR peaks faithfulness at 0.954.
LayoutLMv3 segment classification for Moroccan Official Bulletin PDFs with a live Gradio demo.
PaddleOCR finetunes for Arabic and Latin/French line recognition on Bulletin Officiel documents, with live Gradio demo.
FastAPI + RapidOCR (ONNX) pipeline: upload → OCR/layout/tables → fuzzy structuring → validation UI with bbox overlays.
Interpretable image classification with Kolmogorov–Arnold Networks — Pure KAN and Conv-KAN vs MLP/CNN baselines, plus edge-function visualization and symbolic fits. Part of the KAN Portfolio showcases.
Interactive neural network explorer — React + Express monorepo with Docker/GHCR and Vercel deploy.
Interactive geolocated marker manager with routing, filters, stats dashboard, and JSON import/export.
Browser tool for table-structure bbox annotation with JSON/ZIP export for training.
Bounding-box annotation tool for computer vision datasets.
Vue.js + Babylon.js campus walkthrough with Blender → GLTF models of FST Beni Mellal.
PP-DocLayout region proposal / layout detection for Official Bulletin documents.
Lane detection using Swin Transformers on the TuSimple dataset.
CNN malware image classification on Malimg (bachelor end-of-studies project).
Pygame gridworld Q-learning demo with alpha/gamma/epsilon controls and Q-value overlay.
RAG-powered companion for exploring literature with retrieval-augmented generation.
End-to-end Official Bulletin pipeline combining region proposal, segment classification, and OCR.
3D / cuboid annotation tool for spatial labeling workflows.
Reading tracker / library app deployed on Vercel.
Food / nutrition tracking web app deployed on Vercel.
03 — Experience
DIGEX
Legaliox.com
Code Crafters
Great Debaters
04 — Skills
From Transformers and document AI to full-stack web — evidenced by public repos and Hub models.
05 — Fav Reads
A few titles that shaped how I think about systems, ethics, code, and society.
Understanding Ethics
Torbjörn Tännsjö
Seeing Like a State
James C. Scott
La nature de la nature
Edgar Morin
Clean Code
Robert C. Martin
The Social Contract
Jean-Jacques Rousseau
Theory of Collective Behavior
Neil J. Smelser
06 — Contact
Questions, collabs, or just say hi.
elboubkraoui.farid@gmail.com
Location
Agadir / Souss-Massa, Morocco
