Публичный роадмап «с нуля до junior ML Engineer» на 12 месяцев: классический ML, DL, внутренности LLM, LLM-инженерия и MLOps. Русский язык.

End-to-end diabetes prediction using a 3-layer Neural Network (TensorFlow/Keras vs. pure vectorized NumPy from scratch).

Deep learning-based predictive maintenance system for Remaining Useful Life (RUL) estimation using NASA C-MAPSS FD001 data and an LSTM model.

Master's thesis (UTFSM): interpretable multi-task CNN-Transformer for extensive air shower reconstruction at the CONDOR observatory. LaTeX source, code, trained model and results.

Graduate textbooks that should exist, but don't. Five volumes (1,818 pages), written by Claude, conceived and edited by Xiaoyu Li.

Advanced Real-Time 3D Rigid-Body Physics Engine & Sandbox in C++17 (OBB SAT, 3D Rotational Dynamics, Coulomb Friction, Constraint Solver, Spatial Hash)

Find the one-character Adam bug, build FlashAttention from the online-softmax recurrence, size a 70B model's KV cache three ways. Tested NumPy companion to the book Math for ML Engineers: From Backprop to RLHF. Free Chapter 5 PDF inside.

FP8 (E5M2 and E4M3, any pairing) GEMM, GEMV and fused AdamW for NVIDIA GPUs without native FP8 (sm_50–sm_87, Maxwell to Ampere): exact products, FP32 accumulation, TE HYBRID backward, C API and PyTorch bindings incl. torch._scaled_mm and Float8Tensor

Unofficial Rust port of Spotify's Basic Pitch: polyphonic audio-to-MIDI transcription with pitch bends. Pure Rust inference with tract or ONNX Runtime, bundled model, CLI and WebAssembly. Parity-tested against the Python implementation, note for note.

InSituVision-AI: Predicting microvascular decoupling and osimertinib PFS from routine H&E whole-slide images (< 5 €) via Attention-MIL

A multi-sport computer vision system for real-time object detection, tracking, and sports analytics using YOLO, OpenCV, and deep learning.

Plant disease classification with a four-layer PyTorch CNN, SQLite image indexing, reproducible training and evaluation, and a Streamlit prediction app.

The portfolio of Abdelqoddouss Ellouizi — Computer Science Engineer exploring AI, deep learning, and software engineering through thoughtful, real-world projects.

Physics-aware person re-identification from Wi-Fi CSI using metric learning, signal-structured augmentation, and controlled ablations.

End-to-end research platform for Parkinson's disease classification from single-channel low-cost EEG (NeuroSky TGAM): binary protocol parser, real-time viewer, GAF + CNN with per-patient evaluation.

An optimized indoor localization model using Denoising Autoencoders and deep Neural Networks to achieve 93% accuracy on WiFi RSSI fingerprints of UJIIndoorLoc dataset.

Baseline RSNA Knee MRI: pipeline phát hiện 12 bất thường khớp gối từ ảnh DICOM và báo cáo X-quang (9 ngôn ngữ); dùng làm mốc so sánh cho các mô hình huấn luyện ở repo khác.

🟩 Bare PCB defect detection with YOLO26s — missing hole, mouse bite, open circuit, short, spur, spurious copper, pin hole. Test mAP50 0.928 / mAP50-95 0.617. Cleaned PKU-Market-PCB + DeepPCB merge, weights + Colab included.

CNN for EuroSAT satellite image classification, focused on checkpointing and reloading model weights with TensorFlow 2 / Keras.

Readable 25.7M-parameter GLM-5.3-Flash-style Hybrid-MoE built from scratch in PyTorch. Features hybrid linear/sparse attention, executable-reward RLOO, and recursive self-improvement on verified programs.

Dental charting on panoramic X-rays (DENTEX): FDI tooth numbering, caries / deep caries / periapical lesion / impacted-tooth findings, a HiResCAM anatomical plausibility check and a blinded dentist-vs-AI study. PyTorch Faster R-CNN, Colab-ready.

Deep Hedging of a European call in a Heston market. TensorFlow implementation of Buehler et al.

High-performance distributed vector database & embedding search engine in Rust, Python, TypeScript, and Go with SIMD HNSW indexing.

U-Net and mIoU implemented from scratch in PyTorch for semantic segmentation on Oxford-IIIT Pet.

On-device Android app that repaints photos as Monet paintings: CycleGAN + stroke-based renderer, running on the phone GPU with LiteRT.

Cats vs dogs image classifier: a CNN trained from scratch vs transfer learning with a pretrained MobileNetV2. 98.7% test accuracy with TensorFlow 2 / Keras.

A TensorFlow/Keras CNN for classifying flower photos into five categories, with training and prediction scripts.

Deep-learning visual inspection station for electrical transformers: four-camera Raspberry Pi rig, ResNet-101, 90.8% accuracy, deployed in industry.

A TypeScript neural-network library on TypeShade: lazy tensor graphs fused into single GPU dispatches on WebGPU, WebGL2, or CPU.

Agent-guided PyTorch training projects with readable terminals, live curves, reproducible recovery and model-specific explanations.