Biscuit defect classification using Classical Machine Learning, TinyVGG CNN, and EfficientNet-B0 Transfer Learning.

Ink into Stars: 108 Lectures on Machine Learning | 《墨点成星：机器学习一百零八讲》中英双语讲义与配套代码

Lightweight open-source AI training platform on Kubernetes: submit Python scripts as batch Jobs, queue with priorities and quotas, track curves, register models, run batch and online inference. No CRDs. | 轻量开源 K8s AI 训练平台

Controlled study of a GPT-style language model trained from scratch: epochs, learning rate, depth, attention heads, and LayerNorm / residual / FFN ablations (47 runs, report + Colab notebook).

Alefba Reader: reads Persian letters from a photo (printed, handwritten, coloured, tilted or sideways, marked on a sheet, or several in a row) and gives their romanised spelling. OpenCV + two PyTorch CNNs, Flask web app for laptop and phone with an alphabet chart, dark mode and the Noghte mascot. CIT 401A-IT31S1, TIP Manila.

File LaTeX e appunti universitari: Robotica, Deep Learning, Teoria dell'Informazione e Sistemi di Controllo.

FastAPI + React/TypeScript app serving four ML models: pneumonia X-ray, brain-tumor MRI, Parkinson's voice and stroke risk

DeepKit-ml A lightweight machine learning library built from scratch in Python, implementing core algorithms and optimization techniques.

Watch a language model think: next-word predictions and live attention, visualized as a galaxy — runs 100% in your browser.

End-to-end deep learning pipeline for 7-class dermoscopic skin-lesion classification. Benchmarks Custom CNN, MobileNetV2, and DenseNet121 on HAM10000, achieving 76.6% accuracy with DenseNet121. Includes a Streamlit demo.

nvidia-smi for multiple servers: every NVIDIA GPU and who is using it, in one terminal over SSH. Nothing to install on the servers.

VisionScope is a local-first AI vision desktop app for real-time object detection, tracking, text-based and image-based search, segmentation, and image captioning from live cameras or video files, powered by PyTorch and modern vision models.

Automatic license plate recognition from video or webcam: YOLOv9 plate detection, MobileViT OCR, tracking with confidence voting, and exact + fuzzy watchlist matching. CPU-ready, GPU optional.

Offline Morse code (CW) decoder and encoder for English, Russian and Kazakh: neural network (CTC) + n-gram language models, Windows and Linux

Python toolkit for running pre-quantized Hugging Face Transformers models (LLMs, VLMs, speech recognition) on Mobilint NPUs.

Practical, reproducible training and fine-tuning guides for embeddings, rerankers, LLMs and multimodal models. Learn with Jupyter notebooks, Google Colab and reusable Python modules.

Spatial multimodal world model framework for the physical world (ZDTaichu5.0-9B style): arbitrary-resolution vision + spatial understanding + multimodal fusion + world model + spatial-multimodal data production pipeline, with a zero-dependency deterministic trainable mock backend.

Production-grade deep learning pipeline for multi-class Brain Tumor MRI classification using Transfer Learning (EfficientNetB3, DenseNet201, Xception), Weighted Soft-Voting Ensemble, TTA, and Grad-CAM++ (XAI).

Application mobile Flutter de reconnaissance de fruits par IA embarquée (TensorFlow Lite, inférence on-device hors-ligne), avec authentification Firebase et assistant vocal Google Gemini.

I built an end-to-end deep learning project that classifies brain MRI scans into four categories: glioma, meningioma, pituitary tumor, and no tumor.

Python projects for stock market analysis and prediction using machine learning, deep learning, and NLP.

Shrinking a tiny language model from 32 bits to 2 bits per number, one experiment at a time. Pure Python + NumPy with zero frameworks.

Windows 本機日文漫畫翻譯至繁體中文：Sakura、Magi、OCR、人工審查；Intel 258V 實機配置與完整開發紀錄。

A teaching-scale looped transformer (recurrent-depth) inspired by GPT-6 Astra, built from scratch in PyTorch. Features shared-weight loops, per-token depth policies, test-time loop scaling, and audited runs.

Physics-informed neural network with the fast spectral method to solve the Fokker-Planck-Landau equation. Compared to `opPINN`, this framework does not require surrogate modeling, reducing both total training time and enhancing the accuracy of the trained model.

ListingLens: a computer vision project with notebooks, experiments and results.

Deep learning CRNN-CTC OCR pipeline for Egyptian National ID card digitization (2014–2018)

A 15×15 world, a 9×9 window of fog, 500 steps. Play the exact environment that trains the agents, then see where you rank against a random walker, a scripted forager, trained PPO policies and a planner that sees everything.

From Java to AI: Python AI learning project for Java developers | NumPy -> ML -> Deep Learning -> Transformer

Autonomous Neural Vector Road Network & Topology Graph Extraction from Satellite/Aerial Imagery (SpaceNet 3 Verified)