Knowledge tracing on ASSISTments: baselines, IRT, PFA and DKT with leak-free evaluation.

Implemented Transformer Architecture from scratch

AI & machine learning project: a Keras CNN that recognizes hand-drawn camping items (20 categories, 87.86% test accuracy) trained on Google's Quick, Draw! dataset, with a Tkinter drawing app

8-class vehicle image classifier with Swin-Tiny, a Neysan pickup detector, and a confidence threshold for manual review

Ingress: 110M-parameter Conformer Foundation Model for self-supervised representation learning and Habitable Zone exoplanet discovery (NASA TESS & Kepler).

Curso de PyTorch en español, en notebooks: red neuronal simple, descenso de gradiente, SGD, optimalidad y la matemática detrás de cada función

Interactive notes on computing: math, infrastructure, computer science, machine learning and deep learning, with live figures, a note graph and search.

DART: a degradation-aware recurrent transformer for archival film restoration (ACCV 2026). Pretrained weight, inference on your own footage, training and the AbsoluteDegradation evaluation protocol.

GSDM (Graph-Structural Decision Model): a non-transformer, from-scratch NumPy architecture that reasons over code graphs (AST, data flow) and returns calibrated decisions instead of text. Ships GVF, a Python vulnerability finder with XML output.

A reproducible benchmark comparing Physics-Informed Neural Networks (PINNs) against a Strang split-step Fourier reference solver for the 1D Schrödinger/GPE. Honest failure analysis included.

DCX is a memory layer between an AI agent and its decision-makers, a model or a human. When the agent repeats a choice, like retrying a 429 error, DCX checks whether it has decided this before and whether that decision proved itself. If yes, it answers instantly with no model call. If not, it asks a model or a human and saves the answer for reuse.

🧠 A pure NumPy 4-layer Neural Network from scratch to classify MNIST digits. Open Source. Copyright (c) 2026 Muhammad Al-Muzahid.

Fine-tuning ResNet-50 on Google’s Quick, Draw! doodles to beat an IEEE 2025 benchmark in all 4 runs, plus a CNN built from scratch in NumPy and a live Gradio drawing demo.

拍到一道好菜就存下来：识别菜名、给出配比与步骤，库外菜由大模型生成后落库。FastAPI + React + ConvNeXt ONNX。

Identifying 25 identical LoRa IoT transmitters from raw RF I/Q signals with PyTorch, with a leakage-free, unseen-day evaluation

Multimodal visual search · 92.4% Recall@5 · CLIP + FAISS · FastAPI

Next-generation superhuman chess AI engine & research arena. Features Dual-NNUE, spatial policy ViT, playable bots (Stockfish 17, Lc0, AlphaZero, Maia), and a 17-note research vault.

ResNet-style classifier for Fashion-MNIST built with custom residual layers, model subclassing and a custom training loop (tf.GradientTape) in TensorFlow 2. Reaches 82.76% test accuracy training on only 5,000 images.

Deep learning pneumonia detection on chest X-rays with Grad-CAM visual explainability, PyTorch, and an interactive Streamlit diagnostic interface.

An AI-powered Signature Forgery Detection System that uses deep learning to analyze and verify handwritten signatures. The project includes a backend for model inference and a frontend for user interaction, with machine learning models managed using Git LFS.

✨ Restauración y ampliación de fotografías con IA en Google Colab mediante CodeFormer (rostros) e HYPIR (detalles generativos). Incluye mejora opcional del fondo con Real-ESRGAN, salida 2x/4x, verificación SHA-256 de pesos y caché en Google Drive.📸

Deep learning for EV battery temperature forecasting, health (SOH/RUL) prediction, anomaly detection and a physics-guided digital twin, on the NASA PCoE Li-ion dataset. FastAPI + React dashboard.

Command-line client for the gpu.nz GPU cloud: rent RTX 4090, A100, H100 by the minute, launch ComfyUI, Jupyter, Ollama, vLLM, SSH in. Linux, macOS, Windows.

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

Real-time facial expression recognition using a CNN trained from scratch with PyTorch.

Gemlik zeytininde Halkalı Leke ve Pas Akarı için YOLOv8-Seg & EfficientNet-B0 hibrit yapay zekâ teşhis motoru ve zirai reçete sistemi. (TÜBİTAK 2204-A & TEKNOFEST)

Deep learning audio classification pipeline for wildlife acoustic monitoring using Librosa and TensorFlow CNN

Paper ECG photo → digital signal → interpretation, running entirely in the browser (YOLOv8, 1D U-Net, Pyodide, ONNX Runtime Web)

⚽ K League × University of Seoul open AI competition, Track 1 — pass end-point prediction with a dense-supervised BiLSTM + heatmap CNN (2nd place, Gold Award)