A doctor for your ML training runs: reads TensorBoard/MLflow/W&B/stdout logs and diagnoses NaNs, divergence, overfitting, idle GPUs and memory leaks. CLI + skill for coding agents.

Official PyTorch implementation of LHRF-YOLO: a lightweight model with hybrid receptive field for forest fire detection (Forests 2025, 16, 1095).

NVFP4 GEMM, GEMV, linear layers and quantization for NVIDIA GPUs without native FP4 (sm_50–sm_90): every NVIDIA NVFP4 variant (ModelOpt, Transformer Engine 1x16 / 16x16, 4over6, swizzled scales) bit for bit, exact block sums, tensor cores on sm_75–sm_89, C API and PyTorch bindings

A modular pipeline for EEG signal conditioning, ICA ocular artifact removal, and EEGNet decoding.

Reproducible evaluation of U-Net for colorectal polyp segmentation, focusing on cross-dataset generalization across Kvasir-SEG and CVC-ClinicDB.

Meta's AI audio generation library: create music (MusicGen) or sound effects (AudioGen) from text descriptions. PyTorch inference and training code, EnCodec neural audio codec, pretrained models and demos. Modified copy of facebookresearch/audiocraft.

AI-powered image classification application that identifies clothing, footwear, and accessory products using MobileNetV2 and TensorFlow.

Neural image steganography: hide a short ID inside a photo — invisible to people, readable by a phone camera (PyTorch + iOS)

Federated learning for privacy-preserving ECG arrhythmia classification under non-IID data distributions

Differentiable discrete wavelet transforms for PyTorch: DWT, DTCWT, scattering, SWT and wavelet packets.

Robot vs. Robot Tic-Tac-Toe is an advanced open-source robotics project powered by Niryo Ned2 6-axis arms, robust computer vision, and an unbeatable Alpha-Beta Minimax decision-making engine.

From Subcellular Spatial Biophysics to Routine Digital Pathology AI: Resolving Non-Genetic Osimertinib Resistance via Microvascular Decoupling

Custom CNNs on CIFAR-10 with a leave-one-out ablation of batch norm, augmentation and dropout, plus Grad-CAM

Filling cloud gaps in Baltic sea temperature with calibrated uncertainty, tested on an unseen basin and the 2018 marine heatwave.

Browser-based interview practice tool for Data Science, Machine Learning, and Deep Learning.

Research dataset for AI-driven next-24-hour IoMT cybersecurity risk prediction using network, DICOM, and endpoint telemetry features in healthcare environments.

Open-access course materials on Explainable Artificial Intelligence (XAI), covering foundational concepts, explanation methods, evaluation, and hands-on Python exercises.

Quant ML research platform: deep learning on CME Level-3 order-book data (CNN-Mamba, PatchTST), Rust feature pipeline, Ray cluster + MLflow, walk-forward/OOT validation, RL execution, options & systematic strategy research

TSM-based IPN-Hand gesture recognition (RGB, 8 frames). Adds dynamic 3x3 kernels to CoordAtt, class-balanced sampling and label-swapping horizontal flip. PyTorch.基于 TSM 改造的 IPN-Hand 手势识别（RGB，8 帧），含 CoordAtt 坐标注意力动态卷积核、 类别均衡采样与水平翻转等优化。PyTorch 实现。

Ammu-1.0 FactCheck Need Classifier Models

Top Deep Learning Deployment (Opensource) Star if you like it!

Mechanistic interpretability research suite: attention head ablation, induction head discovery, and activation patching on Arabic LLMs.

Differentially private CNN on MNIST with PyTorch + Opacus DP-SGD: 90.0% test accuracy at ε = 0.30 (δ = 1e-5).

Implicit-GEMM Conv3d for AMD RDNA4 (gfx1201) via Triton WMMA

An interactive visual journey through convolution, feature maps, pooling and prediction in CNNs. Runs entirely in the browser.

Deep-learning reconstruction of subsurface ocean temperature (15 depths, 0–1000 m) from daily satellite surface fields over the North Indian Ocean, validated against GLORYS and Argo floats, with a React + FastAPI explorer.

From-scratch C++20 convolutional neural network for MNIST, with sequential, parallel and CUDA training.

Entropy-guided knowledge distillation (ResNet-34 to ResNet-18) for 9-class wound image classification: code and models behind an ICCI 2025 paper, with honest notes on evaluation caveats.

The first trainable, 100% domain-agnostic PyTorch Cortical MCP & Multi-Agent Engine. Wakes n (default=4) orthogonal specialists from a 1-sentence prompt, gates working memory at 0 routing tokens (-60% context), verifies designs in Python/SciPy before answering, and trains its own weights on the fly via Dream-RSI.

AI Engineer — LLMs & Agents, Machine Learning, Deep Learning, NLP, Computer Vision. Personal GitHub profile README.