Run very large MoE models (100B–1T+) on the hardware you already own, by streaming experts across VRAM/RAM/SSD. Source-available, free for personal use.

A curated list of free & open-source AI tools, models, APIs and learning resources for developers.

Neural noise suppression and echo cancellation engine: bit-exact 16 kHz models, dependency-free Rust, runs natively and in the browser via WebAssembly.

Which speech denoiser works best, when, and at what cost? 11 algorithms, one benchmark, a Simulink model and an interactive web app.

Real-time, animated Streamlit app for document layout parsing: YOLOv10x + DocLayout-YOLO with ensemble fusion and region names in 12 Indian languages. Built on IndicDLP.

ATCNet vs. lightweight Mini-ATCNet for motor imagery EEG classification on BCI Competition IV-2a. Controlled accuracy-vs-efficiency comparison with leakage-free, subject-dependent 5-fold CV.

A Library for Full-Parameter and LoRA Fine-Tuning of Timestral Time Series Foundation Models.

Design, train, and optimize Large Language Models with an AI-powered visual engineering platform. Build custom Transformer architectures, import pretrained models, and generate runnable PyTorch code. An integrated AI agent assists with debugging, while a compute estimator calculates hardware costs.

Deterministic benchmark evaluating Vision-Language Models (VLMs) on geometric perception and SVG vector reconstruction under parsimony budgets.

Founder & Systems Architect @RyNxt. M.Sc. in CE. Low-level OS kernel optimization, DirectStorage, C++ & real-time pathology AI researcher.

A full-process YOLO object detection guide tool for beginners: automatically detects GPU/CPU and recommends training parameters, comes with built-in dataset generation and image labeling, real-time training logs and result visualization, supports real-time camera recognition and batch image recognition. Developed with PyQt5 + PyTorch + Ultralytics,

MUH-920 Artificial Intelligence Applications in Engineering, SDÜ. Open 14-week undergraduate course for all engineering and science departments: Agents, search, fuzzy logic, evolutionary and swarm optimization, machine and deep learning, generative AI and reinforcement learning, with interactive labs, Colab notebooks and a Python track.

How a Vision Transformer works, step by step: one MRI image through patches, self-attention and [CLS], re-built in NumPy

Official code of MonarchGRU (IEEE AICAS 2026): a compact GRU with Monarch weights for always-on keyword spotting, 97.9% on Google Speech Commands v2.

ThinkNet is an all-in-one AI platform to design LLMs, build reasoning models, implement research papers, and automate AI engineering. Explore multi-agent systems, multimodal AI, and robotics with 20+ AI libraries, compute estimation, dataset tools, quantization, and GPU support. Train and optimize AI models while accelerating research.

AI-powered road damage detection and severity assessment using deep learning and computer vision, with an interactive web application for analyzing road conditions.

A curated reading list on feature superposition in neural networks, plus a daily arXiv feed on superposition and mechanistic interpretability.

BacteriaCV - Dataset preparation pipeline for DIBaS bacterial microscopy images with species mapping, leakage-safe 70/20/10 splits, and five-fold training folds.

Explainable AI prototype in MATLAB for diabetic retinopathy screening from fundus images, with quality checks, Grad-CAM explanations, auto-generated reports and a web-assisted referral lookup.

Text-to-SQL with a 852k-parameter transformer trained from scratch, compared with LoRA and a 120B API model on identical questions.

Verified JAX/Equinox ports of Hugging Face LLMs and embedding models: numerically matched against PyTorch, forward and backward, and saved back to Hugging Face formats.

A visual, intuition-first path from basic Python to AI research fundamentals: PyTorch, nanoGPT, DPO, and DeepSeekMath GRPO loss walkthrough.

Draw PyTorch model architectures as clean figures for papers and slides, in one call.

🍕 Recognize 20 Indian food items using MobileNetV2 transfer learning + Streamlit

Non-Transformer language model with a self-organizing memory pool and live introspective signals. Code, checkpoints and controls, 30M–125M.

297 classic multimodal models: educational PyTorch re-implementations with exact arXiv-verified paper references and CPU smoke tests

How a CNN works, step by step: one MRI image through convolution, ReLU, pooling and softmax, built from scratch in NumPy

PoC de cumplimiento de planogramas en góndolas retail con Ultralytics YOLO26: detección de productos y huecos, agrupación por estantes y reporte de cumplimiento.

Claude watches your training runs so you can sleep. A Claude Code plugin for one run or a whole SLURM sweep: catches NaNs, exploding grads, OOMs, crashes, hangs and dead jobs, resumes from a safe checkpoint, and leaves you a morning report.

Interactive simulations of how AI systems work, with the math behind each one and the caveats on where toy models stop matching real LLMs.