Projects · free and open source
The simplest, fastest repository for training and fine-tuning medium-sized GPTs in PyTorch with clean, hackable, zero-abstraction code.
Simple, pure C and CUDA implementation for training GPT-2 and GPT-3 without 245MB of PyTorch runtime overhead.
A tiny 100-line scalar autograd engine implementing backpropagation over a dynamically built DAG with a PyTorch-like API.
End-to-end, production-grade AI agent projects built from scratch with Agno, AutoGen, OpenAI Swarm, Google ADK, and Ollama — from investment advisors to coding agents.
Dozens of practical tutorials and implementations for generative AI agent architectures — from conversational bots to multi-agent workflows with LangGraph, AutoGen and PydanticAI.
Build robust, stateful multi-agent applications with cyclic graphs, human-in-the-loop checkpoints, and streaming execution.
Autonomous web agent framework that converts natural language intent into structured DOM interactions and browser workflows with vision models.
Let language models run code locally in Python, JavaScript, Shell, and AppleScript inside a sandboxed environment.
Open-source autonomous agent runtime executing shell commands, file workflows, and browser tasks with local LLMs and messaging gateways.
A persistent, self-improving memory infrastructure for LLMs and agents — dynamically extracting facts, profiling users, and updating memory over time.
An extensive collection of runnable Jupyter notebooks implementing every major RAG strategy — from basic chunking to corrective RAG, GraphRAG, and agentic retrieval.
Clean, documented Jupyter notebooks building RAG step by step — indexing, chunking, retrieval, query translation, RAG fusion, and generation — with a video series alongside.
Free, open-source fine-tuning framework that makes LoRA/QLoRA 2–5× faster with 80% less memory via hand-optimized CUDA kernels — runs on consumer GPUs.
Comprehensive framework supporting SFT, DPO, PPO, ORPO, and KTO for fine-tuning 100+ LLMs with a web UI and CLI — the Swiss Army knife of LLM training.
Port of LLaMA model in pure C/C++ with 1.5-bit to 8-bit integer quantization, running on Apple Silicon Metal, AVX2 x86, and NVIDIA GPUs.
The industry-standard production inference engine delivering 2x–4x higher throughput via PagedAttention and continuous batching.
High-performance serving engine for complex LLM workflows with RadixAttention automatic KV cache reuse across multiple calls.
A 12-week, 26-lesson curriculum by Microsoft covering classical machine learning with scikit-learn — with quizzes, assignments, and real-world datasets.
Popular machine learning algorithms explained with math and demo notebooks — linear regression, logistic regression, neural networks, anomaly detection and more.
Comprehensive machine learning interview preparation covering ML system design, classical algorithms, deep learning, NLP, and production ML — used by thousands of engineers.
GitHub stars as of November 2026.