Projects Recently Starred by Experts

Starred 1 hour ago by

Luis Capelo(Cofounder of Lightning AI). Also starred by

Jeff Hammerbacher(Cofounder of Cloudera),

Georgios Konstantopoulos(CTO, General Partner at Paradigm),

Andrey Vasnetsov(Cofounder of Qdrant), and

4 more.

rtk by rtk-ai

3.0%

65k

CLI proxy for massive LLM token reduction

Compresses CLI command outputs for LLMs, slashing token consumption by 60-90%.
Integrates via hooks to transparently rewrite commands, optimizing LLM context.
High-performance Rust binary provides detailed token savings analytics and discovery.
Optimizes common dev commands (git, grep, files) with full output recovery.

Created 5 months ago

Updated 1 day ago

Starred 1 hour ago by

Eric Zhang(Founding Engineer at Modal). Also starred by

Mehdi Amini(Author of MLIR; Distinguished Engineer at NVIDIA),

Chip Huyen(Author of "AI Engineering", "Designing Machine Learning Systems"),

Jeff Hammerbacher(Cofounder of Cloudera), and

14 more.

flashinfer by flashinfer-ai

0.6%

Kernel library for LLM serving

High-performance CUDA kernels for attention: FlashAttention, SparseAttention, PageAttention.
Load-balanced scheduling alleviates load imbalance for variable-length inputs.
Memory efficiency via Cascade Attention, Head-Query fusion, low-precision & fused-RoPE.
Customizable attention variants via JIT compilation; CUDAGraph & torch.compile compatible.

Created 2 years ago

Updated 11 hours ago

Starred 12 hours ago by

Jesse Clark(Cofounder of Marqo). Also starred by

Luca Soldaini(Research Scientist at Ai2),

Jeremy Howard(Cofounder of fast.ai),

Shyamal Anadkat(Research Scientist at OpenAI), and

3 more.

llm-datasets by mlabonne

0.2%

Curated datasets/tools for LLM post-training

SFT datasets for turning pre-trained models into helpful assistants.
Math datasets enhance mathematical reasoning and formal logic.
Code datasets improve understanding, generation, and analysis.
Preference datasets align LLM answers with human values/style.

Created 2 years ago

Updated 1 month ago

Starred 13 hours ago by

Matei Zaharia(Cofounder of Databricks). Also starred by

Elvis Saravia(Founder of DAIR.AI),

Michael Chiang(Cofounder of Ollama),

Andy Konwinski (Cofounder of Perplexity, Databricks), and

4 more.

omnigent by omnigent-ai

58.2%

Unify and manage AI agents across platforms

Supervise multiple AI agents (Claude, Codex, Pi, custom) in a single session.
Enables cross-device collaboration and real-time session syncing from terminal to phone.
Supports agent governance via policies for sandboxing, spend caps, and action approvals.
Allows defining custom agents via YAML, including tools and sub-agent delegation.

Created 1 week ago

Updated 8 hours ago

Starred 13 hours ago by

Matei Zaharia(Cofounder of Databricks).

PixelRAG by StarTrail-org

1250.8%

Visual RAG for documents and web pages

Search and retrieve information from document screenshots, preserving visual structure lost in text parsing.
Supports multimodal inputs (web, PDF, images) using Qwen-VL embeddings for visual understanding.
Offers rendering pipelines, FAISS indexing, and a hosted API for scalable visual search.
Integrates with Claude as a plugin, enabling LLMs to "see" and analyze visual content.

Created 3 weeks ago

Updated 13 hours ago

Starred 1 day ago by

Elie Bursztein(Cybersecurity Lead at Google DeepMind). Also starred by

Emil Ernerfeldt(Cofounder of Rerun) and

Dan Guido(Cofounder of Trail of Bits).

nono by always-further

1.7%

Kernel-enforced capability sandbox for AI agents

Runs untrusted AI agents and CLI processes with OS-level security primitives.
Leverages Linux Landlock and macOS Seatbelt for robust, structural security.
Offers fine-grained control over file access and network permissions.
No escape hatch: once inside, restrictions cannot be bypassed.

Created 4 months ago

Updated 18 hours ago

Starred 1 day ago by

Tim J. Baek(Founder of Open WebUI). Also starred by

Yaowei Zheng(Author of LLaMA-Factory),

Max Howell(Author of Homebrew), and

Amanpreet Singh(Cofounder of Contextual AI).

ponytail by DietrichGebert

101.8%

50k

AI agent code generation optimized for minimal output

Reduces code (80-94%), speeds execution (3-6x), and cuts costs (47-77%) via conciseness.
Integrates as a plugin across AI platforms (Claude, Codex, etc.) and models (Haiku, Sonnet).
Prioritizes YAGNI, stdlib, and existing dependencies before generating minimal, working code.
Offers features like code review for deletion and adjustable agentic verbosity levels.

Created 1 week ago

Updated 19 hours ago

Starred 1 day ago by

Li Jiang(Coauthor of AutoGen; Engineer at Microsoft).

how-to-train-your-gpt by raiyanyahya

0.6%

Build a modern LLM from scratch, line by line

Implement tokenizers, embeddings, attention, and full Transformer architectures from the ground up.
Master LLM training pipelines with AdamW, mixed precision, and custom inference engines.
Explore modern techniques like RoPE, RMSNorm, and SwiGLU powering LLaMA and Mistral.
Gain deep understanding of LLM internals via annotated code and practical, worked examples.

Created 1 month ago

Updated 1 week ago

Starred 1 day ago by

Didier Lopes(Founder of OpenBB). Also starred by

Pawel Garbacki(Cofounder of Fireworks AI),

Yaowei Zheng(Author of LLaMA-Factory),

Dan Guido(Cofounder of Trail of Bits), and

11 more.

harbor by harbor-framework

5.4%

Harness for AI agent evaluation and RL optimization

Evaluate diverse agents (Claude Code, Codex) and LLMs against benchmarks.
Create, share, and run parallel experiments across thousands of environments.
Generate agent rollouts for Reinforcement Learning optimization tasks.
Supports cloud execution via Daytona and Modal providers.

Created 10 months ago

Updated 11 hours ago

Starred 1 day ago by

Jesse Clark(Cofounder of Marqo). Also starred by

Lewis Tunstall(Research Engineer at Hugging Face),

Abubakar Abid(Cofounder of Gradio),

Will DePue(Coauthor of Sora), and

11 more.

modded-nanogpt by KellerJordan

0.6%

Language model training speedrun on 8x H100 GPUs

Trains a language model to 3.28 cross-entropy loss on FineWeb validation set.
Achieves target performance in 3 minutes using modern architectures and Muon optimizer.
Employs techniques like FP8 matmul, skip connections, and FlexAttention.
Includes value embeddings and window size warmup inspired by Gemma 2.

Created 2 years ago

Updated 1 day ago

...