Projects

### Overview: Open Interface is a powerful tool that allows you to control any computer using Large Language Models (LLMs). It automates the process of sending requests to an LLM backend to figure out the necessary steps, simulating keyboard and mouse input, and course-correcting with updated screenshots as needed. ### Key Features: - **Self-Drives Your Computer**: Automatically executes tasks by sending requests to an LLM backend. - **Automated Steps**: Simulates keyboard and mouse input to complete tasks. - **Progress Tracking**: Sends screenshots to the backend to monitor progress. ### Tech Stack: Python ### Why It Matters: Open Interface simplifies the process of controlling computers using LLMs, making it accessible and efficient for a wide range of tasks, including solving Wordle, creating meal plans, and developing web applications.

### Overview: FractalAndroid is a human-centric, resource-aware Android node designed for decentralized intelligence and federated learning. It orchestrates distributed computing with privacy-first localization and hardware management. ### Key features: - Resource-aware execution - Federated learning capabilities - Distributed compute node - Privacy-preserving localization - Empathetic hardware management ### Tech stack: Kotlin ### Why it matters: FractalAndroid leverages the power of Android devices to execute localized machine learning workloads, ensuring zero impact on user experience and battery longevity. It integrates with the Fractal decentralized network, making it ideal for decentralized intelligence and federated learning applications.

### Overview: ml.js is a collection of machine learning tools written in JavaScript, designed for use in both the browser and Node.js environments. ### Key features: - Principal component analysis (PCA) - Hierarchical clustering - K-means clustering - Naive Bayes - K-Nearest Neighbor (KNN) - Partial least squares (PLS) - K-OPLS - Cross-validation - Confusion matrix - Decision tree classifier - Random forest classifier - Feedforward Neural Networks - Self-organizing map / Kohonen networks - Simple linear regression - Polynomial regression - Multivariate linear regression - Power regression - Exponential regression - Theil-Sen regression - Robust polynomial regression - Decision tree regression - Random forest regression - Levenberg-Marquardt - Fast Combinatorial Non-negative Least Squares - Matrix - Singular value decomposition (SVD) - Eigenvalue decomposition (EVD) - Cholesky decomposition - Lu decomposition - Qr decomposition - Sparse matrix - Kernels - Distance functions - Similarity functions - Distance matrix - XORShift-add RNG - non-negative Generalized Morphological Component Analysis ### Tech stack: JavaScript ### Why it matters: ml.js provides a comprehensive set of machine learning tools for JavaScript developers, making it easier to implement and experiment with various machine learning algorithms directly in the browser or Node.js environment.

### Overview: Intel® Neural Compressor is an open-source Python library designed for compressing models across various deep learning frameworks, including PyTorch, TensorFlow, and ONNX Runtime. It supports leading model compression techniques such as static and dynamic quantization, and provides support for advanced quantization of Large Language Models and Vision-Language Models. ### Key Features: - **Static Quantization**: Compresses models by converting weights to lower precision. - **Dynamic Quantization**: Automatically adjusts the precision of weights based on the model's usage. - **SmoothQuant**: A technique that reduces quantization error by smoothly interpolating between quantized values. - **Weight-Only Quantization**: Quantizes only the weights of the model, leaving the activations in their original precision. - **Quantization-Aware Training**: Trains the model with quantized weights to further improve compression. - **Mixed Precision**: Utilizes both FP32 and FP16 for efficient computation. - **Supports Multiple Hardware**: Intel Neural Compressor is compatible with Intel Gaudi AI Accelerators, Intel Core Ultra Processors, Intel Xeon Scalable Processors, Intel Xeon CPU Max Series, and Intel Data Center GPU Flex Series. ### Tech Stack: Python ### Why it Matters: Intel® Neural Compressor simplifies the process of compressing and quantizing deep learning models, making it easier for developers to create more efficient and faster AI applications. Its support for a wide range of hardware and frameworks makes it a versatile tool for various AI projects.

### Overview: Trieve is an all-in-one platform offering search, recommendations, RAG (Reinforcement Learning with Attention), and analytics via APIs. It provides a self-hosted solution in VPC or on-premises, integrating with OpenAI and Jina embedding models for semantic vector search, and leveraging Qdrant for quality neural sparse-vector search. ### Key features: - **Self-Hosting in VPC or on-premises**: Full self-hosting guides available for AWS, GCP, Kubernetes, and Docker Compose. - **Semantic Dense Vector Search**: Integrates with OpenAI or Jina embedding models and Qdrant for semantic vector search. - **Typo Tolerant Full-Text/Neural Search**: Every uploaded chunk is vector'ized with naver/efficient-splade-VI-BT-large-query for typo tolerant, quality neural sparse-vector search. - **Sub-Sentence Highlighting**: Highlights matching words or sentences within a chunk and bolds them on search for enhanced user experience. - **Recommendations**: Find similar chunks or files with the recommendation API, useful for platforms with user favorites, bookmarks, or upvotes. - **Convenient RAG API Routes**: Access to any LLM via OpenRouter for RAG, including fully-managed RAG with topic-based memory management and select context RAG. - **Bring Your Own Models**: Allows integration with your own models. ### Tech stack: Rust ### Why it matters: Trieve is a powerful, all-in-one platform for search, recommendations, and analytics, offering a

Overview: SkillClaw is an AI agent skills evolution platform that enables skills to grow and improve through every interaction across sessions, agents, and devices. It supports a wide range of compatible agents and APIs, including Hermes, OpenClaw, QwenPaw, IronClaw, PicoClaw, ZeroClaw, and more. Key features: - Skills evolve from every session, agent, and context. - Integrates with multiple OpenAI-compatible APIs. - Supports a broad compatibility with various agents and devices. - Enables collective skill evolution for teams. Tech stack: Python Why it matters: SkillClaw simplifies the process of skill evolution by making it seamless and efficient, allowing users to focus on their goals without worrying about the underlying technicalities.

### Overview: 本仓库整理了人工智能会议（如 ACL、EMNLP、NAACL、COLING、AAAI、IJCAI、ICLR、NeurIPS、ICML 等）中开源代码的论文，便于检索与参考。 ### Key features: - 收集了顶会论文及其开源代码资源 - 包括 ACL、EMNLP、NAACL、COLING、AAAI、IJCAI、ICLR、NeurIPS 和 ICML 等会议 - 提供了详细的论文和代码整理 ### Tech stack: - 代码语言：Python、Java、C++ 等 ### Why it matters: 论文篇幅有限，通常很难覆盖全部实现细节。配套开源代码能提供更多可复现信息，方便后续跟进与复现。

Overview: SecretFlow is a unified framework designed for privacy-preserving data analysis and machine learning. Key features: - An abstract device layer with plain and secret devices for encapsulating cryptographic protocols. - Device flow modeling for higher algorithms as device object flows and Directed Acyclic Graphs (DAGs). - Horizontal or vertical partitioned data for algorithm execution. - Seamless integration of data processing, model training, and hyperparameter tuning. Tech stack: Python Why it matters: SecretFlow simplifies privacy-preserving data analysis and machine learning by providing a unified framework that ensures data privacy while enabling secure and efficient computation. Repository: secretflow/secretflow GitHub stars: 2716 GitHub description: A unified framework for privacy-preserving data analysis and machine learning

### Overview: This repository provides a production-ready FastAPI template for building AI agent applications with LangGraph integration. It includes all the necessary components to handle stateful conversations, long-term memory, tool calling, observability, rate limiting, and authentication, allowing developers to focus on their agent logic. ### Key Features: - **LangGraph Stateful Agent**: A robust foundation for building scalable, secure, and maintainable AI agent services. - **LangGraph Integration**: Seamless integration with LangGraph for stateful conversations, tool calling, and long-term memory. - **Rate Limiting and Authentication**: Built-in rate limiting and authentication mechanisms to ensure secure and efficient service usage. - **LLM Registry**: A unified endpoint for accessing multiple LLM models, including OpenAI, Anthropic, DeepSeek, Qwen, and more. - **Easy Setup**: Quick setup with minimal configuration required, including a free API key from Atlas Cloud for seamless integration with LangGraph. ### Tech Stack: - **Python**: The primary language used for this template. - **FastAPI**: The backend framework powering the application. - **LangGraph**: The stateful agent framework providing essential functionalities. ### Why It Matters: This template is designed for AI engineers who want a solid foundation for their projects, not a tutorial project. It handles the hard parts, ensuring that developers can focus on their core logic while leveraging the comprehensive features provided by LangGraph and Atlas Cloud. ### GitHub Stars: 2715 ### Repository: - **Repository**: wassim249/fastapi-langgraph-agent-production-ready-template

### Overview dLLM is a library that simplifies the training and evaluation of diffusion language models, making it easier to reproduce and verify results. It provides scalable training pipelines and unified evaluation pipelines, abstracting away inference details for simplicity. ### Key features: - Scalable training pipelines based on `transformers` and `Trainer`. - Support for LoRA, DeepSpeed, and FSDP. - Unified evaluation pipelines based on `lm-evaluation-harness`. - Minimal training, inference, and evaluation recipes for open-weight models. - Implementations of training algorithms like MDLM, BD3LM, and Edit Flows. ### Tech stack: Python ### Why it matters: dLLM simplifies the development process by providing transparent and reproducible training and evaluation pipelines, making it easier to understand and verify the results of diffusion language models.

Overview: Torch-Points3D is a PyTorch framework designed for deep learning on point clouds, built on PyTorch Geometric and Facebook Hydra. It provides a high-level API for democratizing deep learning on point clouds, with support for 3D sparse convolution. Key features: - Implementation of PointNet from Charles R. Qi et al. - Support for various tasks such as segmentation, classification, registration, object detection, and panoptic segmentation - A modular design for easy integration and reproducibility - Compatibility with CUDA 10 or higher and Python 3.7 or higher Tech stack: PyTorch, PyTorch Geometric, Facebook Hydra Why it matters: Torch-Points3D offers a streamlined approach to deep learning on point clouds, making it accessible and efficient for researchers and developers. Its modular design and compatibility with popular deep learning frameworks make it a valuable tool for researchers and practitioners alike. Repository: torch-points3d/torch-points3d GitHub stars: 2714 GitHub description: Pytorch framework for doing deep learning on point clouds.

Overview: Semantra is a multipurpose tool designed for semantically searching documents, making it easy, friendly, and secure to run specialized semantic search engines. Key features: - Analyzes specified text and PDF files on your computer - Launches a local web search application for interactive querying - Built for individuals seeking needles in haystacks, including journalists, researchers, students, and historians Tech stack: Python Why it matters: Semantra simplifies the process of running a semantic search engine, making it accessible and user-friendly for a wide range of applications. Repository: freedmand/semantra-python Primary language: Python GitHub stars: 2712 GitHub description: Multi-tool for semantic search

Overview: This is a multi-agent collaboration system inspired by the Ming Dynasty's Cabinet and the Six Ministries, built using the OpenClaw framework. It allows for efficient management of AI teams and seamless coordination across various departments. Key features: - Multi-Agent Collaboration System - OpenClaw Framework Integration - 18 Agents for Modern Governance - 14 Agents for Traditional Governance - 5 Agent Collaboration in Writing Fiction (Optional) - 60+ Skills for Skillful Management - Cron Jobs for Automation - Sandbox for Secure Testing Tech stack: TypeScript Why it matters: This system provides a practical solution for managing AI teams in both traditional and modern governance setups, offering a balance between efficiency and security. Repository: wanikua/danghuangshang Primary language: TypeScript GitHub stars: 2709 GitHub description: Open-source multi-agent collaboration system inspired by Chinese governance — deploy and coordinate specialized AI agents with OpenClaw. README excerpt: [](https://mseep.ai/app/wanikua-danghuangshang) [English Version](./README_EN.md) | [日本語版](./README_JA.md) | [🏢 企业版 Become CEO (English)](https://github.com/wanikua/become-ceo) | [🔧 天工开物 TianGong (Terminal版)](https://github.com/wanikua/t

### Overview: MUNIT is a state-of-the-art multimodal unsupervised image-to-image translation model that achieves state-of-the-art performance across various image-to-image translation tasks. ### Key features: - **Multimodal Translation**: MUNIT can translate images from one modality to another, including image-to-image translation, edge-to-shoes/handbags translation, animal image translation, street scene translation, and summer-to-winter translation. - **Unsupervised Learning**: The model is trained entirely without labeled data, making it highly effective for real-world applications where labeled data is scarce. - **High-Quality Results**: MUNIT produces high-quality translations that are visually appealing and contextually appropriate. ### Tech stack: Python ### Why it matters: MUNIT is a groundbreaking model that has significantly advanced the field of image-to-image translation. Its ability to perform tasks like edge-to-shoes/handbags translation, animal image translation, and street scene translation with high accuracy makes it a valuable tool for researchers and practitioners alike. Its unsupervised learning approach also makes it highly applicable in scenarios where labeled data is not readily available, making it a versatile and powerful tool in the field of computer vision.

### Overview: YAO is a rigorous engineering, evaluation, governance, and portability system designed for creating, evaluating, packaging, and governing reusable agent skills. The 1.0 version focused on turning workflows into installable, readable, cross-platform skill packages. The 2.0 version expands this factory into a Skill OS, providing a governed system for modeling, compiling, testing, reviewing, and tracking skill iterations. ### Key features: - **Skill OS 2.0 Upgrade**: Explicitly tracks the lifecycle of skills, adding semantic contracts, target compilers, evaluation evidence, release gates, and operation reports. - **Skill IR**: A platform-neutral intermediate representation for intent, triggers, inputs, outputs, boundaries, references, and expected artifacts. - **Target Compilers and Adapters**: Generated surfaces for OpenAI, Claude, generic agent skills, and VS Code-oriented workflows. - **Output Eval Lab**: Checks trigger checks, output assertions, execution evidence, timing and token evidence, benchmark reproducibility, blind-review packs, answer keys, and adjudication reports. - **Review Studio 2.0**: A single HTML gate page for intent, triggers, output eval, context cost, runtime checks, trust, Skill Atlas signals, adoption drift, waivers, annotations, release evidence, warnings, blockers, and fix actions. - **Evidence and Release Governance**: Consistency checks, package verification, install simulation, runtime permission probes, intake of world-class evidence, ledger, operator runbook, and public claim guard. - **SkillOps Loop**: Metadata-only adoption drift, telemetry hooks,

Overview: Time-LLM is an official implementation of "Time-LLM: Time Series Forecasting by Reprogramming Large Language Models," presented at ICLR 2024. It aims to reprogram large language models for time series forecasting tasks. Key features: - Reprogramming input time series into text prototype representations. - Augmenting input context with declarative prompts to guide LLM reasoning. - Supports LLM models like Llama-7B and PLMs like GPT-2 and BERT. Tech stack: Python 3.11, torch, accelerate, einops, matplotlib, numpy, pandas, scikit_learn. Why it matters: Time-LLM provides a novel approach to time series forecasting by leveraging the strengths of large language models, making it a valuable tool for researchers and practitioners in the field.

### Overview: OpenPAI is a resource scheduling and cluster management platform designed for AI applications. It offers a robust, modular architecture and a user-friendly interface for AI tasks. ### Key features: - **Resource Scheduling**: Efficiently manage and allocate resources across multiple clusters. - **Cluster Management**: Simplify the setup and management of AI clusters. - **Modular Architecture**: Easily customize and expand the platform to meet new needs. ### Tech stack: JavaScript ### Why it matters: OpenPAI is a crucial tool for researchers and developers working on AI projects, providing a reliable and scalable environment for AI tasks. Its modular design and user-friendly features make it accessible and efficient for both beginners and advanced users. ### Repository: microsoft/pai ### GitHub stars: 2,700 ### GitHub description: Resource scheduling and cluster management for AI ### Release Notes: After the release of v1.8.1, OpenPAI has entered stable mode with no major feature releases planned. The repository is now in read-only mode for collaboration. For further information, please contact the repository admin.

### Overview: YT Pro is a YouTube client application designed for older Android versions, featuring background player functionality, Google Gemini, and numerous additional features. It supports video summarization, video downloader, shorts downloader, video and audio muxer, thumbnail downloader, caption downloader, ads blocker, minimize video, picture in picture mode, gesture control, dislike count display, background audio player, custom heart feature, media codec management, skip sponsor, force zoom, hide shorts, 10x video speed, minimal APK size, adaptive UI icons, and minimal internal dependencies. The application is open-source and available on GitHub. ### Key Features: - **Google Gemini**: Summarizes videos and allows customization of prompts and models. - **Video Downloader**: Downloads videos directly from YouTube. - **Shorts Downloader**: Downloads short videos. - **Video and Audio Muxer**: Merges video and audio streams. - **Thumbnails Downloader**: Downloads video thumbnails. - **Captions Downloader**: Downloads video captions. - **Ads Blocker**: Blocks ads. - **Minimize Video**: Allows users to minimize the video window. - **Picture in Picture Mode**: Enables users to watch videos in a separate window. - **Gesture Control**: Controls volume and brightness using gestures. - **Shows Number of Dislikes**: Displays the number of dislikes for a video. - **Background Audio Player**: Plays audio in the background. - **Custom Heart Feature**: Saves videos without logging into your account. - **Enable / Disable Media Codecs**: Manages media codecs. - **Skip Sponsors**:

Overview: InvoiceNet is a deep neural network designed to extract intelligent information from invoice documents, making it easier to view and extract details from PDF, JPG, and PNG invoices. Key features: - An easy-to-use interface for viewing and extracting information from invoices. - The ability to train custom models using the Trainer UI on your own dataset. - Customizable fields for adding or removing as needed. - Save extracted information directly into your system with a single click. Tech stack: Python Why it matters: InvoiceNet simplifies the process of extracting valuable information from invoice documents, making it accessible to developers and businesses alike. The project addresses the challenge of collecting large datasets for training general invoice models, which is currently a bottleneck in the community.

Overview: maestro is a streamlined tool designed to accelerate the fine-tuning of multimodal models, specifically focusing on vision-language models such as Florence-2, PaliGemma 2, and Qwen2.5-VL. Key features: - Streamlines configuration, data loading, reproducibility, and training loop setup. - Offers ready-to-use recipes for popular models including Florence-2, PaliGemma 2, and Qwen2.5-VL. - Supports LoRA, QLoRA, and graph freezing techniques to optimize hardware requirements. - Provides a consistent JSONL format for data handling, making it easier to manage and process data. Tech stack: Python Why it matters: maestro simplifies the complex process of fine-tuning multimodal models, making it accessible and efficient for developers and researchers alike.