Projects

### Overview: 「摘星」· 微信助手是一款集关键词提醒、多平台消息推送、公众号摘要与提醒、聊天归档、收藏整理、朋友圈归档、AI 对话、朋友圈归档、MCP 工具扩展等功能于一体的微信助手，旨在帮助用户高效管理微信群聊、公众号和朋友圈信息。 ### Key features: - **关键词提醒**：在指定群聊设置关注词，命中后即时推送到已绑定的平台。 - **多平台消息推送**：支持绑定微信、QQ 机器人与飞书，扫码即用；摘要、提醒等推送可自由选择投递渠道。 - **群聊 AI 摘要**：多个群聊打包成一个分组，定时生成一份合并摘要（按群分段），消息量大时自动降级，AI 标注待办与决定。 - **公众号助手**：按主题分组，AI 定时汇总文章生成摘要。 - **公众号实时提醒**：关注的公众号发布新文章时，即时推送 AI 速读摘要到微信。 - **手机扫码访问**：同一局域网下，手机扫描二维码即可访问 Web 控制台。 - **聊天导出**：打包聊天记录为离线 HTML/Json 档案，含文字、图片、语音、视频。 - **收藏整理**：浏览、搜索、筛选收藏内容，AI 对话查询，一键导出。 - **朋友圈归档**：导出为离线 HTML，图片视频完整

### Overview: VoiceBlender is a programmable voice AI platform that integrates SIP, WebRTC, WhatsApp, and WebSocket call legs, providing multi-party mixing, recording, speech AI integrations, and a REST API for easy integration. It supports various codecs, media routing, AI agents, and real-time events, making it a versatile solution for voice communication. ### Key Features: - **Legs (Call Endpoints):** - **SIP** - Inbound and outbound calls over UDP, TCP, and TLS. - **SIP Trunks & Registrations** - Register to PBXs or peers with static IP. - **Codecs** - PCMU, PCMA, G.722, Opus, AMR-WB, AMR-NB. - **WebRTC** - Browser voice via SDP offer/answer with ICE. - **WhatsApp Business Calling** - Inbound and outbound WhatsApp voice calls. - **WebSocket Legs** - Raw PCM over WebSocket for bots and custom clients. - **LiveKit** - Bridge a LiveKit room, one leg per remote participant. - **MoQ (Experimental)** - Media-over-QUIC legs over WebTransport. - **SIPREC** - Session recording server and client. - **Multi-stream SIP Calls** - Several audio streams in one call. - **Media:** - **Rooms** - Multi-party mixing at 8, 16, or 48 kHz. - **Room Bridging** - Connect two rooms with

### Overview: World Arcade is an open-source platform that allows developers to easily access and play generative games locally. These games are typically run on complex deterministic engines, but World Arcade simplifies this process by providing a local environment for users to play. ### Key features: - **Browse Open-Source Game Models:** Collect and browse all available games made open-source by Diamond and other world creators. - **Play Games:** Access and play games like Yume Nikki, CS:GO, Minecraft, and more. - **Change Parameters:** Adjust settings to improve performance on lower-end GPUs, at the cost of game stability/quality. ### Tech stack: Python, Docker, WebRTC, and NVIDIA Container Toolkit. ### Why it matters: World Arcade makes it easy for developers to play and experiment with generative games, even those that are not widely available. It simplifies the process of finding and running these games, making them accessible to a broader audience.

Overview: D-AR is a project that recasts diffusion on pixels as sequential token generation using a Llama backbone, bridging diffusion models and autoregressive models. It aims to improve tokenizers trained on larger datasets and explore text-to-image generation models. Key features: - **D-AR Tokenizer**: A tokenizer trained on ImageNet with 300M parameters. - **DAR Models**: Models trained on ImageNet with 343M and 775M parameters, respectively, using the D-AR tokenizer v1. Tech stack: Python. Why it matters: D-AR represents a significant advancement in the field of diffusion models and autoregressive models, offering improved tokenizers and potential for text-to-image generation.

Overview: xPrivo is an open-source AI chat interface that prioritizes privacy and control, allowing users to chat with any local or external AI model without requiring an account. Key features: - Connects to any OpenAI-compatible API endpoint - Supports local models and custom API keys - No account required - Self-hosted or hosted solution available - Dark/light mode, multi-language support, file processing, and AI experts Tech stack: TypeScript Why it matters: xPrivo offers a privacy-first approach to AI chat, ensuring users retain control over their conversations and data. Repository: xprivo/ai-chat GitHub stars: 137 GitHub description: xPrivo is the privacy-first AI assistant that you can self-host. No accounts, never. Free as in freedom.

### Overview: GaussCtrl is an innovative text-driven 3D Gaussian splatting editing framework presented at ECCV 2024. This project leverages stable diffusion models to achieve high-quality 3D text-based editing, providing a versatile solution for various applications. ### Key features: - **Text-Driven Editing**: GaussCtrl uses text prompts to guide the editing process, ensuring consistent results across different views. - **Multi-View Consistency**: The system maintains the consistency of the edited scene across multiple camera perspectives, enhancing the realism and coherence of the final output. - **NeRF-based Splatting**: The editing is performed using NeRF-based splatting techniques, offering high-quality textures and realistic effects. ### Tech stack: - **Python**: The primary programming language used for the project implementation. - **NeRFStudio**: A powerful tool for generating 3D scenes and models, essential for the NeRF-based splatting technique. - **LangSAM**: A state-of-the-art mask extraction model used for enhancing the editing quality. ### Why it matters: GaussCtrl provides a robust and versatile solution for 3D text-driven editing, offering high-quality results with minimal manual intervention. This project showcases the potential of combining text-based editing with advanced machine learning techniques, making it a valuable contribution to the field of computer graphics and virtual reality. Repository: ActiveVisionLab/gaussctrl

### Overview: This repository provides a lean implementation of various multi-agent LLM methods, focusing on Iteration of Thought (IoT). It includes methods developed by Agnostiq Inc. and other state-of-the-art techniques, designed for quick experimentation and caution in production use. ### Key features: - **AIoT (Autonomous Iteration of Thought)**: Dynamically adapts reasoning paths based on the evolving state of the conversation. - **GIoT (Guided Iteration of Thought)**: Ensures a thorough exploration of reasoning paths by forcing the LLM to continue iterating for a predefined number of steps. ### Tech stack: Python ### Why it matters: This repository offers a modular and easy-to-use approach for experimenting with multi-agent architectures and LLMs, focusing on IoT methods developed by Agnostiq Inc.

Overview: This project fine-tunes Qwen2.5-Coder-1.5B-Instruct for Chinese sentiment analysis using the freeze training method, improving accuracy from 91.6% to 97.8% (+6.2%). Key features: - Fine-tuning Qwen2.5-Coder-1.5B-Instruct for Chinese sentiment analysis - Freeze training method, reducing training time by 60-70% and GPU memory usage by 40-50% - Improved accuracy from 91.6% to 97.8% (+6.2%) Tech stack: Python, HuggingFace, Google Colab, local setup Why it matters: The project demonstrates the effectiveness of freeze training in reducing computational and memory requirements while maintaining high performance in sentiment analysis tasks.

### Overview: SEPIA Speech-To-Text (STT) Server is a WebSocket-based, real-time Python server designed for automatic speech recognition (ASR) using open-source engines. It supports multiple ASR engines like Vosk, Coqui, Deepspeech, and Scribosermo, offering standardized APIs and post-processing options. ### Key Features: - WebSocket server capable of receiving audio streams and sending transcribed text in real-time. - Modular architecture supporting multiple ASR engines. - Optional post-processing for enhanced results. - Standardized API for all engines and support for individual engine features. - On-the-fly server and engine configuration via HTTP REST API and WebSocket 'welcome' event. - User authentication via simple token or individual tokens for multiple users. - Docker containers available for all major platform architectures, including Raspberry Pi. - Fast performance to run even on a 2GB Raspberry Pi 4 (realtime). - Compatible with the SEPIA Framework client (v0.24+). ### Tech Stack: Python Fast-API ### Why It Matters: This project aims to provide a standardized, secure, and real-time interface for open-source ASR tools, offering a comprehensive solution for developers and researchers interested in speech recognition.

### Overview: TrustGen is a dynamic and comprehensive framework for evaluating the trustworthiness of Generative Foundation Models (GenFMs) across dimensions such as safety, fairness, robustness, privacy, and more. ### Key features: - **Dynamic Dataset Generation**: Automatically generate datasets tailored for evaluation tasks. - **Multi-Model Compatibility**: Evaluate LLMs, VLMs, T2I models, and more. - **Local & API Model Support**: Run models locally or via API endpoints for maximum flexibility. - **Multi-GPU Acceleration**: Concurrent inference across multiple GPUs for 7-8x speed improvement. - **Advanced T2I Models**: Support for Stable Diffusion 3.5, FLUX.1-dev, HunyuanDiT, Kolors, and more. - **Performance Optimizations**: xformers memory efficiency, attention slicing, and optimized pipelines. - **Customizable Metrics**: Configure workflows with flexible metrics and evaluation methods. - **Metadata-Driven Pipelines**: Design and execute test cases efficiently using metadata. - **Comprehensive Dimensions**: Evaluate models across safety, fairness, robustness, privacy, and trustworthiness. ### Tech stack: Python ### Why it matters: TrustGen provides a robust, flexible, and comprehensive solution for evaluating the trustworthiness of Generative Foundation Models, enabling developers to ensure the safety, fairness, and robustness of their models.

Overview: The Airbyte Agent SDK provides a suite of connectors and tools for working with third-party APIs, offering full IDE autocomplete and local connector support. Key features: - Typed connector tools for working with third-party APIs - Local connector support with user credentials - Tool builders and decorators for turning connector calls into LLM tools - Retrying on transient failures and converting output over max_output_chars Tech stack: Python Why it matters: The SDK simplifies integrating with third-party APIs and provides reliable, permission-aware access with full integration support.

### Overview: This repository contains official skills for FLUX image and video generation from Black Forest Labs. It includes comprehensive guides and API integration patterns for Claude Code, Codex, and agentskills.io-compatible agents. ### Key Features: - **Prompt Guides**: Detailed instructions for crafting effective prompts for FLUX models. - **API Integration Patterns**: Best practices for integrating the BFL API to ensure smooth communication with the model. - **Video Generation**: Tools for creating video content using native audio and a single shot to a finished product ad. ### Tech Stack: TypeScript ### Why it Matters: These skills provide developers with the necessary tools and best practices to create high-quality image and video content using FLUX models, ensuring consistent and effective results.

### Overview: Redbox is an AI-driven application designed to interact with and summarize civil service documents, leveraging Generative AI to streamline administrative processes. ### Key Features: - **AI Chatbot Integration**: Redbox integrates with a Generative AI model to provide real-time summaries and insights. - **Customizable Responses**: Users can customize responses to specific documents, enhancing the chatbot's functionality. - **Document Summarization**: Redbox can summarize various types of administrative documents, including letters, briefings, minutes, and speech transcripts. ### Tech Stack: The project is built using Jupyter Notebook, Python, and LangChain AI. ### Why It Matters: Redbox aims to revolutionize how the civil service works by automating administrative tasks through AI, making the process more efficient and accessible.

Overview: DiffuseMix is an official PyTorch implementation of the paper "DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models" presented at CVPR 2024. This project offers a comprehensive solution for enhancing data augmentation techniques to preserve labels, significantly improving the quality and diversity of generated images. Key features: - Label-Preserving Data Augmentation - Utilizes Diffusion Models for efficient and effective augmentation - Supports various prompts for diverse image styles and aesthetics Tech stack: Python Why it matters: DiffuseMix provides a robust and versatile approach to data augmentation, ensuring that the generated images maintain their original labels, thus preserving the integrity and reliability of the dataset. This makes it an invaluable tool for researchers and practitioners in the field of computer vision and image processing. Repository: khawar-islam/diffuseMix GitHub stars: 136

### Overview: Aria is a local and uncensored AI entity designed for seamless interaction and development. ### Key features: - **Server and Client Mode**: Supports both server and client configurations, allowing for easy integration and deployment. - **Model Support**: Compatible with various models, including larger and more advanced ones. - **Resource Efficiency**: Designed to be efficient and resource-friendly, ensuring smooth operation even on limited hardware. ### Tech stack: Python ### Why it matters: Aria is a powerful tool for developers looking to integrate AI capabilities into their projects without the need for external dependencies. ### Repository: lef-fan/aria ### GitHub stars: 135 ### GitHub description: A local and uncensored AI entity.

Overview: CUREBench, a benchmarking project for AI reasoning in therapeutic decision-making, is designed to evaluate the performance of models in a realistic clinical setting. The project aims to provide a comprehensive and challenging test for AI systems in healthcare, focusing on their ability to make informed decisions based on medical data. Key features: - Access to a large-scale private dataset for both validation and testing. - A detailed leaderboard to track the performance of different models. - Support for various AI models, including both zero-shot and fine-tuned ones. - Easy-to-use command-line interface for quick evaluation and submission. Tech stack: Python, utilizing libraries such as transformers for model inference and evaluation. Why it matters: CUREBench is crucial for advancing the field of AI in healthcare, as it provides a realistic and challenging environment for evaluating AI models' capabilities in medical decision-making. The project's focus on therapeutic decision-making ensures that the models' performance is directly applicable to real-world clinical scenarios, making it an essential resource for researchers and developers in the field. Repository: mims-harvard/CUREBench GitHub stars: 136 GitHub description: CUREBench @ NeurIPS 2025: Benchmarking AI reasoning for therapeutic decision-making at scale.

### Overview: BanglaSpeech2Text is an open-source offline speech-to-text package designed for the Bangla language. It fine-tunes the latest Whisper model for optimal performance, enabling accurate transcription without internet connectivity. ### Key features: - Fine-tuned on the latest Whisper model for Bangla language - Supports multiple audio formats including mp3, mp4, mpeg, mpga, m4a, wav, webm, and more - Automatically converts audio files to the required format - Compatible with SpeechRecognition package for microphone-based transcription - Available in Python 3.7 or higher ### Tech stack: Python ### Why it matters: BanglaSpeech2Text offers a reliable and efficient solution for Bangla language speech-to-text transcription, making it ideal for applications that require offline processing and accuracy.

### Overview: Audiomate is a Python library designed to handle audio datasets efficiently. It provides a generic interface for accessing and loading various audio datasets, making it easier to integrate audio data into machine learning projects. ### Key Features: - **Data Access**: Provides data structures for accessing and loading different audio datasets. - **Corpus Management**: Facilitates the management of audio datasets, including validation, splitting, subsets, merging, and filtering. - **Feature Extraction**: Offers tools for extracting features from audio samples. - **Model Training**: Enables feeding samples for training machine learning models. ### Tech Stack: - Python ### Why It Matters: Audiomate simplifies the process of working with audio datasets, making it easier to integrate audio data into machine learning projects. It is particularly useful for tasks requiring the use of audio datasets, such as speech recognition or acoustic event detection. ### Repository: - Repository: ynop/audiomate - GitHub Stars: 138 - GitHub Description: Python library for handling audio datasets. ### Documentation: - [Documentation](https://audiomate.readthedocs.io) - [Examples](https://github.com/ynop/audiomate/tree/master/examples) - [Changelog](https://audiomate.readthedocs.io/en/latest/notes/changelog.html) ### Supported Datasets: - **Acoustic Event Dataset** - **AudioMNIST** - **Mozilla Common Voice** - **ESC-50** - **Fluent Speech Commands** - **Free Spoken Digit Dataset** - **German Distant

### Overview: This repository provides a tutorial for finetuning LLaMA-7B with Chinese instruction datasets. The datasets are combined using English-instruction and Chinese-output construction, making it suitable for complex NLP tasks such as summarization, question answering, and text generation. The repository focuses on low GPU consumption methods, making it ideal for those with limited GPU resources. ### Key features: - Combined Chinese instruction and English output datasets - Supports finetuning with LLaMA-7B and BLOOM models - Includes finetuning scripts for both LLaMA and BLOOM models - Provides detailed instructions on how to install and use the repository ### Tech stack: Python ### Why it matters: This repository is valuable for developers looking to fine-tune large language models with Chinese datasets, making it easier to work with limited GPU resources. The combination of English-instruction and Chinese-output datasets allows for more complex NLP tasks, such as summarization and question answering.

Overview: Evolve is a self-improving agent development tool that leverages on-the-job learning to enhance AI agent performance. It transforms past experiences into reusable guidelines, improving future task completion without altering model weights. Key features: - Guideline consolidation and clustering - Memory-based guideline storage - Efficient learning with high model coverage - Consistency improvement for repeated runs Tech stack: Python Why it matters: Evolve enhances AI agents' performance by providing robust, memory-based guidelines that can be easily integrated into existing agent assistants, such as Claude Code and Codex. This tool is particularly valuable for improving task completion rates and ensuring consistent performance across multiple runs. Repository: AgentToolkit/altk-evolve Primary language: Python GitHub stars: 135