Projects

Overview: Facemoji's mocap4face is a cross-platform SDK designed for facial motion capture, generating blendshapes and rigid head poses in real-time from photos or videos. Key features: Real-time facial motion capture, blendshapes, and rigid head poses. Tech stack: C++ and OpenCV. Why it matters: Facemoji's mocap4face is essential for developers looking to integrate facial motion capture into their projects, enabling realistic and dynamic animations. Repository: facemoji/mocap4face Primary language: Unknown GitHub stars: 614

Overview: Harvester is a Python library designed to simplify the image acquisition process in computer vision applications, making it easy to gather and manipulate images using the GenICam standard. Key features: - Image acquisition through GenTL Producers - Multiple loading of GenTL Producers in a single Python script - GenICam feature node manipulation Tech stack: Python Why it matters: Harvester streamlines the image acquisition process, making it accessible and efficient for developers working on computer vision projects. It supports various GenTL Producers and allows for easy manipulation of GenICam compliant entities, enhancing flexibility and ease of use.

Overview: This repository is a minimal, Swift implementation of the FlashAttention algorithm, ported to Apple silicon, focusing on the core bottlenecks of attention algorithms. Key features: - Minimal source files - JIT compiled at runtime - Single-headed attention only - Uses less memory than the official implementation - Achieves 100% parallelization efficiency - Consistent performance on M1 Max (4400 gigainstructions per second) Tech stack: Swift Why it matters: This implementation provides a simpler, more maintainable solution compared to the original FlashAttention algorithm, focusing on the key performance bottlenecks and achieving consistent results on modern Apple silicon hardware. Repository: philipturner/metal-flash-attention GitHub stars: 614

Overview: Real-time audio-to-text transcription using the faster-whisper library. Key features: - Transcription speed of less than a second - Support for multiple audio formats (MP3, OGG, WAV) - Integration with the OpenAI API for text proofreading - Configurable settings for faster-whisper models and VAD settings Tech stack: Python, Sounddevice, faster-whisper, Silero VAD Why it matters: This project offers a high-speed, accurate real-time transcription solution for developers, making it useful in applications requiring real-time text processing, such as live streaming platforms, voice assistants, and more.

### Overview: Voice input for Fcitx5 — local and cloud ASR, LLM rewriting, cross-distro packages ### Key features: - **Two trigger modes** — tap to toggle recording, or hold to push-to-talk - **Local & cloud ASR** — offline [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) models or cloud providers (Doubao, Aliyun Bailian, ElevenLabs, OpenAI-compatible), switchable at runtime from the command palette (`Shift_R` → `/asr`) - **LLM post-processing** — error correction, formatting, translation via scenes - **Command mode** — select text, speak an instruction, release to apply - **GUI & CLI** — `vinput-gui` for quick setup, `vinput` CLI for full control - **Cross-distro** — Arch, Fedora, Ubuntu/Debian, Nix, Flatpak ### Tech stack: C++ ### Why it matters: Improves productivity and accessibility by enabling voice input across various platforms and languages.

Overview: LightReasoner is a groundbreaking project that explores the potential of small language models (SMALL) to teach large language models (LARGE) reasoning. This project aims to demonstrate that strategic token selection, rather than exhaustive training, can unlock the latent potential of LLMs, leading to superior performance with remarkable token efficiency. Key features: - **Efficient Token Selection**: LightReasoner leverages smart token selection to achieve consistent improvements in zero-shot pass@1 accuracy while significantly reducing computational overhead. - **Pre-Collected Samples**: LightReasoner samples are pre-collected for immediate fine-tuning, making it easier to reproduce and accelerate downstream research workflows. - **Enhanced Models**: LightReasoner Enhanced Models are available on Hugging Face Hub, fine-tuned with the project's efficient reasoning enhancement approach for easy deployment and experimentation. Tech stack: Python Why it matters: This project showcases that smarter, not blindly harder, is the key to scalable AI improvement, demonstrating the potential of small language models to enhance large language models' reasoning capabilities.

Overview: `multilspy` is a Python library designed to simplify the creation of language server clients, allowing developers to easily integrate static analysis tools into their applications. Key features: - Automatically handles platform-specific server binaries and language server setup/teardown - Supports JSON-RPC communication between the client and server - Manages and passes hand-tuned server and language-specific configuration parameters - Provides a simple API for the user, executing all steps required by the server-specific protocol Tech stack: Python Why it matters: `multilspy` simplifies the integration of static analysis tools into applications, making it easier to build language server clients that can provide useful information to developers.

Overview: This project is a TensorFlow implementation of a Convolutional Neural Network (SRCNN) for image super-resolution, based on the original research from the University of Hong Kong. Key features: - TensorFlow implementation - Supports training and testing - Compatible with Scipy for image processing Tech stack: Python Why it matters: This project provides a practical and efficient solution for image super-resolution using deep convolutional networks, demonstrating the effectiveness of SRCNN in real-world applications.

Overview: Facetorch is a Python library designed for facial detection and analysis, utilizing PyTorch for deep neural network capabilities. It aims to provide open-source face analysis models, packaging them as portable `torch.export` models for efficient inference. Key features: - Customizable configuration using Hydra and OmegaConf - Reproducible environments with uv, conda-lock, and Docker - Portable models serialized as `.pt2` format - Governed extensibility with immutable manifest entries and model rights approval - Deterministic input handling, accepting local paths, tensors, NumPy arrays, PIL images, and bytes Tech stack: Python, PyTorch, Hydra, OmegaConf, uv, conda-lock, Docker Why it matters: Facetorch offers a user-friendly and efficient solution for facial analysis tasks, suitable for developers and researchers looking for flexibility and performance.

Overview: 中文Mixtral混合专家大模型（Chinese Mixtral MoE LLMs）是基于Mistral.ai发布的Mixtral模型开发的，该模型使用了稀疏混合专家模型（Sparse MoE）架构。通过大规模中文无标注数据进行中文增量训练，得到了中文Mixtral基础模型，并进一步通过指令精调，得到了中文Mixtral-Instruct指令模型。该模型原生支持32K上下文（实测可达128K），在处理长文本和数学推理、代码生成等方面表现显著。 Key features: - 中文Mixtral基础模型 - 中文Mixtral-Instruct指令模型 - 预训练脚本和指令精调脚本 - 快速本地量化和部署教程 - 支持LLaMA.cpp、LLaMA、LangChain、PrivateGPT、vLLM等Mixtral生态 Tech stack: Python Why it matters: 该模型在处理长文本和数学推理、代码生成等方面表现显著，支持多种语言模型生态，适用于多种应用场景，是开发和研究中文语言模型的有力工具。

Overview: This repository contains open-source SEO skills designed to enhance the performance and SEO optimization of Claude Code, Cursor, Codex, and any AI agent that reads `SKILL.md` files. It includes daily SEO checks, title tag optimization, blog refreshes, backlink management, and more. Key features: - SEO autopilot skill - Blitz SEO skill - SEO title optimization skill - Blog refresh skill - Linkdr skill - Intro skill - Scribiz skill - Customer journey skill - Diff-marketing skill - Copywriting blitz skill - Understand customers skill Tech stack: Node.js Why it matters: This tool streamlines the SEO optimization process, ensuring that AI agents can maintain and improve their performance effectively, making them more efficient and effective in their tasks.

### Overview: HEX is a local-first voice dictation tool designed for macOS, built with Rust. It allows you to dictate directly into any app on your Mac, without requiring cloud services or external models. The app is available now and includes features like on-device transcription, customizable shortcuts, and optional voice tools. ### Key Features: - **On-Device Transcription**: Download a local speech model during setup; ordinary dictation does not require an external service. - **Customizable Shortcuts**: Choose a shortcut, language, and model. Modes for corrections, application-specific behavior, and optional text processing. - **Voice Tools**: Enable Voice Action or voice commands when needed. Both start off, with provider-backed rewriting separate from local speech recognition. ### Tech Stack: Rust ### Why It Matters: HEX is a powerful and efficient solution for voice dictation, designed for macOS users who prefer local processing and minimal dependency on external services. Its on-device transcription feature ensures that all data is kept locally, making it ideal for privacy-conscious users.

### Overview: AgentChain is a powerful tool for orchestrating large language models (LLMs) and multiple agents to tackle complex tasks. It leverages state-of-the-art LLMs for natural language processing and multimodal capabilities to handle various inputs and outputs. ### Key Features: - **🧠 LLMs as the brain:** AgentChain uses state-of-the-art LLMs to plan and make decisions based on natural language inputs. - **🌟 Fully Multimodal IO:** It accepts and outputs from various modalities such as text, image, audio, and tabular data. - **🤝 Orchestrate Versatile Agents:** AgentChain can orchestrate multiple agents to perform complex tasks, using composability and hierarchical structuring of tools. - **🔧 Customizable for Ad-hoc Needs:** It can be customized to fit specific project requirements, enhancing capabilities with new agents and a distributed architecture. ### Tech Stack: AgentChain is built using Python, leveraging LLMs and multimodal capabilities. ### Why it Matters: AgentChain is a versatile tool for a wide range of applications, including task execution with natural language instructions, data understanding and generation, and complex combination of tools. It is designed to be highly customizable and adaptable to various project needs. ### Repository: The repository for AgentChain is hosted on GitHub at <https://github.com/jina-ai/agentchain>. It has been starred by 612 developers, indicating its popularity and utility in the field of AI and machine learning.

### Overview: Sharing Updatable Models (SUM) on Blockchain is a framework designed to host and train publicly available machine learning models. The project aims to make model predictions free, ensuring data validation through a series of steps on the blockchain, and allowing anyone to query the model for predictions for free. ### Key Features: - **Blockchain Integration**: Models can be stored and updated on the blockchain, ensuring data accessibility and security. - **Public Model Sharing**: Models can be hosted on the blockchain, making them accessible to anyone. - **Incentive Mechanisms**: Users can validate requests to add data, and the system can be incentivized through deposits or virtual points. - **Model Updates**: The model can be updated using predefined training algorithms, and anyone can query the model for predictions for free. ### Tech Stack: - **Blockchain**: Ethereum blockchain for storing and updating models. - **Smart Contracts**: Written in Solidity for implementing the incentive mechanisms and data handling. - **Python**: Used for simulation tools and demonstration purposes. ### Why It Matters: The project aims to democratize access to machine learning models by making them publicly available on the blockchain, ensuring that users can train their own models based on the shared data without the need for complex computations. ### Repository: - **GitHub Repository**: microsoft/0xDeCA10B - **Stars**: 612 - **Description**: Sharing Updatable Models (SUM) on Blockchain (formerly Decentralized & Collaborative AI on Blockchain) ### Security: - **Build Status**: [![Build Status](https://dev.azure.com

### Overview: Nomi is an open-source, local-first AI video studio that allows users to bring any model or their own ComfyUI to create videos. It enables direct interaction with Claude Code, Codex, or Cursor, generating keyframes and placing them on a timeline. The system is designed for a local-first approach, keeping projects, prompts, and keys on the user's machine, eliminating the need for an account or telemetry. ### Key Features: - **Local-first Workflow**: Projects, prompts, and keys are stored locally on the user's machine. - **AI Agent**: The agent reads the canvas, splits the story into shots, generates keyframes, and arranges them on a timeline. - **Direct Interaction**: Users can interact with Claude Code, Codex, or Cursor, and the system generates keyframes and places them on the timeline. - **Storyboard**: A storyboard feature is available to plan the shots and reference cards for characters and locations. - **Auto-Editing**: Users can choose to edit each step, auto-edit, or fully auto-generate the video. ### Tech Stack: The backend is built using TypeScript, leveraging the power of the AI agent to generate keyframes and place them on the timeline. ### Why It Matters: Nomi offers a pro-grade workflow on top of users' own models, allowing them to pay their provider directly at their price. The system is designed for a local-first approach, eliminating the need for an account or telemetry. This makes it an attractive option for those looking to create videos on a budget, without the hassle of managing an

### Overview: Harness Score is an AI coding agent reliability measurement tool that provides a comprehensive assessment of a harness's maturity level. It measures the context, skills, hooks, sensors, and feedback mechanisms of an AI coding tool, offering a detailed 108-point breakdown and a precise list of what needs to be fixed. ### Key Features: - **Maturity Level (L0–L4)**: Measures the overall reliability of the harness. - **108-point Breakdown**: Six dimensions of the harness, including context, skills, hooks, sensors, and feedback. - **Detailed Diagnosis**: Identifies and explains the missing components, providing a recipe for fixing them. - **No LLM Calls or Network Access**: Provides a consistent and accurate score every time. - **CI Integration**: Measures the pipeline's hygiene and safety, ensuring the code is clean and secure. ### Tech Stack: TypeScript ### Why It Matters: Harness Score provides a reliable and consistent way to measure the reliability of an AI coding harness, ensuring that the agent's performance is up to par. It helps developers gate their pipelines on a standardized metric, improving overall code quality and reducing the risk of deployment failures. ### Repository: paladini/harness-score GitHub Stars: 609 ### Primary Language: TypeScript ### GitHub Description: Your AI coding agent is only as reliable as the harness around it. Measure that harness in seconds with Harness Score. Know exactly what to fix. Watch the number go up. ### Why This Exists: Vibe-coding without a harness works, right up

Overview: Tock is an open-source conversational AI toolkit designed for building Bots using Generative AI, Natural Language Processing, and Kotlin. Key features: - Supports OpenAI, Azure OpenAI, AWS Bedrock, and Ollama models for RAG answers. - Compatible with OpenNLP, Rasa, Amazon SageMaker, and other NLP stacks. - Kotlin-based user interface for building stories, configuring generative AI, and analyzing conversations. - Built-in connectors for various channels like Web, WhatsApp, Messenger, Teams, Slack, Google Chat, and more. - Provides toolkits for custom Web integration with React and Vue. - Deployed in the Cloud or On-Premise with Docker or Kubernetes (Helm chart). Tech stack: Kotlin. Why it matters: Tock offers a comprehensive solution for building conversational AI bots, leveraging Generative AI and NLP to create engaging and context-aware interactions. Repository: theopenconversationkit/tock GitHub stars: 612 GitHub description: Tock, the open source conversational AI toolkit.

Overview: kani (カニ) is a lightweight and highly hackable framework designed for chat-based language models, offering fine-grained control over the parts of the control flow that matter. Key features: - Model support for OpenAI, Anthropic, Google AI, Hugging Face transformers, llama.cpp, vLLM, and more. - Model-agnostic framework for adding support for additional models. - Model zoo example for loading popular models. - Documentation available on ReadTheDocs and arXiv. - Community extensions for extending functionality. Tech stack: Python Why it matters: kani provides a flexible and customizable environment for NLP researchers, hobbyists, and developers, making it suitable for a wide range of applications in natural language processing. Repository: zhudotexe/kani Primary language: Python GitHub stars: 610 GitHub description: kani (カニ) is a lightweight and highly hackable framework for chat-based language models with **tool usage/function calling**.

Overview: RAI is a vendor-agnostic AI framework designed for developing and deploying Embodied AI features in physical robots, utilizing ROS 2 tools for complex actions, scenarios, and interface execution. Key features: - Seamlessly integrate Gen AI capabilities into robots - Enhance robotic systems with smart features - Support diverse interaction capabilities Tech stack: Python Why it matters: RAI simplifies the integration of AI into physical robotics, making it accessible for developers across different vendors. Repository: RobotecAI/rai Primary language: Python GitHub stars: 610 GitHub description: RAI is a flexible AI agent framework for developing and deploying Embodied AI features in physical robots. It utilizes ROS 2 tools for complex actions, scenarios, and interface execution, and supports diverse interaction capabilities.

**Overview:** Kwipu is a powerful tool that turns your Markdown notes into a fully local Graph RAG engine. It extracts semantic and structural relations from wikilinks and YAML frontmatter, allowing you to ask questions directly from a terminal, web interface, or MCP client. **Key features:** - **Graph RAG Engine:** Kwipu builds a graph from Markdown files to provide a searchable and interactive interface. - **Semantic Relations:** Extracts semantic relations using an LLM. - **Structural Relations:** Extracts structural relations from wikilinks and YAML frontmatter. - **Hybrid Search:** Combines vector similarity, BM25, temporal metadata, and optional synonym retrieval. - **Multilingual Support:** Supports English, Italian, French, German, Spanish, and Portuguese patterns. - **Local Execution:** Runs on local Ollama, ensuring no cloud usage. - **Obsidian Compatibility:** Works seamlessly with Obsidian-style vaults. - **Terminal Interface:** Direct interaction from a terminal. - **Web Interface:** Explore and query the graph from a browser. - **MCP Client Integration:** Connects to an MCP server for enhanced functionality. - **Configuration and Deployment:** Configurable and deployable for various environments. **Tech stack:** Python **Why it matters:** Kwipu simplifies knowledge retrieval by turning your notes into a searchable, interactive graph. It leverages LLMs for semantic understanding and structural relations, providing a powerful tool for anyone looking to enhance their note-taking and knowledge management practices. **Repository:** benmaster82/Kwip