Projects

### Overview: agent-manager is a lightweight terminal application that seamlessly integrates with your existing AI coding tools, providing a unified workflow for developers working with various AI agents. ### Key Features: - **Live Status**: Real-time monitoring of your AI agent's performance. - **Quick Prompts**: Instant access to prompts for various AI tools. - **Worktrees**: Efficient management of multiple sessions and projects. - **Diff Review**: Review changes and diffs directly within the terminal. ### Tech Stack: Go ### Why It Matters: agent-manager simplifies the development process by providing a unified interface for managing and interacting with multiple AI coding tools. This allows developers to focus on their work without worrying about the underlying infrastructure, making the workflow more efficient and enjoyable.

Overview: This project is a Python-based tool designed to generate bulleted notes summaries of books and other long texts, leveraging metadata available in epub and pdf files. It optimizes for the effective use of large language models by utilizing 2000 token chunks, aiming to maximize reasoning performance. Key features: - Automatically extracts chapters from most books - Splits chapters into 2000 token chunks - Supports epub and pdf files with ToC metadata - Compatible with Python 3.11.9 Tech stack: Python 3.11.9 Why it matters: The project addresses the need for efficient summarization of long texts, particularly in the context of large language models, by leveraging the optimal token length of 2000 tokens, which aligns with findings from research on reasoning capacity. Repository: [cognitivetech/ollama-ebook-summary](https://github.com/cognitivetech/ollama-ebook-summary) GitHub stars: 623 GitHub description: LLM for Long Text Summary (Comprehensive Bulleted Notes)

### Overview: Chandan, a Senior Researcher at Microsoft Research, shares his notes and research overviews on AI, machine learning, and statistics. These resources are useful for anyone interested in these fields. ### Key features: - **Slides**: Source for presentations, including ML slides from Berkeley CS 189 and AI slides from CS 188. - **Research Overviews**: Overviews and summaries of recent papers in various research areas. - **Notes**: Markdown notes and cheat sheets for courses and areas such as computational neuroscience, causal inference, classification, linear algebra, and more. - **Blog Posts**: Posts on advancements in machine learning, statistics, and neuroscience. ### Tech stack: HTML ### Why it matters: Chandan's resources are valuable for students, researchers, and anyone interested in the latest developments in AI, machine learning, and statistics.

Overview: Dreamer is a Python implementation of the Dreamer agent designed for learning behaviors in Atari and DMControl environments. It efficiently learns long-horizon policies by predicting ahead in a compact feature space. Key features: - Efficient learning of long-horizon policies - Compact feature space for fast prediction - Support for both Atari and DMControl environments - Code available on [GitHub](https://github.com/danijar/dreamer) Tech stack: Python Why it matters: Dreamer offers a fast and simple solution for learning complex behaviors in reinforcement learning environments, making it valuable for researchers and practitioners looking to improve agent performance in challenging tasks. GitHub stars: 625 Repository: danijar/dreamer

### Overview: LLM Hub is an open-source mobile app designed for on-device language model (LLM) chat, image generation, video generation, and music generation. It supports multiple model formats and runs powerful models locally, making it ideal for private and optimized mobile usage. ### Key Features: - **💬 Chat**: Multi-turn conversations with RAG memory, web search, TTS auto-readout, and multimodal input. - **🤖 AI Agent**: Autonomous device agent with function calling, interactive maps, device tools, MCP tools, and Termux terminal command generation. - **🤖 creAItor**: Design custom AI personas with specialized system prompts. - **💻 Vibe Coder**: Build apps with live HTML/JS preview; Android also integrates Codex through Termux using selected local model. - **✍️ Writing Aid**: Summarize, expand, rewrite, improve grammar, or generate code from descriptions. - **🎨 Image Generator**: Create images from text prompts using Stable Diffusion 1.5 with swipeable gallery. - **🎵 Music Generator**: Generate music and sound effects locally with SoundGen on Android and Magenta Realtime 2 on iOS, including live playback. - **🔎 Instant Media Search**: Search photos and videos by description or a reference image. - **🎞️ Video Moment Finder**: Describe a moment in a video and jump to matching timestamps. - **🔊 Text to Speech & Voice Cloning**: Android supports Kokoro, multilingual Supertonic 3 with voice-profile import, and Pocket TTS with on-device voice cloning. - **🔍

Overview: This repository contains a TensorFlow implementation of the paper "Deep Image Matting," focusing on a specific image processing technique. Key features: - TensorFlow implementation - Supports training and inference - Includes a pretrained model Tech stack: Python Why it matters: This implementation provides a practical tool for image matting, which is useful in various applications such as image editing and enhancement.

### Overview: VisualThinker-R1-Zero is the first to observe an "aha moment" and increased response length in visual reasoning on a 2B non-SFT model. ### Key features: - **First to observe "aha moment" and increased response length** for multimodal reasoning on a non-SFT 2B model. - **Vision-centric tasks** benefit from improved reasoning capabilities. - **Emergent ability to rethink and correct mistakes** during self-reflection behavior. - **First to successfully produce the emergent "aha moment" and increased response length** for multimodal reasoning on just a non-SFT 2B model. ### Tech stack: Python ### Why it matters: This project showcases the potential of non-SFT models in achieving remarkable results in multimodal reasoning tasks, highlighting the importance of exploring new approaches in AI research.

### Overview: Synthesizer is a multi-purpose LLM framework designed for generating datasets and integrating with retrieval-Augmented Generation (RAG) pipelines. It supports various LLM providers and is optimized for data creation and training. ### Key Features: - **Custom Data Creation**: Generates datasets tailored to specific needs using LLMs from Anthropic, OpenAI, vLLM, and HuggingFace. - **Retrieval-Augmented Generation (RAG) on Demand**: Built-in interface for integrating with Agent Search API, enhancing the accuracy of generated data. - **RAG Pipeline Performance Evaluation**: Measures and evaluates the performance of RAG pipelines using the Agent Search API. ### Tech Stack: Python ### Why it Matters: Synthesizer simplifies the process of generating datasets and integrating with RAG pipelines, making it a valuable tool for researchers and developers looking to enhance their data creation and analysis capabilities.

Overview: Detecto is a Python package that simplifies the process of building computer vision models using PyTorch, allowing users to perform tasks like still image and video inference, transfer learning on custom datasets, and serialization of models to files. Key features: - Build fully-functioning computer vision models with just 5 lines of code. - Perform inference on still images and videos. - Transfer learning on custom datasets with 5 lines of code. - Serialize models to files for easy transfer between libraries. - Easy installation via pip or manual download of dependencies. Tech stack: Python, PyTorch. Why it matters: Detecto simplifies the complex process of computer vision model development, making it accessible to developers without extensive knowledge in deep learning or PyTorch. Repository: alankbi/detecto Primary language: Python GitHub stars: 623 GitHub description: Build fully-functioning computer vision models with PyTorch README excerpt: 

### Overview: AI00 RWKV Server is a fast, efficient, and easy-to-use inference API server for the RWKV language model based on the web-rwkv inference engine. It supports Vulkan parallel and concurrent batched inference, GPU acceleration, and is compatible with OpenAI's ChatGPT API interface. The server is open-source, commercially usable, and available under the MIT license. It can be used for various tasks such as chatbots, text generation, translation, and Q&A. ### Key features: - **Vulkan Acceleration**: Supports GPU acceleration without the need for CUDA, making it compatible with AMD cards and integrated graphics. - **Compact and Easy to Use**: No bulky runtime environments required, making it ready to use out of the box. - **OpenAI Compatible**: Compatible with OpenAI's ChatGPT API interface. - **Model Compatibility**: Supports multiple versions of RWKV, including V5, V6, and V7. - **Model Download**: Available for download from the Hugging Face repository. ### Tech stack: Rust 1.78.0 ### Why it matters: - **High Performance and Accuracy**: Based on the RWKV model, it offers high performance and accuracy. - **GPU Acceleration**: Supports Vulkan parallel and concurrent batched inference, providing GPU acceleration without the need for CUDA. - **Open Source and Commercially Usable**: Open-source and commercially usable under the MIT license, making it accessible and cost-effective. - **Versatile Applications**: Compatible with various tasks such as chatbots, text

### Overview: Design systems for 200 of the world's best apps, framework-neutral and compatible with SwiftUI, Jetpack Compose, and Expo. Hand your AI agent a `DESIGN.md` and it ships pixel-matched UI. ### Key Features: - **Framework-Neutral**: Design systems for any iOS app framework. - **SwiftUI, Jetpack Compose, Expo**: Compatible with SwiftUI, Jetpack Compose, and Expo. - **Production-Grade**: Hand one to your AI agent, ship pixel-matched UI. - **Four Flavors per App**: Each app folder contains four design system files: `DESIGN.md`, `DESIGN-swiftui.md`, `DESIGN-jetpack-compose.md`, and `DESIGN-expo.md`. ### Tech Stack: - **Framework-neutral**: Design systems for any iOS app framework. ### Why It Matters: - **Consistent UI**: Hand your AI coding agent a `DESIGN.md` and it stops guessing. Each file is a complete, reverse-engineered design system for a real app — exact hex values, full type ramps, component states, motion curves, and haptics — in plain markdown that Claude, Cursor, Codex, and Stitch read natively. No Figma exports. No JSON. No tooling. No setup. - **Framework-Specific**: Framework-neutral design systems for SwiftUI, Jetpack Compose, and Expo. - **Live Gallery**: Browse the live gallery to see the 200 apps, every screen, rendered free in your browser. - **Request

### Overview: SPVNAS is an efficient 3D architecture for deep learning that achieves state-of-the-art performance on SemanticKITTI, winning second and first place in the NuScenes LiDAR Semantic Segmentation and Panoptic Segmentation challenges, respectively. ### Key Features: - **Efficient 3D Convolution**: Utilizes Sparse Point-Voxel Convolution to process 3D data efficiently. - **Model Zoo**: Pretrained models for MinkowskiNets, SPVCNN, and SPVNAS are available in the model zoo. - **Performance on SemanticKITTI**: Achieves high mIoU scores on the SemanticKITTI leaderboard. ### Tech Stack: Python ### Why it Matters: SPVNAS is a groundbreaking approach that significantly reduces computational complexity while maintaining high accuracy, making it highly relevant for real-world applications in autonomous driving and robotics.

Overview: FBOW is an optimized version of the DBow2/DBow3 libraries, significantly improving the speed of Bag of Words creation and vocabulary loading. Key features: - Only depends on OpenCV - Supports any type of descriptors (binary and real) - Corrects bugs found in DBOW2/3 - Extremely fast bag of words creation using AVX, SSE, and MMX instructions - Very fast vocabulary loading Tech stack: C++ Why it matters: FBOW is significantly faster than the original DBow2/3 libraries, making it an essential tool for applications requiring high-speed processing of image data.

### Overview: SmoothNLP is an open-source NLP toolset designed for explainable inference. It focuses on creating knowledge graphs and handling various NLP tasks such as tokenization, part-of-speech tagging, named entity recognition, and more. ### Key Features: - **Knowledge Graph (KG) Extraction**: Supports the extraction of knowledge graphs using SmoothNLP's V0.4 and above. - **Named Entity Recognition (NER)**: Recognizes and classifies named entities like companies, people, and locations. - **Financial Entity Recognition**: Identifies financial entities in text. - **Tokenization**: Splits text into tokens. - **Postagging**: Assigns part-of-speech tags to words. - **Dependency Parsing**: Analyzes sentences to understand their grammatical structure. - **Multi-threading Support**: Facilitates parallel processing for better performance. - **Logging**: Provides detailed logs for debugging and monitoring. ### Tech Stack: Java ### Why it Matters: SmoothNLP is a valuable tool for developers looking to enhance their NLP projects with explainable inference capabilities. It simplifies the process of creating knowledge graphs and handling various NLP tasks, making it easier to build robust and interpretable models in Python and Java environments. ### Repository: SmoothNLP ### Primary Language: Java ### GitHub Stars: 622 ### GitHub Description: SmoothNLP is an open-source NLP toolset designed for explainable inference. It focuses on creating knowledge graphs and handling various NLP tasks such as tokenization, part-of-speech tagging, named entity

### Overview: MDA Open Spec is a Markdown superset designed for agent-facing documents. It compiles a single `.mda` source into multiple `.md` files, including `SKILL.md`, `AGENTS.md`, `MCP-SERVER.md`, and `CLAUDE.md`. The resulting artifacts are tamper-evident, validated with JSON Schema, and signed using Sigstore, ensuring reproducibility and trustworthiness. ### Key Features: - **One Source, Multiple Targets**: Compiles into `SKILL.md`, `AGENTS.md`, `MCP-SERVER.md`, and `CLAUDE.md` for various agent runtimes. - **Tamper-Evident**: Every artifact carries a reproducible content digest and signed artifacts with Sigstore signatures. - **Typed Dependency Graph**: Includes typed dependency relationships for better readability and maintainability. - **Footnote Relationships**: Supports footnote relationships for referencing external content. - **Sigstore Anchored Signatures**: Ensures the integrity and authenticity of the compiled artifacts. ### Tech Stack: TypeScript ### Why It Matters: MDA Open Spec simplifies the process of creating agent-facing documents, ensuring that the same content is consistently delivered across different runtimes. It addresses the challenges of maintaining multiple frontmatter shapes and provides a tamper-evident and signed solution, enhancing the trustworthiness and reliability of agent skills.

### Overview: This repository uses PyTorch to implement several classic papers and models in the field of computer vision and deep learning. ### Key features: - **Classification on MNIST**: Implements two-layer CNN, Network in Network, and ResNet18. - **Generative Models**: Includes DCGAN, pix2pix, DiscoGAN, CycleGAN, BEGAN, and VAE. - **Neural Style Transfer**: Features Neural Style Transfer, preserving color in neural art style transfer, neural style transfer by aligning BN statistics, and fast neural style transfer. - **LSTM**: Implements a Char RNN model. ### Tech stack: Python ### Why it matters: This repository serves as a practical guide for developers to understand and implement classic models and frameworks using PyTorch, making it valuable for both learning and practical applications in computer vision and deep learning.

Overview: OpenHarness is a code-first, composable SDK designed to build powerful AI agents using TypeScript. It leverages Vercel's AI SDK and is inspired by Claude Code, Codex, and similar harnesses. Key features: - Agent - Session - Conversation - Middleware - Virtual Filesystem Provider - React and Vue 3 Composables - Multi-Turn with Sessions - Composable Middleware - Subagents with Dynamic Subagent Catalogs and Resumable Sessions Tech stack: TypeScript Why it matters: OpenHarness simplifies the process of building AI agents by providing a composable SDK that allows developers to create powerful, stateful AI agents using TypeScript. It streamlines the development process by abstracting away the complexities of integrating with various AI models and providers, making it easier for developers to focus on the core functionality of their applications. Repository: MaxGfeller/open-harness GitHub stars: 619 GitHub description: A code-first, composable SDK to build powerful AI agents based on Vercel's AI SDK and inspired by Claude Code, Codex, and similar harnesses.

### Overview: Nubase is an open-source, AI-native backend platform designed specifically for modern product teams. It combines memory, database, storage, and authentication functionalities into a self-hostable service, enabling developers to turn AI-generated code into real apps with minimal setup. ### Key Features: - **Memory**: High-performance in-memory storage for caching and session management. - **Database**: Supports PostgreSQL and other relational databases for efficient data storage. - **Storage**: Object storage for scalable data management. - **Auth**: Secure authentication and authorization mechanisms for secure access. - **Functions**: Deploy backend logic using edge functions for microservices architecture. - **Storage**: Persistent storage for storing generated frontend assets. - **AI Gateway**: Facilitates communication between AI-generated code and the frontend. - **Memory**: Durable memory for storing data and state. - **Cron**: Schedule recurring tasks and jobs. ### Tech Stack: - **Java**: The primary language used for developing and deploying Nubase. - **Docker**: Used for containerization and deployment. ### Why It Matters: Nubase simplifies the process of turning AI-generated code into real applications, making it easier for developers to build modern, scalable, and secure products. Its self-hostable nature ensures that teams can quickly get started without the need for separate hosting accounts, making it ideal for small to medium-sized teams.

### Overview: TEN is an open-source framework designed for conversational voice AI agents. It includes key components such as VAD (Voice Activity Detection), Turn Detection, and Portal, making it a comprehensive solution for full-duplex dialogue communication. ### Key features: - **Turn Detection**: Identifies and classifies turns in dialogue sequences, crucial for understanding conversational flow. - **VAD**: Detects and segments speech activity in audio recordings, essential for processing dialogue. - **Portal**: Provides a unified interface for managing and accessing multiple components within the framework. ### Tech stack: Python ### Why it matters: TEN is a valuable tool for developers looking to enhance their conversational AI projects, offering robust solutions for dialogue analysis and turn detection in full-duplex communication. Its open-source nature ensures that it can be easily integrated into various projects and contributes to the broader development of conversational AI technology. ### Repository: TEN-framework/ten-turn-detection ### Primary language: Python ### GitHub stars: 620 ### GitHub description: Turn detection for full-duplex dialogue communication

Overview: This project demonstrates how to self-host n8n on Google Cloud Run without subscription fees or server headaches, offering complete control over your automation workflows and data. Key features: - Deploy n8n on Google Cloud Run - Use Cloud SQL PostgreSQL for database persistence - Connect to Google services via OAuth - Implement queue mode deployment for production scaling Tech stack: Google Cloud Run, Cloud SQL PostgreSQL, Google Auth Platform Why it matters: This setup provides complete control over your automation workflows and data, eliminating the need for monthly subscription fees and server management headaches. Repository: datawranglerai/self-host-n8n-on-gcr Primary language: HCL GitHub stars: 619 GitHub description: Self-host n8n on Google Cloud without the subscription fees or server headaches - because your automation workflows shouldn't cost more than your coffee budget