Projects

### Overview: Meka Agent is an open-source, autonomous computer-using agent that excels in state-of-the-art web browsing capabilities. It leverages vision and a full computer context to perform tasks efficiently, achieving remarkable performance in benchmark tests. ### Key Features: - **State-of-the-Art Browsing**: Performs web browsing tasks with unmatched performance. - **Extensible Framework**: Allows flexibility in choosing models, tools, and infrastructure providers. - **Benchmark Results**: Achieves 72.7% performance in the WebArena Benchmark. ### Tech Stack: TypeScript ### Why It Matters: Meka Agent is a powerful tool for automating web browsing tasks, offering superior performance and flexibility compared to traditional methods. Its ability to integrate with various models and infrastructure providers makes it a versatile solution for developers and businesses looking to enhance their web automation capabilities. ### Repository: Repository: trymeka/agent GitHub Stars: 362 GitHub Description: state of the art browsing agent (WebArena 72.7%) ### Getting Started: To get started with Meka Agent, follow these steps: 1. Install the main components of the SDK: ```bash npm install @trymeka/core @trymeka/ai-provider-vercel @ai-sdk/openai @trymeka/computer-provider-anchor-browser playwright-core ``` 2. Create your `.env` file and enter your API keys: ```bash OPENAI_API_KEY=GET FROM https://platform.openai.com/settings/organization/api-keys ANCHOR_BROWSER_API_KEY=

Overview: Seedance 2.0, Seedance 2.5, and Seedance 2 Mini APIs are Python wrappers for ByteDance's advanced AI video generation models. These models allow for the creation of cinematic, high-fidelity videos from text prompts and static images, with realistic human face generation. The Seedance 2.0 API offers the highest quality, while Seedance 2.5 and Seedance 2 Mini provide faster, more affordable generation at a fraction of the cost. Key features: - Seedance 2.0: Highest quality video generation - Seedance 2.5: Native 4K, 30-second clips, native audio, 30 images/10 videos/10 audio references - Seedance 2 Mini: Fast, affordable generation at a fraction of the cost Tech stack: Python Why it matters: These APIs enable developers to create professional-grade AI videos quickly and efficiently, making them ideal for various applications such as marketing, education, and more. Repository: Anil-matcha/Seedance-2-API Primary language: Python GitHub stars: 360 GitHub description: Python wrapper for Seedance 2.0, Seedance 2.5, and Seedance 2 Mini API — Text-to-Video, Image-to-Video, realistic human faces, 1080p, consistent character generation. README excerpt: The most comprehensive Python wrapper for the Seedance 2.0 API, Seedance 2.5 API, and Seedance 2

### Overview: This repository is a comprehensive survey of scientific large language models, focusing on biological and chemical domains. ### Key features: - **Textual Scientific Large Language Models (Text-Sci-LLMs)**: Covers models for textual scientific languages. - **Molecular Large Language Models (Mol-LLMs)**: Includes models for predicting molecule properties, interaction predictions, and molecule generation/design/edit. - **Protein Large Language Models (Prot-LLMs)**: Focuses on protein sequence representation, sequence generation/design, and prediction of variants and evolution. - **Genomic Large Language Models (Gene-LLMs)**: Covers models for general genomic predictions, function prediction, variants and evolution prediction, DNA-protein interaction prediction, and RNA prediction. - **Multimodal Scientific Large Language Models (MM-Sci-LLMs)**: Combines molecular and textual information for comprehensive predictions. ### Tech stack: Python, Jupyter Notebook, and various scientific libraries. ### Why it matters: This survey provides a detailed overview of the current state of scientific large language models, highlighting their applications in biological and chemical domains.

Overview: This project aims to develop and train Universal Transformer Memory Models, focusing on creating universally applicable memory systems for transformers. These models are designed to enhance the capabilities of transformers by incorporating memory mechanisms. Key features: - Training and evaluation of Neural Attention Memory Models (NAMMs) - Support for LongBench and ChouBun tasks - Installation and usage through conda environment management - Replication of training and evaluation stages - Use of wandb for logging results - Support for gated models via hugging face hub authentication Tech stack: Python Why it matters: This project is significant as it aims to enhance the memory capabilities of transformers, potentially improving their performance and adaptability across various tasks and datasets. Repository: SakanaAI/evo-memory GitHub stars: 361 GitHub description: Code to train and evaluate Neural Attention Memory Models to obtain universally-applicable memory systems for transformers.

Overview: KVSplit is a project that optimizes the memory usage of larger context windows and heavy LLMs on Apple Silicon Macs by differentiating the precision of keys and values in the KV cache. This reduces memory usage by up to 72% with minimal quality loss and allows for 2-3x longer context lengths in the same memory budget. Key features: - Differentiated precision for keys and values in the KV cache - Optimized for Apple Silicon with Metal support - Comprehensive benchmarking suite with perplexity measurement - Memory usage and performance analysis tools - Easy setup and usage Tech stack: Python Why it matters: KVSplit enables developers to run larger context windows and heavier LLMs on Apple Silicon Macs with reduced memory usage, improved inference speed, and optimized performance, making it a valuable tool for AI and language model optimization on Apple Silicon hardware.

Overview: This project aims to reproduce the MUSE model for fast text-to-image generation, focusing on deduped datasets and using PyTorch and Hugging Face's Transformers library. Key features: - Reproduction of the MUSE model - Use of deduped LAION-2B + COYO-700M dataset - Support for class-conditional generation - Training of improved VQGAN models - Full model training on LAION + COYO datasets Tech stack: Python, PyTorch, Transformers, Hugging Face, Webdataset Why it matters: This project provides a simplified and scalable solution for reproducing and understanding the MUSE model, which can be useful for researchers and developers looking to enhance their text-to-image generation capabilities.

Overview: Ataru is a Rust-based search core designed to unify and retrieve local agent history, providing a stable and efficient way to access past conversations and context. Key features: - Unified indexing of Claude, Codex, and other local session sources. - Incremental indexing for new messages to keep the index up-to-date without scanning the entire history. - Comprehensive Chinese full-text search using Tantivy and Jieba. - Hybrid recall for natural language queries using semantic indexing when available. - Hierarchical aggregation of hits across different levels (Turn, Run, Session, Project). - Contextual re-read of stable Project/Session/Run/Turn identifiers, roles, times, and line numbers. - Offline execution; semantic recall is optional and falls back to keyword search on timeout. Why it matters: Ataru simplifies the process of accessing and reusing past conversations and context by providing a unified and searchable memory layer. It enables developers to quickly and efficiently retrieve the relevant information from their agent history, improving productivity and decision-making. Repository: lovstudio/Ataru GitHub stars: 361 Tech stack: Rust, Tauri, React, Apache 2.0 license Why it matters: Ataru aims to make past conversations and context accessible and reusable, enhancing productivity and decision-making in AI development.

Overview: This repository contains a curated collection of educational marimo notebooks, organized into courses covering fundamental concepts in Python, optimization, probability, linear algebra, and more. Key features: - 🐍 Python notebooks for learning the basics of Python programming - ⚖️ Optimization notebooks for solving optimization problems - 🎲 Probability notebooks - 📏 Linear algebra notebooks - ❄️ Polars notebooks - 🔥 Pytorch notebooks - 🗄️ Duckdb notebooks - 💜 Daft notebooks - 📈 Altair notebooks - 📈 Plotly notebooks - 📈 matplotlib notebooks Tech stack: Python Why it matters: This repository serves as a central repository for high-quality educational notebooks, making it easier for educators, students, and practitioners to learn and understand various topics in computer science, linear algebra, probability, and more.

### Overview This repository contains the source code and pretrained model for a webcam-based real-time image-to-image translation demo. The model uses deep learning techniques, specifically a deep convolutional conditional generative adversarial network autoencoder, to translate images from one style to another in real-time. ### Key Features - **Webcam Input**: Translates images from a live webcam feed in real-time. - **Pretrained Model**: Uses a pre-trained TensorFlow model for image translation. - **Training Data**: Scraped images from various art collections around the world, preprocessed, and used for training. ### Tech Stack - **Python**: The primary programming language used. - **TensorFlow**: The deep learning framework used for the model training and inference. ### Why It Matters This project showcases the versatility of deep learning models in real-time applications, demonstrating how they can be used to translate images from one style to another in a practical and interactive way. The use of a webcam input makes it accessible and engaging for developers and enthusiasts alike. ### Repository Details - **GitHub Stars**: 361 - **Primary Language**: Python - **GitHub Description**: Source code and pretrained model for running pix2pix in realtime on a webcam feed. - **README Excerpt**: This is the source code and pretrained model for the webcam pix2pix demo I posted recently on [twitter](https://twitter.com/memotv/status/858397873712623616) and vimeo. It uses deep learning, or to throw in

Overview: DoReMi is a PyTorch implementation designed to optimize data mixtures for language modeling datasets, addressing the uncertainty in how much of each domain to train on. It uses Distributionally Robust Optimization (DRO) to tune the data mixture, enabling more efficient training of larger models. Key features: - Domain-specific data mixture optimization - Dynamic upweighting or downweighting of domains based on proxy model's excess loss - Improved training efficiency and faster model performance - Fast, resumable dataloader with domain-level weighted sampling - Simple downstream evaluation harness - Integration with HuggingFace Trainer and FlashAttention2 Tech stack: Python, PyTorch, HuggingFace Transformers Why it matters: By optimizing the data mixture, DoReMi helps in training large language models more efficiently, allowing them to achieve better performance on a variety of downstream tasks without being overly biased towards any specific domain. This approach can lead to more robust and versatile models that perform well across different applications.

### Overview Vectorlite is a fast, SQL-powered vector search extension for SQLite that supports any language with a SQLite driver. It leverages the hnswlib library for approximate nearest neighbor search, providing efficient vector similarity queries. ### Key Features - **Fast Vector Search**: Utilizes hnswlib for approximate nearest neighbor search, offering significant performance improvements over other vector search solutions. - **SQL Interface**: Provides a SQL-like interface for vector queries, making it easy to integrate with existing SQL-based applications. - **Cross-Language Compatibility**: Works seamlessly with various languages that support SQLite, including Rust, Python, Node.js, and more. ### Tech Stack - **Main Implementation**: Rust - **Native C++**: Retained for hnswlib, SIMD operations, and thin C ABI shim - **Cargo**: Used for Rust implementation, manages SQLite virtual tables, scalar functions, parsers, and index policy - **CMake and scikit-build-core**: Invoked for normal source and wheel builds - **Public Package Names**: `vectorlite.so` / `vectorlite.dylib` / `vectorlite.dll` filenames remain unchanged - **Architecture and Contributor Guide**: Provides detailed information on the architecture and contributor guidelines ### Why It Matters - **Performance**: Offers faster vector search capabilities compared to other solutions like SQLite-Vec and SQLite-VSS. - **Cross-Language Support**: Works with multiple languages, making it highly versatile and useful in a variety of applications. - **Efficiency**: The SIMD-accelerated vector search is particularly beneficial for applications requiring high performance and low latency

Overview: ContinualAI is an open-source, collaborative wiki and repository for Continual Learning tutorials and demos running on Google Colaboratory. Key features: - Tutorials and demos on Continual Learning - Introduces various techniques such as Replay Strategy, Generative Replay, Progressive Neural Networks, and Genetic Pseudo Rehearsal - Central idea: Collaboratively create and import notebooks directly into Google Colab for easy access and experimentation Tech stack: Jupyter Notebook Why it matters: Continual Learning is crucial for developing AI systems that can learn continuously from new data without forgetting previously learned knowledge, making it essential for applications requiring long-term data processing and adaptation.

Overview: LiveWhisper is a nearly-live implementation of OpenAI's Whisper model, utilizing sounddevice for audio processing. It outputs sentence-by-sentence dictation from a microphone and can be used as an alternative to SpeechRecognition. Key features: - Real-time sentence-by-sentence transcription - Uses Whisper model for accurate speech recognition - Stores audio files if volume and frequency thresholds are met - Sends stored audio to Whisper for transcription Tech stack: Python, Whisper, numpy, scipy, sounddevice Why it matters: LiveWhisper provides a practical alternative to SpeechRecognition for real-time dictation, making it useful in various applications such as note-taking, translation, and more. Repository: Nikorasu/LiveWhisper Primary language: Python GitHub stars: 360 GitHub description: A nearly-live implementation of OpenAI's Whisper, using sounddevice. Requires existing Whisper install. README excerpt:

### Overview: CodeFuse-Query is a comprehensive data-centric static code analysis system designed to handle large-scale software development. It optimizes resource utilization, prioritizes data reusability, and introduces tasks types specifically for code changes, making it highly efficient and scalable. ### Key Features: - **Code Data Model (COREF):** Converts source code into data facts using Datalog and a unique two-tiered schema, COREF. - **Incremental Code Extraction:** Optimizes resource utilization by extracting code in an incremental manner. - **Task Types:** Specializes in tasks related to code changes, enhancing its domain-optimized design. - **Scanning Capabilities:** Supports scanning over 10 billion lines of code daily and more than 300 different tasks. ### Tech Stack: Java ### Why It Matters: CodeFuse-Query revolutionizes static code analysis by reimagining it as a data computation task, enabling efficient and scalable analysis of large codebases. Its domain-optimized design and logic-oriented facet make it highly effective for large-scale software development.

### Overview: Trajectory Consistency Distillation is an official repository for the paper "Trajectory Consistency Distillation," published on arXiv. The repository contains the source code, data, and documentation necessary for reproducing the results and understanding the methodology described in the paper. ### Key features: - **Code Repository**: Contains the implementation of the Trajectory Consistency Distillation algorithm. - **Data**: Includes the dataset used in the paper for training and evaluation. - **Documentation**: Provides detailed explanations of the model architecture, training process, and experimental results. ### Tech stack: Python ### Why it matters: This repository serves as a definitive source for the research presented in the paper, ensuring that the findings and methods can be replicated and verified by other researchers. It also highlights the importance of proper citation and acknowledgment of contributions in academic work, emphasizing the value of academic integrity.

### Overview: This project is a PyTorch implementation of diffusion models applied to SE(3) for generating 6D grasp pose distributions. The library is designed to learn 6D grasp distributions and can be used as a cost function for grasp and motion optimization problems. The project includes a detailed README that outlines installation steps, data preparation, and evaluation of trained models. ### Key Features: - PyTorch implementation of diffusion models in SE(3) - Training and sampling diffusion models for SE(3) - Applying learned distributions as cost functions for grasp optimization - Training and evaluation of models on various datasets including Shapenet and custom datasets - Tools for generating 6D grasp poses from point clouds and partial point clouds ### Tech Stack: Python with PyTorch ### Why it Matters: This project offers a novel approach to grasp pose generation by leveraging diffusion models on Lie groups, which can improve the efficiency and accuracy of grasp optimization algorithms. The ability to generate 6D grasp poses from point clouds makes it particularly useful in applications where precise control of grasps is crucial. ### Repository: The repository is hosted on GitHub at [robotgradient/grasp_diffusion](https://github.com/robotgradient/grasp_diffusion). It has been starred by 360 developers, indicating its popularity and utility in the field of robotics and machine learning. ### Installation: - Create a conda environment - Install the library using pip - Clone the repository and install dependencies - Download the dataset and trained models ### Sample Usage: -

### Overview: This project implements the Gibbs sampling algorithm for a Dirichlet Mixture Model (DMM) as described in the paper by Yin and Wang (2014). The GSDMM is designed for clustering short text documents efficiently and is known for its simplicity and effectiveness. ### Key Features: - **Gibbs Sampling Algorithm**: Implements the Gibbs sampling algorithm for the Dirichlet Mixture Model. - **Dirichlet Mixture Model (DMM)**: Clusters short text documents using a Dirichlet Mixture Model. - **Parameter Selection**: Requires an upper bound `K` on the number of clusters. - **Convergence**: The model converges quickly with good parameter selection. - **Space Efficiency**: Efficient space usage and scalability. ### Tech Stack: Python ### Why It Matters: The GSDMM is an effective and efficient algorithm for clustering short text documents, particularly useful for applications where space and computational resources are limited. Its simplicity and effectiveness make it a valuable tool in text analysis and clustering tasks. ### Repository: rwalk/gsdmm ### GitHub Stars: 360 ### Primary Language: Python ### GitHub Description: GSDMM: Short text clustering ### Usage: To use the Movie Group Process (MGP) for clustering short texts, initialize the MGP with the following parameters: ```python from gsdmm import MovieGroupProcess mgp = MovieGroupProcess(K=8, alpha=0.1, beta=0.1, n_iters=30) ``` Choose `K` to be

Overview: MagNet is a high-level Deep Learning API designed to simplify the process of building and training neural networks. It aims to reduce boilerplate code and make the creation of complex architectures more intuitive. Key features: - **Intelligent Nodes**: Nodes in MagNet are self-aware and can change their properties dynamically based on the computational graph, enabling the creation of complex architectures easily. - **Simple Workflow**: MagNet's API allows for a straightforward workflow, integrating data, defining models, training, debugging, testing, and deploying models. - **Extensibility**: Written on top of PyTorch, MagNet can mix with lower-level details, making it versatile and extensible. - **Ready for Any Experiment**: The Trainer accommodates any training logic, allowing for a wide range of experiments and research. Tech stack: MagNet is built on top of PyTorch. Why it matters: MagNet simplifies the process of building and training neural networks, making it easier to create complex architectures and experiments without the need for extensive boilerplate code.

Overview: HunyuanPortrait is a diffusion-based framework designed for generating lifelike, temporally consistent portrait animations. It decouples identity and motion using pre-trained encoders, enabling detailed and style-flexible animation from a single reference image. Key features: - Implicit Condition Control for Enhanced Portrait Animation - Decoupling Identity and Motion - Detailed and Style-Flexible Animation Tech stack: Python Why it matters: HunyuanPortrait offers a significant advancement in portrait animation by enabling detailed and style-flexible control over the animation, making it a valuable tool for artists and developers in the field of computer graphics and animation. Repository: [Tencent-Hunyuan/HunyuanPortrait](https://github.com/Tencent-Hunyuan/HunyuanPortrait) Primary language: Python GitHub stars: 359 GitHub description: [CVPR-2025] The code of HunyuanPortrait: Implicit Condition Control for Enhanced Portrait Animation

### Overview: 心迹 Vesti 是一款本地优先的 AI 对话记忆与知识管理工具，旨在捕捉、搜索、总结和导出跨多个 AI 平台的聊天记录。它通过整合本地向量数据库和 RAG 技术，为用户提供一个集中、安全且高效的对话记忆与知识管理平台。 ### Key features: - **本地优先的对话记忆与知识管理**：Vesti 通过本地向量数据库和 RAG 技术，确保用户的数据和思维主权不受其他平台的影响。 - **语义增强的独立知识看板**：依托本地向量数据库，Vesti 实现跨平台 RAG 检索，精准追溯长时段内的思考演进与认知转折。 - **多智能体驱动的结构化洞察**：后台自动提取复杂对话的关键脉络，生成多维度清晰摘要，一键将思维火花沉淀为永久策展笔记。 ### Tech stack: TypeScript ### Why it matters: 心迹 Vesti 为用户提供了一个集中、安全且高效的对话记忆与知识管理平台，帮助用户捕捉、搜索、总结和导出跨多个 AI 平台的聊天记录。它通过整合本地向量数据库和 RAG 技术，确保用户的数据和思维主权不受其他平台的影响。这种功能对于需要频繁跨平台协作和记录思考过程的用户来说尤为重要。