Projects

Overview: xpander.ai is a comprehensive runtime and control plane designed to build, run, and ship reliable AI agents quickly and anywhere. Key features: - Chat (generalist AI agent for easy access to agents) - Workbench (control plane for any agent framework) - AgentOS (deploy agents on VPC/Kubernetes or xpander cloud) - Connector Hub (generate and run MCP servers) - AI Tools (OCR, browser automation, code interpreter, etc.) Tech stack: Jupyter Notebook Why it matters: xpander.ai simplifies the process of building, deploying, and managing AI agents, making it accessible and efficient for developers and businesses alike.

### Overview: MiniGPT-5 is an innovative model that leverages the concept of "generative vokens" to achieve interleaved vision-and-language generation. This approach enhances the ability of large language models to generate coherent images and textual narratives, improving performance across various benchmarks, including MMDialog and VIST datasets. ### Key Features: - **Interleaved Vision-and-Language Generation**: Combines vision and language tasks in a single model to generate images and text. - **Generative Vokens**: Acts as a bridge between vision and language, facilitating harmonized image-text outputs. - **Two-Stage Training Strategy**: Focuses on description-free multimodal generation, requiring no comprehensive descriptions of images. - **Classifier-Free Guidance**: Enhances the effectiveness of vokens on image generation, improving model integrity. ### Tech Stack: Python ### Why it Matters: MiniGPT-5 demonstrates significant improvements over the baseline Divter model on the MMDialog dataset and consistently delivers superior or comparable multimodal outputs in human evaluations on the VIST dataset, highlighting its efficacy across diverse benchmarks.

Overview: IceVision is an agnostic computer vision framework that provides a curated collection of high-quality pre-trained models from various training libraries, including PyTorch-Lightning, Fastai, MMDetection, YOLOv5, EfficientDet, and more. It simplifies the deep learning workflow by orchestrating end-to-end deep learning tasks. Key features: - Access to hundreds of neural net models - Pluggable transforms for better model generalization - Data curation/cleaning with auto-fix - Multiple training loop libraries - Multi-task training for efficient model combination Tech stack: Python Why it matters: IceVision simplifies the complex process of computer vision by providing a unified framework that can be easily integrated with various training libraries, making it accessible to developers and researchers alike.

Overview: Temps is an AI-native open-source self-hosted alternative to popular SaaS tools like Vercel, Sentry, PostHog, Pingdom, Resend, and E2B. It provides 440+ CLI operations with drop-in skills for Claude Code, Codex, and OpenCode, enabling developers to deploy, inspect traces, run migrations, and more. Key features: - 440+ CLI operations across 69 groups - Skills for deploying, inspecting traces, running migrations, and more - Framework presets auto-detected for importing public repositories - AI-native chat with read-only access by default - AI-compatible gateway for BYOK provider keys - Built-in web analytics and session replay - Uptime monitoring and alerts for deploy failures, runtime crashes, and backup health Tech stack: Rust Why it matters: Temps offers a comprehensive solution for developers looking to replace multiple SaaS tools with a single self-hosted binary, providing a unified platform for deployment, analytics, session replay, error tracking, OpenTelemetry, email, and AI gateway capabilities. Repository: gotempsh/temps GitHub stars: 867 GitHub description: AI-native open-source alternative to Vercel + Sentry + PostHog + Pingdom + Resend + E2B. 440+ CLI operations with drop-in skills for Claude Code, Codex & OpenCode — deployments, analytics, session replay, error tracking, OpenTelemetry, email, sandboxes & AI gateway in one self-hosted Rust

### Overview: PoseNet is an official PyTorch implementation of the Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation from a Single RGB Image, presented at ICCV 2019. This repository provides a flexible and simple codebase for pose estimation, compatible with various datasets including Human3.6M, MPII, MS COCO 2017, MuCo-3DHP, and MuPoTS-3D. ### Key features: - **PoseNet Implementation**: Provides a PyTorch implementation of the Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation. - **Compatibility**: Compatible with popular datasets including Human3.6M, MPII, MS COCO 2017, MuCo-3DHP, and MuPoTS-3D. - **Visualization**: Includes visualization scripts for 3D pose estimation. ### Tech stack: - **Python 3.6.5 with Anaconda 3** - **PyTorch** - **CUDA 9.0** - **cuDNN 7.1** - **COCO API** ### Why it matters: This repository offers a practical and efficient solution for 3D multi-person pose estimation, making it suitable for researchers and practitioners looking to improve their 3D pose estimation models.

Overview: PointFlow is a PyTorch implementation of a paper that proposes a principled probabilistic framework for generating 3D point clouds. The model learns a two-level hierarchy of distributions where the first level is the distribution of shapes and the second level is the distribution of points given a shape. This allows for both shape sampling and point sampling from a shape. The model is trained with continuous normalizing flows, enabling likelihood computation and variational inference. Key features: - Continuous normalizing flows for likelihood computation - Hierarchical distribution modeling for shape and point generation - State-of-the-art performance in point cloud generation - Ability to faithfully reconstruct point clouds and learn useful representations in an unsupervised manner Tech stack: Python Why it matters: PointFlow provides a principled approach to generating high-fidelity 3D point clouds, making it valuable for applications requiring accurate and detailed point cloud representations. Its ability to faithfully reconstruct point clouds and learn useful representations in an unsupervised manner sets it apart from other generative models.

Overview: Finetune-Anything is a tool designed for fine-tuning the Segment Anything Model (SAM) for various computer vision tasks such as semantic segmentation, matting, and detection. It simplifies the process by providing datasets and support for customizing the SAM model. Key features: - Fine-tuning of SAM for specific computer vision tasks - Support for semantic segmentation, matting, detection, and instance segmentation - Customizable SAM model through design of extend-SAM modules - Onnx export capability for model deployment Tech stack: Python Why it matters: Finetune-Anything streamlines the fine-tuning process of SAM, making it accessible for researchers and developers to tackle basic computer vision tasks efficiently. Repository: ziqi-jin/finetune-anything Primary language: Python GitHub stars: 867

Overview: 深度学习与大模型的可视化学习路线，从1986年的RNN到2025年的最新进展，按19个架构家族梳理演进脉络。 Key features: - 深度学习与大模型的可视化学习路线 - 从1986年的RNN到2025年的最新进展 - 按19个架构家族梳理演进脉络 Tech stack: TypeScript Why it matters: 了解深度学习与大模型的发展历程，掌握其核心技术，为未来技术趋势做好准备。

### Overview: Yai is an AI-powered terminal assistant that uses OpenAI ChatGPT to run commands for you, enhancing your command line experience. ### Key features: - **Command Execution**: Yai understands your commands and executes them directly in your terminal. - **Configuration**: Fine-tunes your experience with supplementary preferences. - **Integration**: Seamless integration with your operating system and preferred tools. ### Tech stack: Go ### Why it matters: Yai streamlines your command line experience by automating repetitive tasks, making it easier and more efficient to manage your terminal environment.

### Overview: This toolkit provides a comprehensive voice activity detection (VAD) solution, including DNN, bDNN, LSTM, and ACAM-based VAD. It also includes a dataset of directly recorded speech data. The toolkit supports various classifiers and is compatible with both Python and MATLAB. ### Key features: - **DNN-based VAD**: Uses a deep neural network for accurate detection. - **bDNN-based VAD**: Boosted deep neural network for improved performance. - **LSTM-based VAD**: Long short-term memory recurrent neural network for dynamic context analysis. - **ACAM-based VAD**: Adaptive context attention model for enhanced accuracy. - **Recorded Dataset**: Directly recorded speech data for training and testing. ### Tech stack: MATLAB and Python. ### Why it matters: This toolkit is essential for researchers and developers working on speech processing and analysis, offering a robust solution for VAD with various classifier options. The dataset is valuable for training and testing purposes, ensuring high accuracy in real-world applications.

## Overview: Recursive Self-Improvement (RSI) is a research field focused on enhancing the capabilities of artificial intelligence systems through self-improvement mechanisms. This includes self-training, agent memory optimization, harness evolution, and embodied self-improvement. RSI is relevant as it has advanced to the point where it is becoming an empirical research direction rather than just a theoretical idea. ## Key Features: - **Self-Training & Self-Reward** - **Synthetic Data & Self-Distillation** - **Self-Play & Iterative Fine-tuning** - **Self-Taught Reasoning** - **Prompt & Program Optimization** - **Context & Memory Evolution** - **Harness & Scaffold Evolution** - **Extensible Harness Substrates** - **Self-Verification & Self-Correction — Enabling Foundations** - **Self-Evolving Agent Frameworks** - **Co-Evolution** - **Inference-time Debate** - **Self-Modifying Coding Agents** - **Iterative Repair & Training** - **Automated AI R&D** - **Embodied & Physical Self-Improvement** - **Evolutionary & Open-Ended RSI** - **Safety, Alignment & Theory** - **Supporting Safety Foundations** - **Introspection & Self-Modeling** - **Benchmarks & Evaluations** - **Direct RSI & Self-Improvement Evaluations** - **Frontier Lab Self-Improvement & AI R&D Evaluation Frameworks** - **AI R&D Capability Proxies** - **

Overview: Cerebellum is a lightweight browser automation system that uses AI-driven planning to navigate web pages and perform goals. Key features: - Compatible with any Selenium-supported browser. - Fills forms using user-provided JSON data. - Accepts runtime instructions to dynamically adjust browsing strategies and actions. - TODO: Create training datasets from browsing sessions. Tech stack: Python Why it matters: Cerebellum simplifies web browsing by transforming it into a directed graph, making it easier to achieve user-defined goals using keyboard and mouse actions. Repository: theredsix/cerebellum GitHub stars: 865 GitHub description: Browser automation system that uses AI-driven planning to navigate web pages and perform goals.

### Overview: A curated list of awesome LLM frameworks, libraries, and software, covering a wide range of topics from tutorials and books to machine learning and AI applications. ### Key Features: - Tutorials and Books - Examples and Exercises - Awesome Lists and Collections - Editor and IDE Support - Version Control - Frontend and UI Components - Scraping and Crawling - Databases - Serialization and Formats - LLM and Inference - Machine Learning Frameworks - Computer Vision - Natural Language Processing - Data Science and Analytics - Agent Frameworks and Runtimes - Agent Skills and Tooling - Memory and Context - Evaluation and Benchmarks - Applications and End User Tools - Image and Video - Security Tools - Performance and Optimization - Command Line Tools - Business and Productivity - Other ### Tech Stack: Python, JavaScript, and various other programming languages. ### Why it Matters: This curated list provides a comprehensive overview of the latest advancements in Large Language Models (LLMs), offering practical insights and step-by-step tutorials for developers looking to get started with or deepen their understanding of LLMs and related technologies.

Overview: This sample demonstrates how to build a serverless AI chat experience using LangChain.js and Azure, focusing on Retrieval-Augmented Generation. The application leverages Azure Static Web Apps and Azure Functions, with Azure Cosmos DB for NoSQL as the vector database. It provides a starting point for developing more complex AI applications. Key features: - A chatbot experience powered by LangChain.js - Integration with Azure Static Web Apps and Azure Functions - Use of Azure Cosmos DB for vector database - Sample data for enterprise documents and support guide Tech stack: TypeScript, LangChain.js, Azure Why it matters: This sample simplifies the process of building AI applications by utilizing LangChain.js and Azure's serverless technologies, making it easier and more cost-effective to create serverless chatbots.

Overview: This repository contains tutorials designed to fine-tune transformers for various natural language processing (NLP) tasks, leveraging the power of PyTorch and Hugging Face's Transformers library. Key features: - Text Classification: Multi-Class - Text Classification: Multi-Label Tech stack: PyTorch and Hugging Face's Transformers library. Why it matters: These tutorials aim to simplify the process of fine-tuning language models for specific NLP tasks, making it accessible to developers and data scientists without requiring deep knowledge of NLP or transformer models.

### Overview: MOMENT: A Family of Open Time-series Foundation Models MOMENT is a comprehensive framework designed to address the challenges of pre-training large models on time-series data. This project introduces a family of open-source foundation models that can be pre-trained on a large and diverse collection of public time-series data, known as the Time-series Pile. The repository includes both the model and dataset, enabling researchers and practitioners to evaluate and improve time-series foundation models in various scenarios. ### Key Features: - **Model Variants:** MOMENT includes three distinct model variants: `MOMENT-1-small`, `MOMENT-1-base`, and `MOMENT-1-large`. - **Data Collection:** The Time-series Pile serves as a comprehensive repository of public time-series data, facilitating large-scale multi-dataset pre-training. - **Benchmark Development:** The project also provides a benchmark dataset and evaluation metrics to assess the performance of time-series foundation models in limited supervision settings. - **Pre-training Performance:** Experiments on the benchmark demonstrate the effectiveness of pre-trained models with minimal data and task-specific fine-tuning. ### Tech Stack: TypeScript The project is implemented using TypeScript, a statically typed language that provides better type safety and tooling support compared to JavaScript. ### Why It Matters: MOMENT is a significant contribution to the field of time-series analysis, offering a robust framework for pre-training large models on diverse datasets and tasks. The inclusion of a comprehensive benchmark and pre-trained models makes it an invaluable resource for researchers and practitioners looking to improve time-series foundation models. ### GitHub

### Overview: This repository contains scripts to crawl and download the BookCorpus corpus, which is a popular large-scale text corpus for unsupervised learning of sentence encoders/decoders. The original corpus is no longer available, but this repository provides a replica that can be used for research purposes. The corpus is sourced from Smashwords, and the scripts include a downloader and post-processing tools to facilitate easy access and use. ### Key Features: - **Crawler**: Downloads books from Smashwords. - **Post-processing**: Combines downloaded text into a single file with sentences per line. - **Tokenization**: Tokenizes the text into segmented words using Microsoft's BlingFire tool. - **Error Handling**: Handles common download errors and provides a list of available books. ### Tech Stack: - Python ### Why it Matters: The repository is useful for researchers and developers looking to replicate the BookCorpus corpus for their own projects, ensuring they can access the same data as the original corpus. It provides a reliable and accessible alternative for those who cannot access the original source. ### Repository: soskek/bookcorpus - **Primary Language**: Python - **GitHub Stars**: 864 - **GitHub Description**: Crawl BookCorpus - **README Excerpt**: This repository includes scripts to reproduce BookCorpus by yourself. It provides a crawler for downloading books from Smashwords, and post-processing tools to combine and tokenize the downloaded text. The scripts are designed to be used responsibly, ensuring compliance with copyright laws and terms of service. ### Disclaimer: The repository is provided

Overview: Aiden is an autonomous AI agent designed to work across your files, terminal, browser, supported applications, APIs, repositories, and connected services, making it easier to plan, act, recover, and verify results. Key features: - Plans and executes tasks across various platforms. - Operates multiple providers and channels. - Supports multiple programming languages and environments. - Bundled skills and tools. - Compatible with Windows, Linux, WSL, and macOS (API mode). Tech stack: TypeScript, JavaScript, Node.js, ESM, esbuild, npm. Why it matters: Aiden simplifies complex tasks by automating them across your entire ecosystem, ensuring efficiency and reliability.

Overview: cvui is a lightweight, header-only UI library built on top of OpenCV drawing primitives. It offers a simple and intuitive user interface without requiring a graphical backend, making it ideal for OpenCV applications. Key features: - Lightweight and simple to use - Header-only with no external dependencies (except OpenCV) - Based on OpenCV drawing primitives only - Friendly and C-like API - Easily render components without worrying about their position - Simple mouse API - Modest number of UI components (11 in total) - Available in C++ and Python Tech stack: C++ and Python Why it matters: cvui simplifies the creation of user interfaces in OpenCV applications by eliminating the need for a graphical backend, such as OpenGL or Qt. This makes it particularly useful for developers working with OpenCV projects where a graphical user interface is essential but the overhead of setting up a separate backend is undesirable.

### Overview: ConsisID is an innovative identity-preserving text-to-video generation model that leverages frequency decomposition for controllable and consistent video synthesis. This project is part of the CVPR 2025 Highlight and aims to preserve human identity across different videos. ### Key Features: - **Frequency Decomposition**: ConsisID uses frequency decomposition to analyze and reconstruct the visual content, ensuring consistent identity across the generated video. - **DiT-based Model**: The model is based on a tuning-free DiT (Disentangled Identity Transformers) architecture, making it efficient and easy to deploy. - **Controllable Generation**: Users can control the identity of the generated video by adjusting the frequency decomposition parameters. ### Tech Stack: Python ### Why It Matters: ConsisID is a significant contribution to the field of video synthesis, providing a robust and controllable solution for preserving human identity in generated videos. Its approach offers a fresh perspective on frequency analysis and diffusion transformers, making it a valuable addition to the research and development community. Repository: PKU-YuanGroup/ConsisID GitHub Stars: 863 GitHub Description: [CVPR 2025 Highlight] Identity-Preserving Text-to-Video Generation by Frequency Decomposition