Projects

### Overview: This repository contains a curated list of prediction market tools, including analytics platforms, trading bots, dashboards, APIs, data feeds, alert systems, and educational resources. It covers popular platforms like Polymarket, Kalshi, Manifold, Hyperliquid, and more. ### Key Features: - **Oddpool:** The Bloomberg for prediction markets, aggregating cross-venue data across platforms. - **Alphascope:** AI-driven market intelligence engine for prediction markets. - **Octagon AI:** Intelligence layer for prediction markets. - **Polyfactual:** AI-powered platform blending prediction markets and social narratives. - **Polytrader:** AI-Powered Prediction Market Trading. - **Sportstensor:** Decentralized AI-powered sports prediction platform. - **BillyBets:** Autonomous AI sports betting agent. - **Fraction AI:** First decentralized auto-training platform for AI agents. - **Bankr:** The leading crypto AI agent with Polymarket integration. - **Pigeon:** AI-powered multi-platform trading terminal. - **Elastics:** AI-native, institutional operating system for automated trading. - **Forcazt:** AI-powered prediction market analytics and aggregator. ### Tech Stack: The repository is built using Markdown, GitHub Pages, and a custom theme. ### Why it Matters: This curated list provides a comprehensive overview of prediction market tools, aiding developers and traders in their decision-making process by offering a wide range of resources and tools for building and analyzing prediction markets.

Overview: Aeon is a fully autonomous AI agent framework built on GitHub Actions, designed to run unattended and self-heal without any approval loops. It seamlessly integrates with popular AI models like Claude, Codex, and more, making it a powerful tool for developers. Key features: - Runs unattended on GitHub Actions - Self-healing capabilities - Integrates with Claude, Codex, and other AI models - No approval loops required Tech stack: JavaScript Why it matters: Aeon simplifies the development process by eliminating the need for manual approval, making it ideal for teams looking to automate and streamline their workflows without compromising on functionality.

Overview: DNABERT is a pre-trained Bidirectional Encoder Representations from Transformers model designed specifically for DNA-language tasks in genomes. This model is part of a comprehensive project aimed at improving genome understanding and analysis through advanced natural language processing techniques. Key features: - Pre-trained DNA-language model - Fine-tuned for specific tasks - Includes source codes, usage examples, pre-trained models, and visualization tools - Under active development with more features planned Tech stack: Python Why it matters: DNABERT offers a powerful tool for analyzing and understanding DNA sequences using advanced natural language processing techniques, making it valuable for researchers and bioinformaticians working with genomic data. Repository: jerryji1993/DNABERT Primary language: Python GitHub stars: 783 GitHub description: DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome GitHub update: 2025/07/08: Links to pretrained DNABERT models have been updated to HuggingFace GitHub update: 2023/06/26: Publicly available DNABERT-2 model and updated usage instructions Citation: Please cite the following publications for using DNABERT in research: [1] Ji, Yanrong et al. "DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome." Bioinformatics (2021). [2] Zhou, Zhihan et al. "DNABERT-2: Efficient Foundation Model

Overview: BayesFlow is a Python library designed for simulation-based inference with deep learning, offering a user-friendly API for amortized Bayesian workflows, rich generative models, and seamless integration with PyMC. Key features: - A user-friendly API for amortized Bayesian workflows - A rich collection of generative models, including diffusion and consistency models - Multi-backend support via Keras 3, allowing use with PyTorch, TensorFlow, or JAX Tech stack: Python Why it matters: BayesFlow simplifies the process of Bayesian inference using deep learning, making it accessible for a wide range of applications, from complex simulators to traditional parametric models, with robust diagnostics and integration capabilities.

### Overview: CountBot 是一个面向中文用户的轻量级 AI Agent 框架，旨在帮助 AI 真正进入执行链路，连接大模型、IM 渠道、工作流与外部工具，实现跨入口协作与任务执行。 ### Key features: - **MCP 客户端模块**：支持多服务器连接、健康检查与工具发现，扩展 AI 执行能力边界。 - **Wiki 知识库模块**：基于 BM25 全文搜索，支持批量获取、相关性过滤与 LRU 缓存加速。 - **MCP、Wiki 管理面板**：组件模块化拆分与国际化支持。 - **WebSocket 状态广播模块**：实时同步 MCP 连接状态至前端会话。 - **上下文管理与心跳机制**：简化问候逻辑，减少 LLM 调用成本。 - **API Key 轮换机制与故障转移能力**：支持多 key 自动切换。 - **工具注册与初始化流程优化**：提升系统启动一致性与可维护性。 - **send_media 工具网页版兼容性优化**：支持在 WEB UI 预览图片和下载附件。 - **模型预设界面更新与构建产物**：发布说明：[https://countbot.cn/docs/releases/v0.9.0](https://countbot.cn/docs/releases/v0.9.0)。 ### Tech stack: Python ### Why it matters: CountBot 是一个连接模型、渠道

### Overview: This repository contains the Keras implementation of the Realtime Multi-Person Pose Estimation project, which was originally developed using Caffe. The updated version aims to remove the dependency on a separate C++ server, making the code more efficient and faster. ### Key Features: - **Training Code for Smaller Model Based on MobilenetV2**: Provides additional training code for a smaller model using the MobilenetV2 architecture. - **Visualisation of Predictions (Heatmaps, PAFs) in Tensorboard**: Enhances visualization of model predictions with heatmaps and PAFs. - **Additional Scripts for Tensorflow Lite Conversion and Model Testing**: Includes scripts to convert and test models for TensorFlow Lite. ### Tech Stack: - Keras ### Why It Matters: This repository offers a more streamlined and efficient implementation of the Realtime Multi-Person Pose Estimation project, making it easier to reproduce and utilize for various applications.

### Overview: OpenAttack is an open-source Python-based textual adversarial attack toolkit designed to handle the entire process of textual adversarial attacking, including preprocessing text, accessing the victim model, generating adversarial examples, and evaluation. ### Key features: - **Support for all attack types**: OpenAttack supports various types of attacks, including sentence-/word-/character-level perturbations and gradient-/score-/decision-based/blind attack models. - **Multilinguality**: OpenAttack supports English and Chinese, and its extensible design enables quick support for more languages. - **Parallel processing**: OpenAttack provides support for multi-process running of attack models to improve attack efficiency. - **Compatibility with 🤗 Hugging Face**: OpenAttack is fully integrated with 🤗 [Transformers](https://github.com/huggingface/transformers) and [Datasets](https://github.com/huggingface/datasets) libraries. - **Great extensibility**: You can easily attack a customized <u>victim model</u> on any customized <u>dataset</u> or develop and evaluate a customized <u>attack model</u>. ### Why it matters: - **Support for all attack types**: OpenAttack provides a comprehensive solution for various types of adversarial attacks, ensuring flexibility and effectiveness. - **Multilinguality**: The toolkit supports multiple languages, making it versatile for international applications. - **Parallel processing**: Efficiently leveraging multiple processes can significantly speed up the attack process. - **Compatibility with 🤗 Hugging Face**: Seamless integration with popular NLP libraries enhances usability and

### Overview: Precise RoI Pooling is an advanced integration-based average pooling method for Region of Interest (RoI) pooling, designed to enhance localization confidence in object detection tasks. This implementation is based on the paper "Acquisition of Localization Confidence for Accurate Object Detection" by Borui Jiang et al., published at ECCV 2018. ### Key Features: 1. **Continuous Gradient on Coordinates**: PrRoI Pooling provides a continuous gradient on bounding box coordinates, allowing for the optimization of RoI coordinates directly. 2. **Avoids Quantization**: Unlike traditional RoI Pooling methods, PrRoI Pooling uses average pooling instead of max pooling, ensuring no quantization errors. 3. **Integration-Based Pooling**: It integrates all pixels within a RoI, providing a more accurate representation of the RoI's features. 4. **Symbolic Links**: The source files inside the `./pytorch/prroi_pool/src/` and `./tensorflow/prroi_pool/src/kernels/external` folders are symbolically linked, ensuring compatibility across different environments. ### Tech Stack: - **Language**: C++ - **Frameworks**: PyTorch, TensorFlow ### Why It Matters: PrRoI Pooling offers significant improvements in localization confidence and accuracy, making it a valuable addition to object detection models. Its ability to handle continuous gradients and avoid quantization errors makes it particularly effective in scenarios where precise localization is crucial. ### Repository: - **GitHub Stars**: 781 - **Repository**: [vacancy/PreciseRoIPooling](

Overview: This Python package, `python_video_stab`, is designed to stabilize videos using OpenCV, providing a comprehensive solution for video stabilization tasks. Key features: - Stabilization of videos using OpenCV - Full searchable documentation available at [adamspannbauer.github.io/python_video_stab](https://adamspannbauer.github.io/python_video_stab) - Installation options for both `vidstab` without OpenCV and with OpenCV Tech stack: Python Why it matters: This package offers a straightforward and efficient way to stabilize videos, making it a valuable tool for anyone working with video stabilization in Python projects. Repository: AdamSpannbauer/python_video_stab Primary language: Python GitHub stars: 780 GitHub description: A Python package to stabilize videos using OpenCV README excerpt: A Python package to stabilize videos using OpenCV. Full [searchable documentation here](https://adamspannbauer.github.io/python_video_stab). This module contains a single class (`VidStab`) used for video stabilization. This class is based on the work presented by Nghia Ho in [SIMPLE VIDEO STABILIZATION USING OPENCV](http://nghiaho.com/?p=2093). The foundation code was found in a comment on Nghia Ho's post by the commenter with username koala. Input: Output:  ![](https://s3.amazonaws.com/python-vidstab/readme/stable_ostr

Overview: Anycost GANs are designed to generate consistent images under various computational budgets, enabling fast preview and interactive image editing. Key features: - Supports diverse computation costs using different channel and resolution configurations. - Sub-generators achieve high output consistency compared to the full generator. - Implements sampling-based multi-resolution training, adaptive-channel training, and generator-conditioned discriminator. Tech stack: Python Why it matters: Anycost GANs offer a flexible and efficient solution for interactive image synthesis and editing by balancing computational efficiency and image quality. Repository: mit-han-lab/anycost-gan GitHub stars: 779

### Overview: LLM Functions is a project that simplifies the creation of LLM (Large Language Models) tools and agents by leveraging familiar languages like Bash, JavaScript, and Python. It allows developers to connect their LLMs directly to custom code, making it easier to integrate complex functionalities without the need for complex integrations. ### Key Features: - **Function Calling**: Connects LLMs to custom code using the `function calling` feature. - **Bash, JavaScript, and Python Support**: Empowers developers to build powerful LLM tools and agents using their preferred languages. - **CLI Tools**: Provides a command-line interface for executing system commands, processing data, and interacting with APIs. ### Tech Stack: The project is built using Bash, JavaScript, and Python. It leverages the `argc` framework for command-line execution and the `jq` tool for JSON processing. ### Why it Matters: LLM Functions simplifies the process of creating LLM tools and agents by abstracting the complexity of integrating LLMs with custom code. It allows developers to focus on the functionality of their tools without worrying about the integration details, making it easier to build and deploy powerful LLM solutions.

### Overview: MagicPose is a novel method for generating realistic human poses and facial expressions that are retargeted to a specific identity. It leverages identity-aware diffusion models to achieve high-quality results, making it particularly useful in applications where the subject's identity needs to be preserved. ### Key features: - Realistic human pose and facial expression generation - Retargeting to a specific identity - Identity-aware diffusion models for high-quality results ### Tech stack: Python ### Why it matters: MagicPose is a significant advancement in the field of video generation, particularly in the areas of realistic human pose and facial expression retargeting. By leveraging identity-aware diffusion models, it offers a more accurate and versatile solution compared to existing methods, making it highly valuable for applications in video editing, animation, and other areas where the subject's identity needs to be preserved while maintaining realism.

Overview: CodeFuse MuAgent is an innovative agent framework driven by KG Engine, designed to assist users in executing complex tasks under human guidance. It supports multi-agent orchestration, Java-based tool and agent registration, and a user-friendly front-end interface. Key features: - One-click deployment capabilities - KG-based agent orchestration - Java-based tool and agent registration and management - User-friendly drag-and-drop front-end product UI Tech stack: Python Why it matters: MuAgent simplifies the development and management of complex agent-based systems, leveraging KG Engine for enhanced reasoning and collaboration capabilities. It is compatible with existing market frameworks and can be integrated with technologies like FunctionCall, RAG, and CodeInterpreter, making it a valuable tool for DevOps scenarios within organizations.

Overview: Memorizz is a Python library designed to optimize memory usage by connecting persistent memory, models, and prompts for AI applications. It leverages popular databases and storage solutions to enhance agent performance. Key features: - **Memory Management:** Supports various types of memory (episodic, semantic, procedural, short-term, and shared) with configurable storage options. - **Agent Framework:** Provides tools and integrations for building persistent agents, including assistant, workflow, and deep research modes. - **Storage Providers:** Supports local filesystem storage and cloud-based providers like MongoDB, Oracle, and Notion. - **Context Efficiency:** Offers tools for managing context, deduplicating memories, summarizing history, and controlling retention and forgetting. Tech stack: Python Why it matters: Memorizz optimizes AI agent performance by efficiently managing memory, ensuring agents can recall relevant information and make better decisions across sessions. Repository: RichmondAlake/memorizz GitHub stars: 776

### Overview: CorridorKey-Runtime is a native AI keying runtime and OFX plugin for DaVinci Resolve and Foundry Nuke, developed in collaboration with Corridor Digital. It provides three product surfaces: an OFX plugin for interactive keying, a CLI for local processing, and a GUI for a graphical workflow. ### Key Features: - **OFX Plugin**: Directly integrates with DaVinci Resolve and Foundry Nuke through the OFX plugin. - **CLI (`corridorkey`)**: Provides environment diagnostics, scripted processing, and JSON output for integration workflows. - **GUI (Tauri-based desktop app)**: Offers a graphical workflow option for users preferring a visual interface. ### Tech Stack: C++ ### Why It Matters: CorridorKey-Runtime simplifies the integration of AI keying into popular video editing software, making it easier to use AI for keying tasks in real-time workflows. The stable surfaces for DaVinci Resolve and Foundry Nuke, along with support for NVIDIA RTX and Apple Silicon, ensure compatibility across different hardware configurations. ### GitHub Stars: 771

### Overview: This repository contains lecture and seminar materials for a course on Deep Learning for Audio Processing (DLA). The course covers various aspects of audio processing, including digital signal processing, automatic speech recognition, and source separation. The materials are organized into weekly folders, and contributions are welcome to improve and expand the course content. ### Key Features: - **Lecture and Seminar Materials**: Each week includes lecture notes and seminar discussions. - **Self-Study Resources**: Basic introduction to PyTorch and basic DevOps. - **Technical Support**: Any technical issues or course materials can be reported as GitHub issues. - **Version History**: The current version is conducted in **autumn 2026** at the [CS Faculty](https://cs.hse.ru/en/) of [HSE](https://www.hse.ru/en/). ### Tech Stack: - **Primary Language**: Jupyter Notebook ### Why it Matters: This course is essential for anyone interested in the field of audio processing and machine learning. It provides a comprehensive understanding of how to apply deep learning techniques to audio data, making it valuable for researchers, developers, and professionals in the field. ### Repository: The repository is named `markovka17/dla` and has been starred by 773 users. The primary language used is Jupyter Notebook. ### Contributors & Course Staff: The course materials and teaching were delivered by a team of instructors including Maxim Kaledin, Georgy Gospodinov, Georgiy Pistsov, Aibek Alanov, Petr Gr

Overview: This repository contains code for the NeurIPS 2023 paper "Emergent Correspondence from Image Diffusion". It provides a framework for establishing semantic correspondence between images using diffusion models. Key features: - Extracts DIFT (Diffusion Correspondence Features) from images - Provides an interactive demo for visualizing semantic correspondences - Offers a command to extract DIFT from given images - Demonstrates how to propagate edits between images sharing semantic correspondences Tech stack: Python Why it matters: This repository is significant for researchers and developers interested in image-to-image translation and semantic correspondence, offering a practical tool for establishing and visualizing semantic correspondences between images using diffusion models. Repository: Tsingularity/dift GitHub stars: 774

Overview: This project provides a GPU/TPU implementation of RingAttention, a model that enables training sequences up to a length of 'number of devices' times longer than those possible with BPT. It achieves this by distributing attention and feedforward computation across multiple devices and overlapping communication with computation. Key features: - Implementation of RingAttention with Blockwise Transformers - Enables training sequences up to a length of 'number of devices' times longer than those possible with BPT - Distributed attention and feedforward computation across multiple devices - Overlaps communication with computation for efficient training Tech stack: Python Why it matters: This project is relevant for researchers and developers interested in large context models, enabling training of sequences up to a million tokens without increasing communication or computation overhead, making it suitable for applications requiring near-infinite context.

### Overview: Lix is an embeddable version control system designed to store files and tables in a single repository. It supports various file formats and SQL tables, enabling real-time collaboration and easy integration into applications. ### Key Features: - **Any File:** Store text, binary files, and large files like Word, PowerPoint, CAD, and video. - **SQL Tables:** Version rows for CRM records, logs, and other application data alongside files. - **Designed as a Database:** Files, tables, and history share a single ACID OLTP database, allowing efficient querying of millions of rows with SQL. - **Real-time Collaboration:** People and agents can share a repository and see changes in real-time. - **Permissions (Planned):** Per file, per group, stored in the repository and versioned like any other change. - **Embeddable:** Runs in-process as a library, and storage can be pluggable: memory, filesystem, browser OPFS, or S3. ### Tech Stack: Rust ### Why It Matters: Lix provides a unified approach to managing files and tables, making it easier to integrate with applications and ensure data integrity. Its real-time collaboration feature and ability to store structured data alongside files make it a valuable tool for teams working with both structured and unstructured data.

### Overview: The Academic Commercialization Assessment Agent is an AI-powered tool that converts research papers into comprehensive assessments of technology readiness, manufacturing potential, patent landscape, market analysis, and risk evaluation. This agent uses a multi-agent system with a Python backend, CrewAI, and FastAPI, providing reproducible benchmarks and a detailed documentation. ### Key Features: 1. **Evidence First, Model Second**: The system retrieves, validates, and freezes the sources before using six LLM stages to reason over the registry. 2. **Measured in the Open**: Protocols, held-out checks, and errata are live in the repository, ensuring transparency and accuracy. 3. **Built for Runs That Fail**: An interrupted run can be resumed from its last validated checkpoint, and paid requests generate receipts, allowing results to be recovered up to 24 hours without re-paying. ### Tech Stack: - **Python**: The backend programming language. - **CrewAI**: The AI framework used for the agent’s reasoning. - **FastAPI**: The backend API framework for the agent’s functionality. ### Why It Matters: The Academic Commercialization Assessment Agent is designed to help researchers and investors quickly assess the commercial potential of new technologies. By providing a detailed, evidence-based assessment, it helps streamline the process of identifying promising technologies and accelerating their commercialization.