Projects

Overview: Lingua is a Rust library designed for accurate natural language detection, suitable for short text and mixed-language text. It simplifies the preprocessing step for applications like text classification and spell checking. Key features: - Supports 75 languages - Detects language in short text snippets - Combines rule-based and statistical Naive Bayes methods - Does not require external APIs or neural networks Tech stack: Rust Why it matters: Lingua provides accurate results on both short and long texts, eliminating the need for lengthy text fragments and improving detection accuracy even for single words and phrases. It is designed to be used offline, making it suitable for applications that require quick and reliable language detection without the need for machine learning frameworks or extensive configuration.

### Overview: STRIDE GPT is an AI-powered threat modeling tool that leverages OpenAI's GPT models to generate threat models for applications based on the STRIDE methodology. Users can provide details such as the application type, authentication methods, and whether it's internet-facing or processes sensitive data. The tool generates threat models and attack trees based on the provided information, offering a comprehensive approach to threat modeling for developers. ### Key features: - **Agentic Codebase Analysis**: Generates an autonomous, deep STRIDE threat model from a codebase. - **CLI and Interactive REPL**: Full-featured terminal experience with tab completion, history, and real-time progress. - **Simple and User-friendly Streamlit Web Interface**: Generates threat models based on the STRIDE methodology. - **Agentic AI Support**: Threat modeling for agentic AI systems with OWASP Top 10 integration. - **Generative AI Support**: Threat modeling for GenAI applications with OWASP LLM Top 10 integration. - **MITRE ATT&CK & ATLAS Mapping**: Threats are annotated with standardized adversary technique IDs (MITRE ATT&CK Enterprise for traditional infrastructure attacks, ATLAS for ML/LLM-specific attacks). - **Architectural Pattern Detection**: Automatically detects RAG pipelines, multi-agent systems, code execution environments, tool ecosystems, and more from application descriptions. - **Embedded Draw.io Diagram Editor**: Create and edit architecture diagrams directly in STRIDE-GPT using the integrated Draw.io editor. - **Multi-modal**: Use architecture diagrams, flowcharts, etc., as inputs for threat modeling

Overview: Numbat is a tool designed to monitor and detect AI agent activity on endpoints, offering live monitoring, endpoint-local detection, optional pre-action blocking, and forensic reconstruction. It uses Go as its primary language and has been featured with 1,144 GitHub stars. Key features: - Live monitoring through hooks and plugins - Endpoint-local detection with built-in and custom rules - Opt-in blocking through supported pre-action hooks - Forensic reconstruction from on-disk session artifacts - Versioned NDJSON records for events, findings, enforcement decisions, and scan summaries - Read-only artifact scanning with secret redaction - Inventory and investigation tools for read-only agent discovery - Single-binary distribution for macOS, Linux, and Windows Tech stack: Go Why it matters: Numbat provides visibility into AI agent activity on endpoints, offering comprehensive monitoring, detection, and forensic capabilities, making it useful for security and compliance purposes. Repository: perplexityai/numbat Primary language: Go GitHub stars: 1144

### Overview: Mysti is an open-source coding workspace that brings together your agents inside VS Code, enabling independent opinions, brainstorming, and collaborative decision-making. ### Key Features: - **Independent Opinions**: Claude Code and Codex provide separate opinions on code snippets. - **Brainstorming**: Agents debate solutions and synthesize the best approach. - **Cross-Agent Handoffs**: Work is passed between agents for review and feedback. - **Proactive Context**: Agents keep track of their responses, context, permissions, and progress in one conversation. ### Tech Stack: TypeScript ### Why it Matters: Mysti allows developers to leverage multiple AI agents for collaborative coding, enhancing problem-solving and decision-making processes within VS Code. ### Repository: DeepMyst/Mysti ### Primary Language: TypeScript ### GitHub Stars: 1145 ### Repository Description: The AI coding dream team of Claude Code and Codex collaborating in brainstorm mode, debate solutions, and synthesize the best approach for your code. Connects the coding tools you choose—CLI agents, local model servers, and the DeepMyst-powered Mysti coordinator. Use one agent when that's enough; bring in more when another perspective helps. [](https://github.com/DeepMyst/Mysti/actions/workflows/ci.yml) ### Get a Second Opinion Without Rewriting the Question - **Claude Code and Codex Respond in Separate Agent Cards**: Compare independently attributed answers in the

Overview: MassGen is an open-source multi-agent scaling system designed to run in your terminal, autonomously orchestrating frontier models and agents to collaborate, reason, and produce high-quality results. Key features: - Cross-Model/Agent Synergy - Parallel Processing - Intelligence Sharing - Consensus Building - Live Visualization Tech stack: Python Why it matters: MassGen leverages the power of collaborative AI to solve complex tasks, scaling AI with continuous refinement and improvement, making it a significant tool in the field of generative AI.

Overview: THUOCL 是由清华大学自然语言处理与社会人文计算实验室整理推出的一套高质量中文词库，包含词频统计信息DF值，词库经过多轮人工筛选，收录准确，开放更新，欢迎专业人士加入协作建设开放词库。 Key features: - 包含词频统计信息DF值，方便用户个性化选择使用。 - 词库经过多轮人工筛选，保证词库收录的准确性。 - 开放更新，将不断更新现有词表，并推出更多类别词表。 - 词库清单详细列出，涵盖IT、财经、成语、地名、历史名人、诗词、医学、饮食、法律、汽车、动物等类别。 Tech stack: Python, NLP, Data Processing Why it matters: THUOCL 是一个开放的中文词库，可以用于中文自动分词，提升中文分词效果，同时提供丰富的词频统计语料库，帮助用户更好地理解和使用中文词汇。

Overview: Fess is an open-source, self-hosted enterprise search server built on OpenSearch, designed for organizations to install and run on any platform with a Java runtime. It supports crawling from web sites, file systems, and data stores, and features a browser-based administration UI and REST API with support for many document formats and multilingual user interfaces. Key features: - Full-text search with faceting, sorting, and search suggestions - Crawlers for web sites, file systems, and data stores - Support for many document formats, including Microsoft Office, PDF, and archives - Browser-based administration UI and REST API - Role- and permission-based filtering of search results - Single sign-on with LDAP, OpenID Connect, SAML, SPNEGO, and Microsoft Entra ID - Multilingual user interface and text analysis (20+ languages) - Extensible through data store, ingest, script, and theme plugins Tech stack: Java Why it matters: Fess is a robust, self-hosted enterprise search solution that can be easily integrated into various environments, providing comprehensive search capabilities with support for multiple languages and formats.

### Overview: TurboOCR is an extremely fast GPU document parser that supports OCR, layout, tables, formulas, and Markdown. It runs at up to 559 images/s on receipts and 20 pages/s on dense documents, making it the fastest GPU document parser by default. ### Key Features: - **OCR, Layout, Tables, Formulas, and Markdown**: TurboOCR parses documents from receipts, forms, and receipts, converting them into Markdown format. - **PP-OCRv6 Detection & Recognition**: One model covers Latin, Chinese, and Japanese, with options for `tiny`, `small`, and `medium` tiers for higher accuracy. - **HTTP + gRPC**: Supports local processing without VLM, and runs on Linux with NVIDIA GPUs. - **High Accuracy**: TurboOCR is accurate on forms and receipts, competitive with PaddleOCR-VL, PaddleOCR-Python, RapidOCR, EasyOCR, and Tesseract. ### Tech Stack: - **C++ / CUDA / TensorRT / PP-OCRv6**: The primary implementation is in C++, using CUDA for parallel processing and TensorRT for optimized inference. ### Why It Matters: - **Speed**: TurboOCR is the fastest GPU document parser, capable of processing up to 559 images/s on receipts and 20 pages/s on dense documents. - **Accuracy**: It is accurate on forms and receipts, making it competitive with leading OCR engines like PaddleOCR-VL, PaddleOCR-Python, RapidOCR, EasyOCR, and Tesseract. ### Repository: -

Overview: StashBase is an open-source local knowledge base designed for writing with AI agents, allowing you to keep your notes and sources on your computer and integrate Claude, Codex, or Antigravity for writing. It supports side-by-side collaboration with your agent, keeping your knowledge base updated, and allows you to bring your preferred agent to the writing process. Key features: - Integrated with Claude, Codex, and Antigravity for seamless writing. - Side-by-side collaboration with your agent. - Up-to-date knowledge base with automatic updates of notes and sources. - Local search and text extraction for PDFs, Word documents, and scanned materials. - Ability to use your existing agent without any additional setup. Tech stack: TypeScript Why it matters: StashBase provides a convenient and efficient way to integrate AI writing assistants directly into your local workspace, ensuring that your notes and sources remain accessible and up-to-date while allowing you to focus on the writing process.

### Overview: This repository is dedicated to a collection of papers and summaries on hallucination detection in Large Language Models (LLMs). It serves as a comprehensive resource for developers interested in understanding and improving the reliability of LLMs through hallucination detection techniques. ### Key Features: - **INTRYGUE:** A method that introduces **INTRYGUE**, a training-free technique to gate predictive entropy with an attention-based estimate of induction-head activity. - **TOHA (Topology-based HAllucination detector):** Represents attention maps as weighted graphs and measures Manifold Topology Divergence between response and prompt subgraphs to detect hallucinations. - **ParamMute:** Suppresses knowledge-critical FFNs for faithful retrieval-augmented generation, improving faithfulness on benchmarks like CoFaithfulQA and ConFiQA. ### Tech Stack: The repository uses Python for its technical implementation. ### Why It Matters: Understanding and improving hallucination detection in LLMs is crucial for maintaining the reliability and trustworthiness of these models, ensuring that they provide accurate and contextually relevant responses. By focusing on these specific techniques, developers can enhance the overall quality and reliability of LLMs in various applications.

Overview: This repository is a curated collection of datasets and tools related to artificial intelligence and data mining in digital games. It serves as a guide for developers looking to apply AI and data mining techniques in game development. Key features: - Datasets and tools for AI and data mining in games - Categories: Mobile, Web, Books, Dataset, Market Research, Miscellaneous, and License Tech stack: Python, JavaScript, and various APIs. Why it matters: This repository provides essential resources for developers interested in integrating AI and data mining techniques into their game projects, making it a valuable resource for anyone working in the field of game development.

Overview: This project focuses on developing a straightforward end-to-end method for question generation using pre-trained transformers, aiming to simplify data processing and training scripts for easier inference. Key features: - Answer-aware question generation - Answer extraction models - Multitask QA-QG - End-to-end question generation (answer agnostic) Tech stack: Jupyter Notebook Why it matters: This project provides a simplified approach to question generation using pre-trained transformers, making it accessible for developers and researchers.

### Overview: Z80-μLM is a 2-bit quantized language model designed to run on an 8-bit Z80 processor. It enables the training and export of conversational models as CP/M .COM binaries, allowing users to chat with their vintage computers. ### Key Features: - Trigram hash encoding for input text hashing into 128 buckets. - 2-bit weight quantization, packed 4 per byte. - 16-bit integer inference using Z80-native 16-bit signed arithmetic. - ~40KB .COM file size, fitting in CP/M's Transient Program Area (TPA). - Autoregressive generation, outputting text character-by-character. - No floating point operations, all math using fixed-point scaling. - Interactive chat mode, just run `CHAT` with no arguments. ### Tech Stack: Python ### Why It Matters: Z80-μLM offers a unique approach to conversational AI, allowing for easy self-hosted distribution and training of models. The 40KB .COM file size and 2-bit quantization make it suitable for 8-bit Z80 processors, enabling retrocomputing enthusiasts to interact with classic computers in a new way.

### Overview: RAG-Retrieval is a comprehensive framework designed for training, inference, and distillation of Richly Augmented Graph-based Retrieval (RAG) models. It supports fine-tuning of various RAG retrieval models, including embedding models, late interactive models, and reranker models, making it a versatile tool for enhancing retrieval efficiency and accuracy. ### Key Features: - **End-to-End Fine-tuning of RAG Models**: RAG-Retrieval supports fine-tuning of any open-source RAG retrieval models, including embedding models (BERT-based, LLM-based), late interactive models (ColBERT), and reranker models (BERT-based, LLM-based). - **Unified Inference Interface**: It provides a lightweight Python library, `rag-retrieval`, that simplifies the process of calling different RAG ranking models, making it easier to integrate with various applications. - **Distillation of Models**: RAG-Retrieval supports distilling models from larger to smaller versions, including support for embedding models and reranker models, facilitating efficient model size reduction. ### Tech Stack: - **Python**: The primary programming language used in the project. ### Why It Matters: RAG-Retrieval offers a unified and efficient way to fine-tune and distill RAG retrieval models, making it a valuable tool for researchers and developers looking to enhance their retrieval systems. Its ability to support a wide range of models and simplify the process of integration makes it a preferred choice for those working on retrieval tasks.

### Overview: 探索顶会论文主图设计灵感，从ICLR、ICML、NeurIPS、CVPR、ACL、AAAI等顶尖会议中精选3,516张精选图，涵盖机器人策略、工作记忆系统、LLM迎合行为分析等主题，助力科研绘图。 ### Key features: - 顶会论文主图灵感画廊 - 机器人策略与执行 - 层级工作记忆 - 工作记忆系统 - LLM迎合行为因果分析 ### Tech stack: 纯静态、零构建、纯HTML/CSS/JS，支持离线访问。 ### Why it matters: 为科研绘图提供灵感与指导，帮助研究人员快速上手，提升论文展示效果。

### Overview: LLM Sandbox is a lightweight and portable library designed to securely execute Large Language Model (LLM) generated code. It provides a common API for container and local Linux and Windows sandbox backends, allowing developers to easily run and manage code execution and file APIs. ### Key Features: - **Isolated Execution**: Container or local OS sandboxing with custom host access and security policies. - **Security Policies**: Define and enforce security rules to control code execution. - **Resource Limits**: Set CPU, memory, and execution time limits. - **Network Isolation**: Control network access for sandboxed code. - **Flexible Sandbox Backends**: Supports Docker, Kubernetes, Podman, Bubblewrap, NsJail, Windows MXC, and WSL2 Bubblewrap. - **Multi-Language Support**: Execute code in Python, JavaScript/Node.js, Java, C++, Go, R, and eleven agent frameworks. - **LLM Framework Integration**: Run examples for eleven different agent frameworks. ### Tech Stack: Python ### Why it Matters: LLM Sandbox simplifies the process of running and managing LLM-generated code, ensuring secure and isolated execution environments. Its flexibility and support for multiple languages make it a valuable tool for developers working with various programming languages and frameworks.

Overview: This repository is an implementation of the Conformer model, a novel transformer architecture designed to enhance the performance of speech recognition tasks. The model combines convolutional neural networks with transformers to capture both local and global dependencies in audio sequences, achieving state-of-the-art accuracy. Key features: - Conformer model architecture - PyTorch implementation - Integrates convolutional neural networks with transformers - Significantly outperforms previous Transformer and CNN-based models - Achieves state-of-the-art accuracy in speech recognition tasks Tech stack: Python, PyTorch Why it matters: Conformer offers a more efficient and effective way to model audio sequences by leveraging the strengths of both convolutional networks and transformers, leading to improved performance in speech recognition applications.

Overview: This project is an application tool based on Edge-Connect, designed to enhance the quality of anime images by performing inpainting and drawing functionalities. It includes features such as automatic image repair, removal of mosaics, and filling in missing parts. Key features: - Anime inpainting - Image repair - Mosaic removal - Missing pixel filling Tech stack: Python Why it matters: This tool offers a convenient way to enhance and improve anime images, making them look more like the original artworks. It is useful for anyone who wants to create high-quality anime content or improve existing images. Repository: youyuge34/Anime-InPainting Primary language: Python GitHub stars: 1132 GitHub description: An application tool of edge-connect, which can do anime inpainting and drawing.

Overview: This project provides a PyTorch implementation of a video classification tool using a 3D ResNet, trained on the Kinetics dataset, with outputs for class scores and features. Key features: - Video classification using 3D ResNet - Outputs class scores and features for each 16 frames - Supports both score and feature modes Tech stack: PyTorch Why it matters: This project showcases the potential of 3D CNNs in video classification, potentially improving upon existing methods by leveraging spatiotemporal information. Repository: kenshohara/video-classification-3d-cnn-pytorch Primary language: Python GitHub stars: 1131

Overview: The Alan AI SDK for Cordova allows developers to quickly create AI agents for their Android applications, enabling human-like conversations and actions through voice commands. Key features: - Embedding an AI agent into Android applications - Generating both business logic and UI in real time - Creating a safe and validated environment for accurate code generation - Responding, evolving, and scaling automatically Tech stack: Ruby Why it matters: This repository provides the Alan AI SDK for Cordova, enabling developers to integrate AI-driven interfaces into their existing apps in days, not months. GitHub stars: 1130