Projects

Overview: LangForge is an open-source toolkit designed to simplify the creation and deployment of LangChain applications. Key features: - Simplified environment setup and API key management - Predefined notebooks for various use cases to help you get started quickly - Instantly chat with your chains using the Jupyter integration - Automatic REST interface generation for your app Tech stack: Jupyter Notebook Why it matters: LangForge streamlines the process of developing LangChain applications, making it easier and more efficient for developers to create and deploy applications quickly and effectively.

Overview: This project is a comprehensive suite of document management tools designed for developers, offering features such as version control, collaboration, and project management functionalities. Key features: - Version control for tracking changes in documents - Collaborative editing capabilities for real-time updates - Project management tools for tracking progress and assigning tasks Tech stack: C# Why it matters: This project is crucial for developers looking to streamline their workflow and ensure the quality of their work by providing robust tools for version control, collaboration, and project management. Repository: EvotecIT/PSWriteOffice GitHub stars: 165

### Overview WebLINX is a benchmark designed to evaluate the performance of web navigation agents with conversational capabilities. It serves as a real-world dataset for researchers and developers to test and improve their models. The benchmark includes a comprehensive dataset and a Python library, making it accessible for both academic and practical applications. ### Key Features - **WebLINX Dataset**: A real-world dataset containing web navigation interactions with conversational agents. - **Python Library**: A library that provides an API to interact with the WebLINX dataset, making it easy to train and evaluate models. - **Colab Notebook**: A Jupyter notebook that demonstrates how to use the library and run experiments. - **GitHub Repository**: A public repository on GitHub where the code, models, and dataset are hosted. ### Tech Stack - Python ### Why It Matters WebLINX is crucial for researchers and developers looking to improve their conversational web navigation agents. It provides a realistic benchmark for evaluating model performance, allowing for the discovery of new techniques and the comparison of existing methods. The dataset and library make it easier to experiment and iterate on models, contributing to the advancement of natural language processing and web interaction technologies. ### Repository - **Repository**: McGill-NLP/weblinx - **GitHub Stars**: 164 - **GitHub Description**: WebLINX is a benchmark for building web navigation agents with conversational capabilities. It serves as a real-world dataset for researchers and developers to test and improve their models. - **Website**: [WebLINX](https://mcgill-nlp.github.io

### Overview: Revideo is a self-hostable Next.js app designed for AI-assisted short-form video creation, offering features such as scene-based scripting, multi-track editing, and previewing in the browser. It supports various social media platforms and YouTube Data API for content discovery. ### Key Features: - **Scene-Based Scripting:** Turn a topic into a short script with scenes, narration, and suggested visuals. - **Multi-Track Timeline:** Edit a project in a multi-track timeline and preview it in the browser. - **Layout Options:** Choose portrait, landscape, or square layouts, caption styles, uploaded music, and safe-zone guides. - **Project Management:** Save and reopen projects with a PostgreSQL database managed through Prisma. - **YouTube Integration:** Search YouTube through the YouTube Data API and search TikTok or Instagram through MuAPI. - **Remix Script Generation:** Generate remix script drafts from discovered videos. ### Tech Stack: - **Frontend:** Next.js 16, React 19, Tailwind CSS 4, Remotion Player 4 - **Backend:** PostgreSQL with Prisma 7, NextAuth with Google OAuth - **Database:** PostgreSQL with Prisma 7 ### Why It Matters: Revideo simplifies the process of creating short-form videos, making it accessible to developers and content creators. It integrates seamlessly with popular social media platforms and YouTube, ensuring a consistent and polished look across different platforms. The ability to generate and remix scripts from discovered videos adds value by providing fresh, relevant content ideas.

Overview: Helix Work is a versatile AI workspace designed for macOS, offering a unified environment for tasks, project management, and AI-driven document creation. It supports multiple AI providers and models, allowing for seamless integration across different platforms. Key features: - Run AI-assisted tasks against local project workspaces. - Access project files, integrated terminal, and browser-based research. - Extend workflows with Skills, Plugins, and MCP-compatible integrations. - Connect to leading AI providers and switch between models. - Review task progress and generated outputs before using or sharing them. Tech stack: macOS, Universal Why it matters: Helix Work provides a comprehensive and flexible workspace for AI-driven tasks, making it ideal for developers and professionals looking to streamline their workflow and access a wide range of AI models. Repository: advai-x/helix-work GitHub stars: 150

Overview: This repository contains an Open-Source RAG application built with TypeScript that integrates with the OpenAI API, supports 100+ LLM providers, and is Dockerized with full type-checking and 100% test coverage. Key features: - 🏎️ FastAPI for building REST APIs - 💻 Typer for command-line interfaces - 🍓 LiteLLM for calling 100+ LLM providers - 🔌 Langfuse for LLM observability - 🔍 Qdrant for vector database - 🚚 UV for project and dependency management - 🏍️ Redis for semantic caching - 🧹 Ruff for linter and formatter - ✅ Mypy for static type checking - 📍 Pydeps for dependency graph generation - 🚀 Pytest for testing framework - 🏗 Testcontainers for integration tests - 📏 Coverage for code coverage - 🗒️ Marimo for next-gen notebook/scripting - 👟 Just for task runner - 🐳 Docker for containerizing the Python application - 🐙 Compose for container orchestration Tech stack: TypeScript Why it matters: This repository provides a comprehensive reference implementation for AI applications that follow best practices, including type-checking, test coverage, and support for multiple LLM providers. It is designed to be easily Dockerized and integrated with Langfuse for observability, making it ideal for learning and development purposes.

### Overview JavAI Workflow, now known as Langchain4j-workflow, is a dynamic, stateful workflow engine tailored for Java developers. It offers granular control over orchestrated workflows as a graph, iteratively and with flexibility. This engine simplifies the creation of sophisticated AI applications, such as RAG-based systems, by enabling custom behavior and reducing hallucinations and response reliability. ### Key Features - **Java-based Workflow Engine**: Built as a Java library, it provides a robust and flexible workflow management system. - **Stateful Bean Management**: Ensures that stateful beans are updated in the correct order, maintaining the integrity of the workflow. - **Conditional Decision Making**: Allows for conditional transitions within the workflow, enhancing the decision-making process. - **Modular Design**: Modules can be reused across different workflows, promoting modularity and reusability. - **Streamlined Workflow Creation**: Simplifies the creation of complex workflows by encapsulating nodes and modules. ### Tech Stack Java ### Why It Matters JavAI Workflow empowers java developers to build and orchestrate sophisticated AI applications efficiently. It reduces the risk of hallucinations and improves the reliability of responses, making it a valuable tool for developers working on AI and RAG systems.

### Overview: KnowledgeOps Agent is a production-oriented Spring AI platform prototype designed for enterprise-level RAG (Relevance and Query Analysis) and knowledge operations. It leverages agent workflow engines, hybrid retrieval methods, and tenant isolation to ensure robust and scalable solutions. ### Key Features: - **Agent Workflow Engine**: Enables efficient agent workflows. - **Hybrid Retrieval**: Combines vector, keyword, graph, and web-based retrieval methods. - **Knowledge Graph**: Supports knowledge graph integration. - **Long/Short-term Memory Persistence**: Ensures data persistence across sessions. - **DeepResearch**: Utilizes deep learning for advanced research capabilities. - **Tenant Isolation**: Provides isolation for different tenants. - **Security**: Implements JWT/RBAC for secure access and tenant isolation. - **Observability**: Offers comprehensive monitoring and observability via Prometheus, Loki, and Tempo. ### Tech Stack: - **Java**: The primary programming language for development. ### Why It Matters: KnowledgeOps Agent is a production-oriented platform prototype that provides enterprise-level RAG and knowledge operations. It ensures robust and scalable solutions through tenant isolation, secure access with JWT/RBAC, and comprehensive observability. The platform is designed to be easily deployable, manageable, and verifiable, making it a valuable tool for businesses looking to enhance their knowledge operations and AI capabilities.

Overview: pdf-mcp is a MCP server designed for Claude Code and other AI agents, allowing them to search, read, and analyze large PDFs in a single tool. Key features: - Hybrid semantic search - Keyword search - Selective page reads - Tables, images, OCR, and multi-column/CJK layouts - CUDA acceleration for large corpora warming Tech stack: Python 3.11+ Why it matters: pdf-mcp simplifies the process of analyzing and summarizing large PDFs, making it easier for AI agents to access and understand information efficiently.

Overview: This project introduces Sentiment Reasoning for Healthcare, a new task that involves predicting sentiment labels and generating rationale explanations for them. The study focuses on improving model transparency by providing human-level rationales for LLM predictions. Key features: - Sentiment Reasoning for both speech and text modalities - A multimodal multitask framework - The world's largest multimodal sentiment analysis dataset Tech stack: Jupyter Notebook Why it matters: Sentiment Reasoning helps users understand LLMs' reasoning, leading to better decision-making in healthcare applications. The project demonstrates a significant improvement in model performance, achieving a 2% increase in accuracy and macro-F1, with no significant difference in semantic quality between human and ASR transcripts. Repository: leduckhai/Sentiment-Reasoning Primary language: Jupyter Notebook GitHub stars: 165

### Overview: Brainwave is a modern, animated SaaS Generative AI Landing Page designed to showcase your product or service. It provides a visually engaging and user-friendly interface, perfect for attracting and converting potential clients. ### Key Features: - **Animated Design**: Brainwave features an animated design to keep users engaged and attract attention. - **Responsive Layout**: The landing page is optimized for various devices, ensuring a seamless experience across all screen sizes. - **Customizable Header**: Users can customize the header with their company logo and tagline, making the landing page more personalized. - **Dynamic Benefits Section**: Displays key benefits of your product or service in an interactive and engaging manner. - **User-Friendly Navigation**: A clean and intuitive navigation system to help users easily find the information they need. - **Pricing and Roadmap**: Offers a clear view of your pricing plans and upcoming roadmap, helping potential clients make informed decisions. - **Social Proof**: Incorporates social media icons and testimonials to build trust and credibility. - **Optimized for Netlify**: Deployed on Netlify for a fast and reliable hosting service. ### Tech Stack: Brainwave is built using the following technologies: - **JavaScript**: The primary programming language used for the frontend. - **React**: Used for building the frontend components. - **Vite**: A modern, efficient build tool that helps in speeding up the development process. - **Tailwind CSS**: A utility-first CSS framework that helps in creating responsive and consistent designs. - **Netlify**: A platform for deploying websites, ensuring a fast and reliable hosting service

### Overview: 这个项目是基于chatglm-6b微调的LORA和PPO模型，用于处理整数和小数的加减乘除运算。模型支持GPU和CPU，并且训练和推理数据集来自自动生成的整数/小数加减乘除运算样本。 ### Key features: - 支持GPU和CPU训练和推理。 - 使用LORA和PPO进行微调和推理。 - 支持自动生成的整数/小数加减乘除运算样本。 - 支持半精度float16和Post-LN + DeepNorm的计算方法。 ### Tech stack: Python ### Why it matters: 这个项目是基于chatglm-6b进行微调和推理的，适用于处理整数和小数的加减乘除运算。它使用了LORA和PPO进行微调和推理，支持半精度float16和Post-LN + DeepNorm的计算方法。

### Overview: Transtation is a translation app built with Kotlin Multiplatform and Compose Multiplatform, providing a seamless translation experience with advanced large language models. ### Key features: - **Multi-Engine Translation**: Supports translation using multiple engines (Bing, Google, Baidu, Tencent). - **LLM Integration**: Utilizes advanced Large Language Models (LLMs) like ChatGPT, GPT-4, ChatGLM, Qwen for accurate translations. - **Long Text Translation**: Translates very long text using LLMs for a smooth experience. - **Plugin Development**: Supports plugin development, allowing easy updates and management. - **Image and Screen Translation**: Translates images and supports screen translation for various use cases. - **Customization**: Offers customization options for themes, languages, and engines. - **Scalability**: High scalability for plugin development and management. ### Tech stack: Kotlin Multiplatform + Compose Multiplatform ### Why it matters: Transtation offers a powerful and user-friendly translation experience, leveraging advanced AI models to provide accurate and efficient translations. The app is designed with Kotlin and Compose to ensure seamless performance and scalability. Repository: FunnySaltyFish/Transtation-KMP Primary language: Kotlin GitHub stars: 163 GitHub description: A translation app on Android/Desktop built by Kotlin Multiplatform + Compose Multiplatform, enjoy amazing experience with LLMs' support.

Overview: AKB is an organizational knowledge base designed for AI agents, providing a unified platform for storing and accessing documents, tables, and files through a URI graph. It is built on PostgreSQL and serves over the Model Context Protocol (MCP), allowing direct communication with agents. Key features: - A unified URI graph for documents, tables, and files - Hybrid semantic and keyword search capabilities - Support for structured tables and files - Direct communication with MCP-aware agents via POST /mcp/ and MCP OAuth Resource Server Tech stack: Python Why it matters: AKB simplifies the management of organizational knowledge for AI agents, providing a seamless and efficient way to store and retrieve information.

### Overview: Basjoo is a self-hosted AI customer-support platform that transforms a website and set of documents into a fully embeddable support agent. It integrates with Qdrant for vector search, multiple LLM providers for chat, and a fastAPI backend for agents and knowledge bases. The platform is built with Python, FastAPI, Vite, SQLite, and Docker Compose, making it scalable and secure. ### Key Features: - **Knowledge Bases as First-Class Citizens**: Agents and knowledge bases can be created, reset, or deleted independently. - **Two Ingestion Paths**: Crawl one page, crawl a site, or upload files for a KB. - **Per-KB Embeddings**: Supports OpenAI-compatible, OpenAI native, Anthropic, Google Gemini, xAI, OpenRouter, z.ai, Volcengine, Moonshot, Alibaba Bailian, and Tencent Cloud LLMs. - **Streaming Chat with SSE**: Provides source citations, keepalives, rate limits, and restricted replies. - **Thirteen LLM Providers**: Supports OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, z.ai, Volcengine, Moonshot, Alibaba Bailian, Tencent Cloud, and SiliconFlow. ### Tech Stack: - **Python**: Backend and frontend development. - **FastAPI**: Backend API. - **Vite**: Frontend development. - **SQLite**: Local storage for agent data. - **Qdrant**: Vector search engine. - **Docker Compose**: Container orchestration. ### Why It Matters

Overview: This project computes GMM-based GOP using Kaldi, a popular speech processing toolkit. Key features: - Computes GOP using GMM-based approach - Supports DNN-based implementation for potentially better performance Tech stack: C++ Why it matters: This implementation provides a GMM-based solution for GOP computation, offering a straightforward and widely-used approach for speech quality assessment.

Overview: VectorRAG.Net is a high-performance .NET-native vector database library designed for semantic search and Retrieval-Augmented Generation (RAG), offering a fast and efficient solution for local AI assistants and low-latency applications. It eliminates the need for external services, providing a fully embedded architecture with no HTTP, Docker, or network latency. Key features: - SIMD-optimized vector search - LSH-based ANN indexing - Hybrid Search (Vector + BM25) - Automatic document chunking - Metadata filtering - Snapshot save/load - Runtime metrics - Fully embedded architecture - .NET 8 native implementation Tech stack: C# Why it matters: VectorRAG.Net simplifies the process of integrating semantic search capabilities into .NET applications, offering a fast, efficient, and secure solution without the need for external services, making it ideal for low-latency applications and environments with restricted network access. Repository: likeslines-maker/VectorRAG.Net GitHub stars: 164

### Overview: This project investigates the emergence of the "attention sink" phenomenon in language models (LMs) during pre-training. The attention sink refers to a situation where the model's ability to focus on specific parts of the input text diminishes over time, leading to a loss of context and understanding. The study analyzes how various factors such as optimization techniques, data distribution, loss functions, and model architectures influence this phenomenon. ### Key features: - Comprehensive evaluation of attention sink in open-sourced LMs. - Script to calculate attention sink metrics for LLaMA2, LLaMA3, Mistral, GPT2, Pythia, and OPT. - Pre-training of LLaMA models on 5B token data to understand pre-training effects. - Detailed documentation and scripts for reproducibility and ease of use. ### Tech stack: Python ### Why it matters: Understanding the attention sink phenomenon is crucial for improving the performance and effectiveness of language models. By analyzing how different pre-training methods and model architectures impact this phenomenon, researchers can develop more robust and context-aware language models.

### Overview: Lucid Memory is a memory system designed specifically for AI applications, offering unparalleled speed, cost-effectiveness, and context-aware retrieval. ### Key Features: - **2.7ms Retrieval**: Lucid Memory is 50 times faster than traditional RAG systems. - **5x Token Efficiency**: It uses 5 times fewer tokens to achieve the same recall rate. - **307 Tests, 0 Tsc Errors**: Continuous improvement with new features like Memory Consolidation and Episodic Memory. - **Zero Cloud, Zero Cost**: No need for cloud storage, and costs are minimal. ### Tech Stack: TypeScript ### Why It Matters: - **Context Understanding**: Unlike traditional RAG, Lucid Memory understands context and context-aware retrieval. - **Performance**: 50% more relevant context per token, 10-20x faster, and $0 cost. - **Recency Awareness**: Recent-but-irrelevant never beats old-but-relevant, unlike systems where recency overwhelms similarity. - **Privacy**: Keeps your code private, unlike sending data to the cloud. ### Tech Stack: TypeScript ### GitHub Stars: 163 ### Primary Language: TypeScript ### Why Lucid Memory? - **Not a Vector Database**: Lucid Memory is the retrieval layer that makes vector databases obsolete for AI memory. - **Self-Maintaining**: Memory Consolidation ensures that recent memories are strengthened, stale ones decay, and weak associations are pruned. - **Zero Cloud, Zero Cost**: No need for cloud storage, and

Overview: VectifyAI/jev-doc-search is a project that integrates Jev's `Choice` feature with PageIndex to perform long-document search, addressing limitations of flat page search on large documents. Key features: - Uses PageIndex to convert a 300-page report into a hierarchical tree structure. - Implements flat page search using Jev's `Choice` feature. - Solves the limits of flat page search by breaking down the document into sections and pages, making it suitable for long documents. Tech stack: Python Why it matters: This project provides a practical solution for searching long documents by leveraging Jev's `Choice` feature with PageIndex to handle the limitations of flat page search, making it suitable for employee handbooks, annual reports, and other large documents.