Projects

### Overview: epub2tts is a Python application designed to convert EPUB or text files into audiobooks using realistic text-to-speech technology. It supports multiple engines and GPU capabilities, making it versatile and highly customizable. ### Key Features: - **Standard Format M4B Audiobook File Creation**: Converts EPUB or text files into a standard format M4B audiobook file. - **Automatic Chapter Break Detection**: Automatically identifies and separates chapters in the input file. - **Embedding Cover Art**: Option to include cover art from the EPUB file. - **GPU Support**: Utilizes GPU for faster processing, especially with the Coqui AI TTS engine. - **Voice Cloning with Coqui XTTS Model**: Allows for seamless voice cloning with the Coqui AI TTS engine. - **58 Studio Quality Voices**: Offers a wide range of high-quality voices from Coqui AI. - **Deepspeed Support**: Optimized for faster processing with the availability of deepspeed. - **Resumable Processing**: Automatically resumes where it left off if interrupted. - **DRM-Free EPUB Requirement**: EPUB files must be DRM-free for the application to work. ### Tech Stack: Python ### Why it Matters: epub2tts is a free and open-source tool that simplifies the process of converting books into audiobooks, offering a variety of high-quality voices and a GPU-friendly option for faster processing. It's particularly useful for those looking to create audiobooks without the need for expensive hardware or specialized software.

Overview: VoiceStreamAI is a Python-based server and JavaScript client solution designed for near-realtime audio streaming and transcription using Whisper and WebSocket. It leverages Huggingface's VAD and OpenAI's Whisper for accurate speech recognition and processing. Key features: - Real-time audio streaming through WebSocket. - Modular design for easy integration of different VAD and ASR technologies. - Factory and strategy pattern implementation for flexible component management. - Unit testing framework for robust development. - Customizable audio chunk processing strategies. - Support for multilingual transcription. - Secure Sockets with optional cert and key file arguments. Tech stack: Python/JS Why it matters: VoiceStreamAI simplifies the process of real-time audio transcription, making it accessible and efficient for developers and businesses looking to integrate speech recognition into their applications.

Overview: Agenttrail is an open-source project designed to provide local observability for AI coding agents, offering two views: Agenttrail Map for project structure and progress, and Agenttrail Kitchen for visualizing tasks and role contributions in a 3D environment. Key features: - Agenttrail Map visualizes project structure and activity. - Agenttrail Kitchen visualizes tasks and role contributions in a 3D environment. - It helps you follow project structure, progress, and which components are changing. - It records agent activity, including file changes and optional Claude Code hooks. - It provides a shared cooking scene for visualizing work as chefs, order tickets, and deliveries. - It can be run in your repo with `npx agenttrail-kitchen .` or `npx agenttrail` for the Map. - It does not run agents, assign work, or decide tasks are finished. Tech stack: JavaScript Why it matters: Agenttrail offers a simple and intuitive way to observe and manage AI coding agents, providing a live view of their work without the need for multiple terminal windows.

### Overview: AgileRL is a comprehensive library for post-training fine-tuning and reinforcement learning (RL) of large language models (LLMs). It seamlessly transitions from supervised fine-tuning to preference tuning and multi-turn agentic RL, with built-in evolutionary hyperparameter optimization. The library supports a wide range of environments, including OpenEnv, and leverages advanced techniques like LoRA, QLoRA, and parallel training to optimize performance and memory usage. ### Key Features: - **Multi-turn Agentic RL**: Train on any OpenEnv environment, including Python classes, library entrypoints, or remote environments. - **LoRA and QLoRA**: Train small adapters on a frozen base, matching full fine-tuning at a fraction of the memory. - **Torch-native Parallelism**: Data parallel or FSDP2 sharding, with a single `FSDPConfig` for every LLM algorithm. - **Memory Optimizations**: Chunked fused log-probs, activation checkpointing, offloading, and trainer offload during rollout. - **Model-Specific Optimizations**: Per-family kernels and settings for high throughput at the lowest memory, on dense, MoE, and hybrid Mamba models. - **Fast Rollouts**: Optimized for efficient training and inference. ### Tech Stack: Python ### Why It Matters: AgileRL simplifies the post-training fine-tuning and RL process for large language models, making it easier to scale up and deploy models effectively. The library's flexibility and advanced optimizations ensure optimal performance and memory usage, making it a valuable tool for

Overview: DISC-LawLLM is an intelligent legal system powered by large language models (LLMs) designed to provide a wide range of legal services. It aims to offer professional, intelligent, and comprehensive legal support. Key features: - 法律文本处理能力：微调数据集构建，增强法律语言理解与生成能力。 - 法律推理思维能力：设计法律三段论指令数据，提高法理推理能力。 - 司法领域知识遵循能力：增强系统对背景知识的检索和遵循能力。 Tech stack: Python Why it matters: DISC-LawLLM is a significant contribution to the field of legal AI, leveraging large language models to enhance legal services, making it accessible and intelligent for various legal professionals and users.

### Overview RAG-Fusion is a search methodology that leverages multiple query generation and Reciprocal Rank Fusion to enhance retrieval-augmented generation. It aims to surface relevant material beyond what a single query would typically find, particularly when the user's vocabulary doesn't match the corpus's indexing. ### Key Features - **Query Generation:** An LLM generates multiple rewrites of the user's query. - **Hybrid Search:** The original query and each rewrite are searched twice, once with BM25 and once with vector search. - **Reciprocal Rank Fusion:** The ranked lists from both searches are merged, and each document scores based on reciprocal rank. ### Tech Stack - **Python** ### Why It Matters Properly configured RAG-Fusion significantly improves retrieval after reranking, particularly with the strongest reranker, achieving an NDCG@10 lift of +0.025 with `bge-reranker-large`, +0.035 to +0.039 with OpenAI's Decisions API and Microsoft-Decision-1, and +0.050 with Jev. The fusion variant outperforms the vector-only variant once a reranker is added. Deploying the hybrid variant with the strongest reranker behind it is recommended for optimal results.

### Overview: Story-Iter is a training-free iterative framework designed for long story visualization, aiming to enhance semantic consistency and fine-grained interactions in narrative generation. ### Key features: - **Global Reference Cross-Attention (GRCA)**: A novel module that models all reference frames with global embeddings, ensuring semantic consistency in long sequences. - **Iterative Paradigm**: Continuously refines each generated image by incorporating all reference images from previous rounds, providing precise and fine-grained interactions. - **Training-Free**: Unlike traditional methods that rely on fixed reference images, Story-Iter uses a training-free approach to enhance long-story generation. ### Tech stack: - Python 3.10.14 - PyTorch 2.2.2 - CUDA 12.1 - cuDNN 8.9.02 ### Why it matters: Story-Iter's state-of-the-art performance in long-story visualization (up to 100 frames) excels in both semantic consistency and fine-grained interactions, making it a significant advancement in the field of narrative generation and visualization.

### Overview: Radiata is a stable diffusion webui based on diffusers, designed for efficient and high-performance rendering of images and videos. ### Key features: - Stable Siffusion - Stable Diffusion XL - Acceleration using TensorRT - ControlNet plugin - Lora & Lycoris ### Tech stack: Python ### Why it matters: Radiata provides a powerful and efficient platform for generating high-quality images and videos using the latest stable diffusion models, with optimized performance and advanced features.

### Overview: WeSight is an open-source desktop control console designed for local AI agents, providing a unified workspace for installing, running, and managing various coding agents like Claude Code, Codex, OpenClaw, and more. ### Key Features: - **Install and Reuse**: Easily install and reuse local agent CLIs with a beginner-friendly UI. - **Visual Workspace**: Run coding agents as a visual chat interface with tool panels, slash commands, and file diffs. - **IM Integration**: Connect agent tasks to IM channels like Feishu with per-engine configuration. - **Metrics Tracking**: Track every task with detailed metrics such as engine, model, token usage, and runtime performance. - **SkillHub Skills**: Extend workflows through built-in and SkillHub skills, scheduled tasks, memory, and a desktop companion. ### Tech Stack: TypeScript ### Why It Matters: WeSight simplifies the setup and management of local AI agents, making it easier for developers to discover and use the project.

Overview: Pipelock is an open-source AI agent firewall designed to secure and monitor mediated HTTP, MCP, A2A, and WebSocket traffic, as well as CONNECT tunnel contents when TLS interception is enabled. It scans for exfiltration, prompt injection, SSRF, and risky tool-call chains, and provides verifiable audit evidence through mediator-signed action receipts. Key features: - Scans mediated HTTP, MCP, A2A, and WebSocket traffic for exfiltration, SSRF, and prompt injection. - Emisses mediator-signed action receipts for verifiable audit evidence. - Scans plain CONNECT without interception at the hostname and URL level. - Configured MCP upstreams are an exception to private-address SSRF blocking. Tech stack: Go Why it matters: Pipelock is crucial for securing AI agents and ensuring compliance with Verifiable Egress Control, providing transparent and auditable evidence of network traffic. Repository: pipelab-org/pipelock GitHub stars: 954

### Overview: OpenEMMA is an open-source implementation of Waymo's EMMA model, providing an end-to-end framework for autonomous driving tasks. It leverages pretrained Vision Language Models (VLMs) to integrate text and front-view camera inputs, enabling precise predictions and decision rationales. ### Key Features: 1. **End-to-End Framework**: OpenEMMA is designed to be an open-source framework for motion planning in autonomous driving. 2. **Multimodal Integration**: It integrates text and front-view camera inputs using pretrained Vision Language Models (VLMs), such as GPT-4 and LLaVA. 3. **Precise Predictions**: OpenEMMA provides precise predictions of future ego waypoints and decision rationales, enhancing the accuracy of autonomous driving tasks. 4. **Accessible Tools**: OpenEMMA aims to provide researchers and developers with accessible tools to advance autonomous driving research and applications. ### Tech Stack: Python ### Why it Matters: OpenEMMA offers a practical solution for researchers and developers looking to advance autonomous driving technologies. By leveraging pretrained VLMs, it simplifies the integration of text and visual inputs, making it easier to create end-to-end autonomous driving systems. The inclusion of decision rationales and precise predictions ensures that the framework is not only effective but also transparent and understandable, contributing to the development of more reliable and ethical autonomous vehicles.

Overview: This project is a simple resume parser written in Python that extracts various information such as name, email, mobile numbers, skills, experience, and more from resumes in PDF and DOCX formats. Key features: - Extracts name - Extracts email - Extracts mobile numbers - Extracts skills - Extracts total experience - Extracts college name - Extracts degree - Extracts designation - Extracts company names Tech stack: Python Why it matters: This tool is useful for developers and recruiters to quickly extract essential information from resumes, making the process of job searching and candidate screening more efficient and accurate.

Overview: Rasa UI is a web application built on top of the Rasa Framework, providing a user-friendly interface for creating and managing training data, models, and stories. Key features: - UI for creating and managing training data (Examples, Intents, Entities, Synonyms, Regex, Stories, Actions, Responses) - Manage multiple bots from a single UI instance - Create, manage, and load different versions of your models for testing and optimization - Log requests for usage tracking, history, and improvements to your models - Easily execute intent parsing using different models - Data is stored in a SQLite DB for backing up/sharing - Can be used with or without a Rasa backend to manage your training data Tech stack: JavaScript Why it matters: Rasa UI simplifies the process of creating and managing Rasa bots, models, and stories, making it easier for developers to build and deploy conversational AI applications.

### Overview: Codex Slides is an open-source AI slide studio designed for coding agents. It allows you to turn a prompt, a repository, or a pile of files into a beautiful, presentation-ready deck without leaving your coding agent. ### Key features: - **10+ high-quality slides** in approximately 4-5 minutes — Fast mode renders every page in parallel. - **45 deck templates** across 15 professional categories, plus 73 community styles across 12 groups. - **24 guided scenarios** in 6 groups — zero-to-one decks, source-to-slides, data stories, deep research, redesign, localization, training, keynotes, and more. - **⚡ Fast mode** — renders every page in parallel, so a 10-slide deck lands in ~4-5 minutes. - **Production export** — real, presentation-ready **PPTX** and print-ready **PDF** with speaker notes preserved; 1K/2K/4K render quality and five aspect ratios (16:9, 4:3, 1:1, 9:16, 3:4). - **Zero extra keys** — runs on your existing `codex login`; every project stays on your disk. ### Tech stack: TypeScript ### Why it matters: Codex Slides is an open-source alternative to Gamma and Tome, designed to live inside your coding agent, without requiring any extra API keys. It allows you to create beautiful, presentation-ready decks in a matter of minutes, making it

### Overview: Local PDF Chat RAG is a transparent, runnable Python implementation designed for developers to inspect and understand the complete retrieval-augmented generation pipeline. It includes modules for document loading, chunking, embeddings, FAISS, BM25, hybrid retrieval, reranking, and answer generation. ### Key Features: - **Transparent Pipeline**: Modules follow the order of processing a RAG request. - **Hybrid Retrieval**: Combines FAISS dense retrieval with BM25 keyword retrieval. - **Optional Reranking**: Supports CrossEncoder or model-based relevance scoring. - **Multiple Model Backends**: Supports local Ollama, SiliconFlow, and OpenAI-compatible APIs. - **Document Support**: Handles PDF, TXT, Markdown, DOCX, XLS/XLSX, and PPTX formats. ### Tech Stack: Python ### Why It Matters: This repository is intended for learning and experimentation. It is not a production-ready knowledge-base service. Before using it with real business data, ensure you add authentication, tenant isolation, persistence, evaluation, security controls, and deployment governance.

### Overview: MocapNET is a real-time method that estimates 3D human pose directly from monocular color images, using a novel 2D pose representation called NSRM and a human body orientation classifier. The project has been featured in the IEEE International Conference on Computer Vision (ICCV) 2023 and successfully defended the Ph.D. thesis. ### Key Features: - Novel 2D pose representation (NSRM) - Human body orientation classifier - Real-time 3D pose estimation - One-click deployment in Google Colab ### Tech Stack: Python ### Why it Matters: MocapNET offers a compact and efficient approach to real-time 3D pose estimation, making it valuable for applications requiring real-time performance and accurate pose tracking. The method's ability to handle complex body and facial orientations is particularly noteworthy, making it suitable for various applications in computer vision and human-computer interaction.

### Overview: 本项目旨在通过Python抓取和分析微博数据，包括文本分析、可视化、情感分析等，涵盖了多种高级文本处理技术和深度学习应用。项目涉及的数据预处理、文本聚类、LDA主题模型分析、情感分析（词典和W2V+LSTM）等技术，旨在为用户提供全面的微博文本分析解决方案。 ### Key features: - **文本分析与可视化:** 使用LDA、树图、关系图、词云、时间趋势、热度地图、词典情感分析、词向量情感分析、tfidf聚类、词向量聚类、关键词提取、文本相似度分析等技术。 - **数据预处理:** 包括数据清洗、预处理、情感分析、词向量转换等。 - **深度学习应用:** 使用LSTM、W2V等技术进行情感分析。 ### Tech stack: Python ### Why it matters: 该项目为用户提供了一个全面的微博文本分析工具，涵盖了从数据抓取到文本处理、情感分析、主题模型分析等多步骤，适用于各种文本分析和可视化需求，尤其适合需要深度学习技术的应用场景。 Repository: HUANGZHIHAO1994/weibo-analysis-and-visualization Primary language: Python GitHub stars: 955 GitHub description: 使用Python抓取微博数据并对微博文本分析和可视化，LDA（树图）、关系图、词云、时间趋势（折线图）、热度地图、词典情感分析（饼图和3

### Overview: AgentDojo is a dynamic environment designed to evaluate prompt injection attacks and defenses for language model agents. It provides a platform to assess the effectiveness of various strategies in mitigating such threats. ### Key Features: - **Dynamic Environment**: Allows for real-time evaluation of agent performance under different attack scenarios. - **Flexible Suite**: Supports multiple tasks and suites to cover a wide range of scenarios. - **Defense Strategies**: Enables the testing of various defense mechanisms against prompt injection attacks. - **Model Compatibility**: Compatible with multiple language models, including GPT-4. ### Tech Stack: - Python ### Why It Matters: AgentDojo offers a practical tool for researchers and developers to evaluate and improve the robustness of language models against prompt injection attacks, thereby contributing to the advancement of AI security and defense mechanisms.

Overview: AgentKit is a TypeScript framework designed for building multi-agent networks, offering deterministic routing and comprehensive tooling for efficient development. Key features: - Deterministic routing - Rich tooling - Easy integration with MCP Tech stack: TypeScript Why it matters: AgentKit simplifies the process of developing complex multi-agent systems, ensuring seamless communication and efficient management of agents.

### Overview: Octocode is a powerful tool for AI agents and developers, designed to research and understand code across multiple repositories. It leverages ripgrep and AST search to provide precise and efficient results, making it a valuable resource for research and development. ### Key features: - **Code Research**: Octocode analyzes your local code and external repositories like GitHub, npm, and more. - **Context Understanding**: It helps developers understand the context of code changes and interactions. - **Token Efficiency**: The Rust engine ensures fast and token-efficient results, making it suitable for large repositories. ### Tech stack: Rust ### Why it matters: Octocode is essential for researchers and developers looking to quickly find and understand context in their code. Its ability to handle large repositories and provide precise results makes it a must-have tool in the research toolkit. Repository: bgauryy/octocode GitHub stars: 955