Projects

Overview: STAPI is an OpenAI-compatible embedding API server that uses Sentence Transformers for embeddings. Key features: - Supports Sentence Transformers for embeddings - Compatible with OpenAI API - Docker installation option - Local Python installation option - Embedding API for text strings - Preloaded models for quick access Tech stack: Python Why it matters: STAPI provides an OpenAI-compatible embedding API, making it useful for applications requiring high-quality text embeddings, such as search engines, chatbots, and more.

Overview: Synapsy Write is a Next.js web application that uses advanced AI models to assist users in generating text content, including paragraphs, blog posts, essays, and more. It leverages models from providers like OpenAI, Mistral, and Anthropic, ensuring high-quality and contextually accurate text. Key features: - **Advanced AI Models**: Synapsy Write uses models from multiple providers to generate high-quality text. - **Versatile Output**: Users can create paragraphs, blog posts, essays, and more with just a few clicks. - **Custom Templates**: Templates allow users to create and reuse custom document structures. - **Conversation Mode**: A conversational interface for interactive content generation. - **User-Friendly Interface**: Easy to use, suitable for both beginners and advanced users. Tech stack: TypeScript Why it matters: Synapsy Write offers a streamlined and efficient way to generate text, making it a valuable tool for writers and content creators.

Overview: AI Health Steward is a self-hosted, private AI health steward that integrates multimodal AI structures, health data into a person-centric profile with a dashboard and AI consultation. It allows users to control their data and AI models, ensuring data is never uploaded to third-party servers. Key features: - Data is stored on users' own servers, not on any SaaS database. - Supports a personalized health profile for each family member. - Allows users to choose their own AI models, whether on the cloud or locally. - Enables mobile access via Capacitor, with a secure pairing code for public network access. Tech stack: Python 3.12, FastAPI, React 18, PostgreSQL 16, pgvector, OpenAI-compatible APIs. Why it matters: It provides a secure, private, and customizable solution for managing health data and AI interactions, allowing users to control their data and AI models without exposing it to third-party services. GitHub stars: 69 Repository: wangzhengpengjay/AI-Health-Steward

### Overview: WhisPaste is a free, open-source desktop application that enables users to speak their words directly into any application without typing. It utilizes Whisper, a whisper.cpp-based voice recognition engine, to convert spoken text into written content. The app is designed for Windows, macOS, and Linux, offering offline functionality and the option to use either Whisper or NVIDIA Parakeet for transcription. It supports multiple languages and provides a seamless user experience with features like hotkey activation, real-time waveform display, and customizable settings. ### Key Features: - **Hotkey Activation**: Press `Ctrl+Shift+D` by default to activate the speech-to-text feature. - **Push-to-Talk**: Speak naturally, and the text appears at your cursor. - **Transcription Engines**: Supports Whisper (99 languages) and NVIDIA Parakeet (~25 languages) with an optional CUDA pack for NVIDIA GPUs on Windows and Linux. - **Offline Functionality**: Transcription is performed locally, ensuring no cloud storage or subscription is required. - **Silence Trimming**: Automatically trims the end of recordings to improve accuracy. - **Smart Mode**: Optional AI post-processing for cleanup, shortening, and translation into 7 languages. - **Voice Snippets**: Supports interactive multi-field templates for text expansion. - **Quick-Paste Panel**: A side panel for quick pasting of selected text. ### Tech Stack: The app is built using Dart, a programming language designed for building native Android and iOS apps, web apps, and desktop applications. ### Why It Matters: WhisPaste offers a

Overview: 一句代码搞定语音合成，文字转语音 Key features: - 文字转语音功能 - 支持多种语言 - 简单易用的API接口 Tech stack: Java Why it matters: 提升开发效率，简化语音合成流程，适用于各种应用场景。

### Overview: OmniRoute AI Explorer is a lightweight, unified dashboard and reporting tool that simplifies access to 237+ AI providers through a single endpoint. It generates code, chats with models, and automates beautiful Markdown reports, saving up to 95% of tokens using OmniRoute's built-in compression. ### Key Features: - **Unified Dashboard**: Access to multiple AI providers through a single endpoint. - **Code Generation**: Automatically generate code snippets. - **Model Chatting**: Interact with models directly. - **Token Savings**: Utilizes OmniRoute's built-in compression to save up to 95% of tokens. ### Tech Stack: - **Python**: Core logic and scripting. - **Streamlit**: Fast, responsive frontend for the visual AI explorer. - **OpenAI SDK**: Native communication with OmniRoute via its OpenAI-compatible `/v1` endpoint. ### Why It Matters: OmniRoute AI Explorer simplifies the process of accessing and utilizing multiple AI providers, making it easier to test models, generate code, and automate reports while significantly reducing token usage. ### Tech Stack: - **Python**: Core logic and scripting. - **Streamlit**: Fast, responsive frontend for the visual AI explorer. - **OpenAI SDK**: Native communication with OmniRoute via its OpenAI-compatible `/v1` endpoint. ### 5 Use Cases: 1. **Unified AI Benchmarking**: Test different models on the same prompt. 2. **Automated Code Reviews**: Generate PR reviews in a CI/CD pipeline. 3. **Token-Optimized Ch

Overview: Vortex is a Python-based platform that enables AI agents to discover and optimize sparse attention algorithms, significantly enhancing the performance of large language models without altering core model code. This makes it an invaluable tool for researchers and developers alike. Key features: - Easy Programming: Program sparse attention with a PyTorch-like frontend, simplifying the process. - High Performance: Built to work efficiently with FlashInfer, CUDA Graph, and Radix Attention. - Agent Native: Designed for autonomous algorithm discovery, with a Claude Code workspace and OpenHands demo included. Tech stack: Python, SGLang, FlashInfer, CUDA Graph, Radix Attention. Why it matters: Vortex turns sparse attention algorithm design into a practical and scalable task, allowing AI agents to explore the design space more efficiently and effectively, leading to improved performance and discovery of new algorithms.

Overview: This project presents an efficient and scalable approach for generating point cloud data using sparse point-voxel diffusion models. The implementation is based on the paper "Efficient and Scalable Point Cloud Generation with Sparse Point-Voxel Diffusion Models" and has been accepted for publication in IEEE Transactions on Neural Network and Learning Systems (TNNLS). Key features: - Implementation of the SPVD model for efficient point cloud generation - Pretrained checkpoint for point cloud completion - Gradio app for real-time completion and super-resolution tasks - Generation checkpoints for all model variants - Release of the SPVD Lightning implementation Tech stack: Jupyter Notebook Why it matters: This project offers a practical solution for generating high-quality point cloud data efficiently, which is crucial in various applications such as computer vision, robotics, and virtual reality. The use of sparse point-voxel diffusion models ensures scalability and speed, making it suitable for large-scale applications. Repository: JohnRomanelis/SPVD GitHub stars: 68 GitHub description: SPVD: Efficient and Scalable Point Cloud Generation with Sparse Point-Voxel Diffusion Models

### Overview: Quote2Image is a Python library designed to convert text quotes into graphical images. It simplifies the process of creating visually appealing illustrations of quotes, making it easy for developers and designers to incorporate these elements into their projects. ### Key Features: 1. **Text to Image Conversion**: Converts text quotes into graphical images. 2. **Customizable Options**: Allows customization of text color, background color, font, and more. 3. **Flexible Backgrounds**: Supports generating and using custom background images with adjustable brightness and blur settings. 4. **Watermark Functionality**: Adds a customizable watermark to the generated image. ### Tech Stack: - Python ### Why It Matters: Quote2Image streamlines the process of turning text into visual content, making it an essential tool for developers looking to add dynamic, engaging elements to their projects. Its flexibility and customization options make it a versatile solution for various applications, from web design to marketing materials. ### GitHub Stars: 69 ### Installation: ```bash pip install Quote2Image ``` ### Usage: The `Convert` function is the primary method for generating images. It accepts various parameters such as the quote, author, text color, background color, font, and size, among others. Here’s a basic example: ```python from Quote2Image import Convert, GenerateColors # Generate Fg and Bg Color fg, bg = GenerateColors() img = Convert( quote="Pooing keeps you healthy", author="Pee", fg=fg, bg=bg, font_size=3

### Overview: OpenBG-IMG 是一个用于电子商务领域的多模态数据集基准，旨在为[CCKS2022面向数字商务的知识处理与应用评测任务三：多模态商品知识图谱链接预测](https://tianchi.aliyun.com/competition/entrance/531957/introduction)提供评测结果。基准使用了多种知识图谱嵌入模型进行链接预测，旨在为相关研究提供参考。 ### Key features: - **Baseline Models**: OpenBG-IMG基准使用了多种知识图谱嵌入模型，包括TransE、TransH、TransD、DistMult、ComplEx、TuckER、TransAE和RSME。 - **Data Set**: 包含27,910个实体和136个关系，数据集包含图片和多模态实体，用于生成评测结果。 - **Environment Configuration**: 使用Python代码进行环境配置，包括安装依赖库和数据集的准备。 ### Tech Stack: - **Python**: 代码编写和环境配置主要使用Python语言。 ### Why it Matters: OpenBG-IMG基准为电子商务领域的多模态知识图谱链接预测任务提供了一个全面的评测基准，有助于研究者评估和改进现有模型的性能。通过基准的公开提供，促进了知识图谱和多模态数据处理领域的研究和应用。 ### Repository: OpenBGBenchmark/OpenBG-IMG - **GitHub Stars**: 69 - **Primary Language**: Python - **GitHub Description

### Overview: DeepSlide is a delivery-first, human-in-the-loop system designed to enhance presentation delivery beyond just slide creation. It supports the entire workflow of presentation preparation and delivery, including narrative planning, logical chain editing, slide generation, script drafting, interactive enhancement, and rehearsal feedback. ### Key Features: - **Narrative Planning**: Generates multiple time-budgeted logical-chain candidates and supports node-level editing and emphasis allocation. - **Logical Chain Editing**: Enhances the logical flow of the presentation through interactive enhancements and attention control. - **Script Drafting**: Produces both `recipe/content.tex` (slides) and `recipe/speech.txt` (script) for a seamless co-delivery. - **Interactive Enhancement**: Enhances the presentation with interactive elements to guide the audience's attention. - **Rehearsal Feedback**: Provides rehearsal feedback to ensure the presentation is ready for delivery. ### Tech Stack: - Python ### Why It Matters: DeepSlide is designed to improve the quality of presentations by focusing on the delivery aspect rather than just the artifact quality. It ensures that information is organized and delivered effectively, considering audience cognition and attention constraints. This makes DeepSlide a valuable tool for anyone looking to enhance their presentation skills and effectiveness.

### Overview: HBFSim is an open-source High-Bandwidth Flash (HBF) simulator designed to evaluate the performance of LLM workloads on real hardware. It runs real LLM inference on a real GPU while injecting modeled HBF timing, capacity, and thermal effects during execution. ### Key Features: - **Real LLM Inference on Real GPU**: HBFSim simulates real LLM workloads directly on a real GPU, ensuring accurate performance metrics. - **Modeling HBF Effects**: It models the timing, capacity, and thermal effects of HBF on the workload, providing a comprehensive evaluation of HBF performance. - **Fast Simulation**: HBFSim is designed to be significantly faster than traditional GPU simulation methods, running about 4,650 times faster than HBF simulation using GPGPU-Sim. ### Tech Stack: Python ### Why It Matters: HBFSim is crucial for evaluating the performance of High-Bandwidth Flash (HBF) in the context of LLM inference. It provides a realistic simulation environment that can help system designers make informed decisions about the placement of data in HBF, ensuring optimal performance and cost-effectiveness in accelerator-based systems.

Overview: TileDB-Vector-Search is a cloud-native library and API for efficient vector similarity search and storage, built on the TileDB Storage Engine. It offers a serverless, scalable solution. Key features: - Vector similarity search - Efficient storage - Cloud-native architecture Tech stack: C++ and Python Why it matters: TileDB-Vector-Search provides a powerful, scalable, and efficient way to store and search vector data, making it ideal for applications requiring high performance and low latency. Repository: TileDB-Inc/TileDB-Vector-Search Primary language: Jupyter Notebook GitHub stars: 69 GitHub description: Cloud-native vector similarity search and storage with efficient, serverless scale-out Quick Links: Build Instructions, Documentation, Python API reference Quick Installation: pip install tiledb-vector-search or conda install -c tiledb -c conda-forge tiledb-vector-search Contributing: Please see `Building` for development instructions, and open issues for feature discussions.

### Overview: This repository collects a comprehensive list of papers on *Deep Hiding*, focusing on generative image steganography techniques. It provides summaries and categorizes various papers, making it a valuable resource for developers interested in enhancing image security and privacy. ### Key Features: - **StegoGAN**: Leveraging Steganography for Non-Bijective Image-to-Image Translation. - **StegaStyleGAN**: Towards Generic and Practical Generative Image Steganography. - **Cross**: Diffusion model makes controllable, robust, and secure image steganography. - **StegaDDPM**: Generative Image Steganography based on Denoising Diffusion Probabilistic Model. - **Generative steganography via auto-generation of semantic object contours**. - **Secret-to-image reversible transformation for generative steganography**. - **Robust steganography without embedding based on secure container synthesis and iterative message recovery**. ### Tech Stack: - Python ### Why It Matters: This repository serves as a valuable resource for developers looking to enhance image security and privacy through generative image steganography techniques. It provides a curated list of papers and summaries, making it an essential tool for anyone interested in this field.

### Overview: The VoicePrivacy Challenge 2022 project aims to provide a baseline recipe for anonymization systems and evaluation software. This repository contains pre-trained models and scripts for the challenge, focusing on anonymization using x-vectors and neural waveform models. ### Key features: - Baseline B1.a: Uses x-vectors and neural waveform models for anonymization. - Baseline B1.b: Directly converts BN features, F0, and x-vector using an NSF model for speech synthesis. ### Tech stack: - Python ### Why it matters: This project is crucial for researchers and developers interested in anonymizing voice data for speech privacy challenges, providing a standardized baseline and evaluation framework to assess anonymization techniques effectively.

### Overview: Heliox OS is an open-source, local-first desktop AI agent designed for voice, gesture, and system automation. It translates user requests into validated plans, ensuring permission gates and verification before executing actions. Heliox supports Windows, macOS, and Linux, offering a secure and controlled environment for desktop automation. ### Key Features: - **Command Center**: A visual interface for managing tasks and actions. - **Permission Gates**: Ensures only authorized actions are performed. - **Local Execution**: Actions are executed locally, without relying on cloud models. - **Verification**: Results are reported only through the guarded execution path. ### Tech Stack: - Python 3.11+ - Windows, macOS, Linux ### Why It Matters: Heliox OS provides a secure and controlled environment for desktop automation, ensuring that only authorized actions are performed. It offers a local-first approach, avoiding the need for cloud models, and includes features like a to-do list, password generator, and focus history, making it a valuable tool for enhancing productivity and control over one's computer.

### Overview: `trellis2.c` is a powerful toolkit for generating textured, segmented, and rigged GLB assets entirely on the GPU. It leverages advanced technologies like Pixal3D, SegViGen, and TokenSkin to create high-quality 3D models from images, including image-to-3D generation, mesh texturing, automatic part decomposition, and automatic rigging. The toolkit runs inference entirely on native CUDA or Vulkan, eliminating the need for Python or PyTorch, making it highly efficient and suitable for real-time applications. ### Key Features: - **Image-to-3D Generation**: Converts images into 3D models. - **Mesh Texturing**: Adds textures to 3D models. - **Automatic Part Decomposition**: Breaks down complex models into simpler parts. - **Automatic Rigging**: Creates realistic rigs for 3D models. ### Tech Stack: - **C/C++**: The primary programming language used. - **CUDA**: For GPU-accelerated inference. - **Vulkan**: For GPU-accelerated rendering. ### Why It Matters: `trellis2.c` is essential for developers looking to create high-quality 3D models quickly and efficiently. Its ability to generate assets entirely on the GPU ensures fast performance and reduces the need for complex Python or PyTorch setups, making it ideal for real-time applications and game development. ### Repository: The repository is accessible at <https://github.com/Wimacs/trellis2.c>. It has been starred by 69 developers, indicating its popularity and

Overview: This project is an official PyTorch implementation of the paper "Towards Adversarial Attack on Vision-Language Pre-training Models" presented at ACM Multimedia 2022. It focuses on adversarial attacks on vision-language pre-training models. Key features: - Co-Attack: An official PyTorch implementation of the paper "Towards Adversarial Attack on Vision-Language Pre-training Models" presented at ACM Multimedia 2022. - Fine-tuned checkpoints for ALBEF and TCL on Baidu Cloud. - Evaluation metrics: Clean accuracy, adversarial accuracy, and ASR (Adversarial Success Rate). - Instruction: 0 (No Attack), 1 (Attack Text), 2 (Attack Image), 3 (Attack Both), 4 (Co-Attack). - Requirements: pytorch 1.10.2, transformers 4.8.1, timm 0.4.9, bert_score 0.3.11. - Data: Dataset json files for downstream tasks, Finetuned checkpoint for ALBEF and TCL, SNLI-VE datasets for Visual Entailment, MSCOCO dataset for Visual Grounding. Tech stack: Python. Why it matters: This project provides a comprehensive solution for evaluating and implementing adversarial attacks on vision-language pre-training models, making it valuable for researchers and practitioners in the field of computer vision and natural language processing.

Overview: This repository is a curated collection of exceptional open-source libraries for generative AI, meticulously reviewed or slated for review by The AI Engineer. It aims to advance the AI engineering discipline and welcome contributions. Key features: - Langfuse - ZenML - MindsDB - FlowiseAI - OpenLLM - Axflow - Haystack - LiteLLM - Ludwig - Jina - LM Format Enforcer - Autogen - Chroma - Weviate - Redisvl - Milvus - Qdrant - RAGAS - SuperDuperDB - Llamafile - pgvector - LMQL - Rags - LLaMa-Factory - LLM-VM - unsloth - DSPy - Ollama - Superagent - Marvin - MetaGPT - GPT-RAG - FastChat - guidance - MiniChain - GPT4ALL - Vectorlink - vLLM - gpt-crawler - PostgresML - Chathub - LLM Evaluation Harness - Vald - Marco - Llama 2 - LLaMA.cpp - GPT-Engineer - FaceSwap - GPT4free - Whisper - LangChain - Stable Difussion WebUI - Transformers - AutoGPT - Continue - Gateway - CrewAI - Danswer - Tabby - Skypilot

Overview: autosrt is a Python script designed to automatically generate subtitle and translated subtitle files for videos and audio files using free speech recognition and translation APIs. Key features: - Automatically generates subtitle and translated subtitle files for videos and audio files - Compatible with various file formats (e.g., MP4, WAV, FLAC) - Supports multiple languages and translations - Can be used as a command-line tool or compiled into an executable for easier execution Tech stack: Python Why it matters: This script streamlines the process of generating subtitles and translations for videos and audio files, making it easier for users to access closed captions and translations without needing to manually transcribe or translate the content.