Projects

### Overview: Taskade is an AI-native workspace platform that allows developers and AI agents to build apps, deploy agents, automate workflows, and more. It connects projects, data, and AI agents as Workspace DNA, transforming one prompt into a live app. ### Key Features: - **Taskade Genesis**: One prompt creates a live app with AI agents, real-time data, and automation workflows. - **AI App Builder**: A powerful platform for developers and AI agents to build and deploy AI-driven applications. - **Real-time Data Integration**: Seamlessly integrates projects, data, and AI agents into a single, connected workspace. - **Automated Workflows**: Automate tasks and workflows with AI-driven agents, streamlining processes and reducing manual effort. - **Project Management**: Track projects, manage tasks, and collaborate with teams in a unified workspace. - **Bug Reporting and Feature Requests**: Easily submit bug reports and feature requests, and have them addressed by the platform. ### Tech Stack: Taskade is built with a combination of modern web technologies, including React, TypeScript, and Node.js. It leverages the power of AI and machine learning to provide intelligent automation and data-driven insights. ### Why it Matters: Taskade is designed to empower developers and AI agents to create intelligent applications that can handle complex tasks and workflows. By connecting projects, data, and AI agents seamlessly, Taskade provides a unified workspace that can quickly respond to prompts and automate repetitive tasks. This platform is ideal for anyone looking to build AI-driven applications or streamline their workflow processes.

### Overview: This project utilizes the Riva SDK and Google Cloud's Audio2Face Avatar technology to create a virtual assistant that can interact with users through voice commands and provide contextually relevant responses. The system is designed to be deployed in a Google Cloud environment, making it scalable and accessible. ### Key Features: - **Voice Interaction**: Users can interact with the virtual assistant via voice commands. - **Contextual Responses**: The assistant provides responses based on the context of the user's query, enhancing the user experience. - **Deployed in Google Cloud**: The project is hosted on Google Cloud, ensuring reliability and scalability. ### Tech Stack: - Python ### Why it Matters: This project offers a practical solution for creating a virtual assistant that can engage users through voice, enhancing user interaction and providing personalized responses based on context. The use of Google Cloud ensures that the project is scalable and reliable, making it suitable for various applications requiring voice-based interaction.

### Overview: 羽人-百川 7B 是一个基于 baichuan-7B 的开源多模态大语言模型，旨在解决多轮对话、开放域问答、角色扮演、文本生成、文本理解、图片理解等多个任务。模型在多个任务上表现出色，特别是在多模态任务上，其图片理解能力尤为突出。 ### Key features: - **多模态融合**: 将 LLM 的语言模态和 SOTA 的 CLIP 模型视觉编码器进行融合，实现卓越的图片理解能力。 - **高质量 SFT 数据集**: 基于 Pleisto 自有的商业多轮对话与指令精调数据集，数据集经过多轮质检，包含高质量的图像和文本数据。 - **商业友好**: 源代码和模型权重均以 Apache-2.0 协议开源，模型训练和推理代码完全继承自 baichuan-7B 模型许可协议，支持免费商业使用。 - **兼容 ChatML**: 全面兼容 GPT-4 同款的 ChatML 格式，减少 Prompt Injection 风险，实现良好的 System Prompt 遵循度。 - **参数量限制**: 受限于较小的参数量，模型在数值计算、逻辑推理类任务上表现不佳，但在多模态任务上仍可发挥 CLIP 的优势。 ### Tech stack: Python ### Why it matters: 羽人-百川 7B 是一个基于 baichuan

Overview: This project provides a framework for fine-tuning diffusion models using trajectory-level optimization, aiming to improve the efficiency and effectiveness of diffusion model training. Key features: - Trajectory-level optimization: Guides the diffusion process toward high-reward regions. - No additional inference cost: Amortizes trajectory optimization into training. - Generalized reward alignment: Works with any reward model, including differentiable and non-differentiable objectives. Tech stack: Python Why it matters: This framework enhances the efficiency and effectiveness of diffusion model training by optimizing sampling paths through reward-guided iterative refinement, making it suitable for various applications in image generation and fine-tuning.

### Overview: A Python wrapper around ComfyUI, allowing the AI Horde to run ComfyUI-designed inference pipelines programmatically. ### Key features: - **ComfyUI Integration**: Wraps a pinned commit of ComfyUI to enable AI Horde to use ComfyUI-designed inference pipelines. - **GPU/CUDA Requirement**: Requires an NVIDIA GPU with CUDA support. - **Public API**: Provides a public API for interacting with ComfyUI inference pipelines. ### Tech stack: Python 3.12+ with GPU/CUDA requirements. ### Why it matters: The AI Horde relies on this wrapper to leverage ComfyUI's capabilities for advanced image generation and manipulation, making it easier for the AI to create high-quality images and designs.

Overview: HaoQChen/iflytek_awaken_asr is a project that utilizes Ifytek's technology to achieve awaken and order recognition. Key features: - Integration with Ifytek's speech recognition and wake word detection capabilities - Real-time processing of audio data for accurate recognition - Support for various languages and dialects Tech stack: C Why it matters: This project is crucial for developers looking to integrate advanced speech recognition and wake word detection into their applications, enhancing user experience and improving operational efficiency.

### Overview: Flash-MinerU is a versioned, RayOrch-powered execution layer for MinerU, turning PDFs into Markdown and reducing results back to the correct document and order. It keeps MinerU's models and output conventions while isolating each supported MinerU runtime/version in an explicit pipeline. ### Key Features: - **Versioned Execution**: Ensures consistent and reproducible results across different MinerU versions. - **RayOrch Integration**: Uses RayOrch for logical dataflow, including explicit fan-out/fan-in, lineage, readiness, and cross-document batching. - **MinerU Isolation**: Each MinerU runtime/version is isolated in an explicit pipeline, allowing applications to upgrade one path at a time. ### Tech Stack: - **Ray**: Supplies the physical distributed runtime, including actors, GPU placement, resources, RPCs, and object storage. - **RayOrch**: Supplies the logical dataflow, including explicit fan-out/fan-in, lineage, readiness, and cross-document batching. ### Why It Matters: Flash-MinerU simplifies the process of converting PDFs into Markdown, making it more scalable and efficient. By keeping MinerU's models and output conventions intact, it allows for easy upgrades and maintenance of applications without disrupting the overall workflow.

Overview: PocketLLM is a Flutter-based AI chat application that offers offline model execution, Ollama/LLMStudio integration, and a modern UI. It supports privacy-focused, cross-platform, and open-source features. Key features: - Unified model management - Bring your own keys - Consistent chat experience across platforms - Dynamic model catalogues - Granular provider operations Tech stack: Flutter, FastAPI, Supabase Why it matters: PocketLLM provides a secure, low-latency access to large language models, allowing users to connect their own provider accounts and enjoy a shared, cross-platform chat experience.

Overview: This repository provides code and tutorials for participants to develop dose prediction models for the GDP-HMM Challenge at AAPM 2025. Key features: - Code and tutorials for training and evaluation - Pretrained model and baseline for comparison - Data understanding and visualization tools - Docker setup for final phase - Evaluation methods and results Tech stack: Python Why it matters: This repository offers a comprehensive solution for participants to tackle the challenging dose prediction task at the AAPM 2025 challenge, providing essential resources and tools to help achieve success. Repository: RiqiangGao/GDP-HMM_AAPMChallenge GitHub stars: 72 GitHub description: The code base for Generalizable Dose Prediction for Heterogenous Multi-Cohort and Multi-Site Radiotherapy Planning (GDP-HMM) challenge at AAPM 2025

### Overview: SynParaSpeech is an automated synthesis framework designed to create large-scale paralinguistic datasets, enhancing speech synthesis and understanding. The project addresses the limitations of existing paralinguistic datasets by generating high-quality data with paralinguistic sounds that are fully aligned with speech, text, and precise timestamps. ### Key Features: 1. **Automated Synthesis of Paralinguistic Speech-Text Pairs**: SynParaSpeech generates timestamped text using multiple ASR models and inserts paralinguistic tags with Deepseek Chat V3. 2. **Audio Synthesis Integration**: Combines paralinguistic audio (voice conversion via SeedVC) with speech audio to maintain timbre consistency. 3. **Manual Verification**: Evaluates 4 dimensions (naturalness, paralinguistic matching, audio quality, temporal alignment) to ensure high-quality clips. 4. **TTS Model Optimization**: Enhances TTS models (CosyVoice2, F5-TTS) using SFT + DPO optimization to improve paralinguistic quality. 5. **Paralinguistic Event Detection**: Prompts MLLMs to detect paralinguistic events, improving the accuracy of paralinguistic speech understanding. ### Tech Stack: JavaScript ### Why It Matters: SynParaSpeech provides a comprehensive solution for paralinguistic dataset construction, addressing the limitations of existing datasets and enabling more realistic speech synthesis and understanding. The framework's automation and high-quality data generation make it a valuable tool for researchers and developers in the field of speech processing and synthesis.

Overview: RAW-Diffusion is an official implementation of the RGB-Guided Diffusion Models for High-Fidelity RAW Image Generation paper presented at WACV 2025. It offers a novel approach to generate high-fidelity RAW images using RGB guidance, addressing the need for comprehensive RAW image datasets. Key features: - RGB-Guided Diffusion Models - RAW Image Generation - Comprehensive RAW Dataset Creation Tech stack: Python Why it matters: RAW-Diffusion provides a practical solution for generating high-fidelity RAW images, addressing the limitations of current deep learning approaches that prioritize RGB data. It addresses the need for diverse RAW datasets by integrating RGB guidance, leading to improved performance across various applications, including object detection in RAW imagery. GitHub stars: 72

### Overview: RaBitQ is a practical and asymptotically optimal quantization algorithm designed for high-dimensional vector quantization in Euclidean space. This project introduces a novel approach to compress high-dimensional vectors with arbitrary compression rates, achieving dominant accuracy under the same compression rates as classical scalar quantization. The algorithm significantly improves recall in the setting from 2-bit to 6-bit, making it suitable for approximate nearest neighbor search applications. ### Key Features: - **Practical Implementation Techniques**: The project provides a library that seamlessly integrates with scalar and binary quantization, offering blazingly fast distance computation with dominant accuracy. - **Asymptotically Optimal Quantization**: RaBitQ supports high-dimensional vector quantization with arbitrary compression rates, achieving dominant accuracy under the same compression rates as classical scalar quantization. - **Significant Improvement in Recall**: The algorithm brings significant improvements in recall, especially from 2-bit to 6-bit, enabling high recall without reranking. ### Tech Stack: C++ ### Why it Matters: RaBitQ is particularly significant for applications requiring high recall in approximate nearest neighbor search, such as recommendation systems, image search, and other scenarios where precise vector quantization is crucial. Its ability to achieve dominant accuracy under the same compression rates as classical scalar quantization makes it a valuable tool for improving recall in various practical applications.

Overview: This project provides a FLUX 2 Pro API, which is an OpenAI-compatible endpoint for generating images from text or references. The API is priced per megapixel and supports various output resolutions, with a minimum charge of $1. Key features: - FLUX 2 Pro API (`flux-2-pro`) with OpenAI-compatible endpoint - Pricing based on megapixels - Supports multiple output resolutions - Minimum charge of $1 - Asynchronous processing Tech stack: Python Why it matters: This API simplifies the process of integrating FLUX 2 Pro with OpenAI-compatible models, offering a single endpoint for image generation without the need for monthly subscriptions or pre-paid plans.

### Overview: A Vue2 application for performing streaming speech recognition using Google Cloud Speech, designed for Progressive Web Apps (PWAs). ### Key features: - Real-time speech recognition - Google Cloud Speech API integration - Progressive Web App (PWA) support ### Tech stack: JavaScript ### Why it matters: This project provides a practical solution for developers looking to implement speech recognition in their PWAs, leveraging the power of Google Cloud Speech for real-time audio processing.

### Overview: Nano Banana 2 (official) API (`gemini-3.1-flash-image-preview-official`) is an OpenAI-compatible endpoint providing text and image generation services. It offers a single key access to multiple models, enabling teams to switch between models without changing their client configurations. The API is token-billed, with prices per million tokens, and supports both text and image outputs. It supports up to 100 tasks per day and offers a $1 minimum charge, with the cost reported in the response for each completed task. ### Key Features: - **OpenAI-Compatible Endpoint**: Access to multiple image, video, and language models via a single API key. - **Token-Billed Pricing**: Charges per million tokens, with a $1 minimum. - **Model Switching**: The same `https://api.apimart.ai/v1` base URL and `Authorization` header can be used to switch between models. - **Batching and Retries**: Tasks are batched and retries are handled as ordinary queue work. - **Async by Design**: Tasks can be submitted, and the `task_id` can be used to poll for completion. ### Tech Stack: - Python ### Why It Matters: - **One Key, Entire Catalog**: The same `https://api.apimart.ai/v1` base URL and `Authorization` header reach Nano Banana 2 (official) and 300+ other models. - **$1 Minimum, Pay As You Go**: No subscription required; pay for each task. - **Charge Backs**: The

Overview: effect-uai is a low-level primitives library for building AI agents using the Effect framework, focusing on explicit control and ease of use. Key features: - Explicit control over the agent loop - Built on Effect, handling retries, streams, and concurrency - Composable primitives for assembling agentic loops - Recipes for handling hard parts like model council, auto-compaction, and pause/resume - Streaming first approach with typed errors - Carrying your own state Tech stack: TypeScript Why it matters: effect-uai provides a low-level, explicit control approach for building AI agents, making it suitable for developers who want to have full control over their agent loops and handle errors and state explicitly. GitHub stars: 71 Repository: betalyra/effect-uai Primary language: TypeScript Status: While in 0.x, minor releases may include breaking changes, with migration guides available for upgrades. Why effect-uai: Most agent libraries decide how your loop works, but effect-uai gives you explicit control and ease of use, focusing on state and error handling.

Overview: Text2Video is a software tool designed to convert text into engaging videos, aiming to enhance learning experiences by making text more visually appealing. Key features: - Converts text to video - Supports various learning styles - Uses AI for better engagement Tech stack: HTML, CSS, JavaScript, Node.js, CCapture.js, ffmpegserver.js, Amazon Polly, Python, Flask, gevent, spaCy, and Pixabay API. Why it matters: Videos are more effective in learning than text for many people, including visual learners. By converting text into videos, Text2Video aims to make learning more engaging and efficient.

Overview: PalmHill.BlazorChat is a Blazor WebAssembly, SignalR, and WebAPI chat application featuring real-time LLM conversations, markdown support, customizable settings, and a responsive design. It supports Llama2 models and was tested with Orca2. Key features: - Real-time chat using SignalR IAsyncEnumerable streaming - Support for Llama3 models - Memory consumption reduction by using a single model for both chat and embedding - Utilizes Microsoft's `semantic-kernel` for chat, embedding, and document retrieval Tech stack: C# Why it matters: This project showcases the integration of cutting-edge AI models and real-time communication technologies, providing a unique and engaging chat experience that leverages the strengths of Llama2 and other large language models.

Overview: Feelime 是一个离线优先的 Android 输入法，支持全拼、双拼、离线语音识别和常用语与剪贴板功能，所有数据都在设备本地处理，无需网络访问。 Key features: - 全拼/双拼、离线语音、常用语与剪贴板 - 离线优先的输入法体验 - 现代界面设计 - 语音输入支持中英混合识别、整句纠错和本地标点恢复 - 中文输入支持全拼、双拼、T9、笔画五键、简体、连续造词、独立用户词库、候选池、联想词库、单手模式 - 多语言支持，包括英语、法语、俄语、日语罗马字 - 键盘界面可定制，支持符号键定制和控制键层 - 键盘源文件在本地，无需重新打包 APK - 键盘热更新功能，无需重启即可快速切换输入法 - 文档丰富，包括工程约定、交互语义、外观方案、验证方法等 Tech stack: JavaScript Why it matters: Feelime 是一个专注于离线优先的输入法解决方案，提供全面的中文和多语言支持，确保用户在任何网络环境下都能保持高效和便捷的输入体验。其现代界面设计和强大的语音识别功能使其成为开发者和用户的理想选择。

Overview: DART is a low-latency speculative decoding approach for LLMs, inspired by diffusion-based models, designed to surpass EAGLE3 by 30% on average. It achieves up to 65% improvement on certain code-centric workloads with a single forward pass of a single transformer layer and a fast cpp-based tree search. Key features: - **Fast Drafting Forward**: Produces multiple logits simultaneously with 1 forward of 1 layer. - **Fast Tree Search**: Uses n-gram-based tree search to build the final draft tree (C++ based). - **Low Drafting Cost**: Results in extremely low drafting cost for efficient inference. - **Relatively High $\tau$**: Average Acceptance Length is competitive with EAGLE3. Tech stack: Python Why it matters: DART offers significant speedup and improved efficiency in LLM inference, making it a valuable tool for researchers and developers working on large language models. Repository: fvliang/DART Primary language: Python GitHub stars: 72 GitHub description: Official Implementation of DART (DART: Low-Latency Parallel Drafting with Continuity-Aware Tree Pruning for Speculative Decoding, EMNLP26). README excerpt: <img src="figs/logo.png" alt="DART" width="220"> [](LICENSE) [![Version](https://img.shields.io/badge/version-0