Projects

Overview: 本项目是一个基于 RAG 的生猪健康管理智慧医药系统，集成了 AI 兽医诊断、疾病管理、药品管理、文章资讯管理等功能模块，旨在通过数字化手段提升生猪养殖的医疗管理水平。 Key features: - AI 兽医诊断 - 疾病管理 - 药品管理 - 文章资讯管理 Tech stack: Spring Boot 3 + Spring AI + Milvus + MySQL 8 + MyBatis-Plus + Ollama / DeepSeek / 通义千问 / SiliconFlow Why it matters: 项目提供智能兽医诊断服务，帮助养殖户或兽医快速识别生猪病情并推荐相应的治疗方案，提升生猪养殖的医疗管理水平。

Overview: Companion is a Generative-AI-powered foreign-language private tutor that allows users to learn a new language by writing, talking, reading, and listening in both their native language and the selected foreign language. Key features: - Uses OpenAI ChatGPT and Whisper for language translation and text generation. - Supports Python 3.8 and above. - Corrects any mistakes made by the user. - Available both locally and as a cloud service. - Configured to run on mobile devices. Tech stack: Python. Why it matters: Companion is a powerful tool for anyone looking to learn a new language, offering a personalized learning experience with real-time corrections and the ability to practice in both languages.

Overview: This is an 8-module course designed to help developers build an agentic RAG application from scratch by collaborating with Claude Code. The course covers topics such as chat interfaces, document ingestion, full RAG pipeline, agentic patterns, and more. Key features: - Chat interface with threaded conversations, streaming, tool calls, and subagent reasoning - Document ingestion with drag-and-drop upload and processing status - Full RAG pipeline: chunking, embedding, hybrid search, reranking - Agentic patterns: text-to-SQL, web search, subagents with isolated context Tech stack: React, TypeScript, Tailwind, Vite, Python, FastAPI, Supabase, Docling, OpenAI, OpenRouter, LangSmith Why it matters: This course provides a hands-on experience for developers to build a full-featured RAG system, focusing on collaboration with Claude Code and practical application of various AI and database technologies.

### Overview Chunky is an open-source toolkit designed for reliable Retrieval-Augmented Generation (RAG) pipelines. It simplifies the process of converting PDFs to Markdown, cleaning documents, inspecting chunks, comparing different chunking strategies, and enriching metadata for LLM applications. ### Key Features - **Document Review Workspace**: Compare PDF, Markdown, and chunks side by side before indexing. - **Multiple Conversion Engines**: Support PyMuPDF, Docling, MarkItDown, LiteParse, VLM, and Cloud API. - **Batch Processing**: Convert, enrich, and chunk multiple documents from the sidebar. - **Chunking Strategy Comparison**: Test LangChain, Chonkie, and Docling splitters with configurable size and overlap. - **Saved Chunk Versions**: Persist and reload chunk sets by Markdown source and splitter configuration. - **Markdown Enrichment**: Clean conversion artifacts with deterministic cleanup plus LLM correction. - **Chunk Enrichment**: Generate context-aware titles, summaries, keywords, and retrieval questions. - **Pluggable Backend**: Add converters or splitters through the registry without frontend changes. ### Tech Stack Python ### Why It Matters Chunky helps ensure that documents are prepared correctly before they reach a vector store, reducing the likelihood of RAG failures. It provides a clear view of the document's structure and content, allowing for better decision-making and improved results for LLM applications. ### Repository Repository: GiovanniPasq/chunky GitHub Stars: 192 GitHub Description: Open-source toolkit for reliable RAG pipelines: convert

### Overview: This repository is an open-source survey of the latest video reasoning tasks, paradigms, and benchmarks. ### Key features: - 🧠 Reasoning Paradigms: Textual Chain-of-Thought, Chain-of-Frames, Structured & Latent Reasoning, Code World Models - 🤖 Agentic & Interactive Reasoning: Thinking with Videos, Agentic Tool Use, Unified Multimodal Models, Streaming & Proactive Reasoning - 🌍 Perception & World Models: Long-Video Perception, Predictive World Models for Action ### Tech stack: Python ### Why it matters: This repository provides a comprehensive overview of video reasoning tasks and paradigms, helping researchers and practitioners stay updated on the latest advancements in the field.

Overview: This repository provides tutorials on using PandasAI, an advanced Python library for interacting with data through generative AI. Key features: - Ask questions about data in natural language. - Generate plots for data visualization. - Clean datasets by addressing missing values. - Enhance data quality through feature generation. - Connect to various data sources like CSV, XLSX, PostgreSQL, MySQL, BigQuery, Databricks, Snowflake, etc. Tech stack: Python, PandasAI, Ollama, Streamlit, Chainlit, Llama3, Meta. Why it matters: PandasAI simplifies data exploration and analysis, making it easier to work with large datasets and generate insights through AI-driven solutions.

### Overview: 本项目精选了涵盖大语言模型长期记忆实现的各类开源和商业项目，从底层检索/存储范式到多模态一致性、训练与对齐的完整生态，全面覆盖了AI记忆领域的关键技术与应用。 ### Key features: - **记忆管理 RL 化**: AgeMem 等项目将长期/短期记忆的增删改查转化为agent可训练的工具动作，采用三阶段RL（step-wise GRPO）端到端优化，取代启发式控制器。 - **从"原子事实"到多粒度**: TriMem 同时保留原始对话片段、抽取事实与合成画像三层粒度，使用TextGrad迭代优化提示词，实现无参数更新的终身演化。 - **冲突感知与事实修订**: MOSAIC、Infini Memory针对"事实随时间变化"设计可维护记忆（主题文档 / 冲突检测），直击写路径质量控制难题。 - **评测从"召回"转向"行动"**: MemoryAgentBench、MemoryArena等基准测试将评估推向操作级幻觉、多会话agentic任务与选择性遗忘。 - **测试时记忆层化**: Titans-as-a-Layer 把神经长期记忆做成即插即用适配器，无需改动主干即可为多模态模型注入长期记忆。 - **本地优先跨工具 MCP 记忆**: Memorix、OKF标准的mcp-memory等以SQLite/knowledge graph + Markdown落地，强调私有化与跨

### Overview: CampusX offers a collection of free courses designed to help developers master key technologies in AI, ML, Data Science, and more. These courses are tailored to provide practical skills and real-world applications. ### Key Features: - **SQL for Data Analysis and Problem Solving** - **Building AI Agents** - **Credit Risk Modeling Techniques** - **Data Analysis using MS Excel** - **Data Science Interview Preparation** - **Advanced Deep Learning with PyTorch** - **e-KYC Solutions Using Computer Vision** - **Emotion Detection with Convolutional Neural Networks (CNNs)** - **Building End-to-End ML Systems with MLOps** - **Fine Tuning Transformers** - **Flask for Machine Learning** - **End-to-End Machine Learning Projects with AWS Sagemaker** - **Introduction to Generative AI with LangChain** - **Leveraging LlamaIndex for Generative AI Tasks** - **Generative AI for Vision** - **IPL Dashboard using Power BI** - **MLFlow for MLOps** - **Object Detection using Deep Learning** - **Retrieval Augmented Generation (RAG)** ### Tech Stack: Jupyter Notebook ### Why It Matters: These courses are designed to equip developers with the skills necessary to tackle real-world problems and projects in the field of AI, ML, and Data Science. Whether you're looking to enhance your existing skills or start from scratch, these courses provide a solid foundation and practical experience.

### Overview: This repository contains materials for the INTERSPEECH 2019 Tutorial titled "Advanced methods for neural end-to-end speech processing – unification, integration, and implementation". It provides hands-on demonstrations of Text-to-Speech (TTS) and Automatic Speech Recognition (ASR) using Jupyter Notebooks. ### Key Features: - TTS demonstration: A Jupyter Notebook for text-to-speech conversion. - ASR demonstration: A Jupyter Notebook for automatic speech recognition. ### Tech Stack: Jupyter Notebooks ### Why It Matters: This tutorial is essential for developers and researchers interested in advanced techniques for neural end-to-end speech processing, providing practical examples and insights into real-world applications. ### Repository: espnet/interspeech2019-tutorial ### GitHub Stars: 194 ### GitHub Description: INTERSPEECH 2019 Tutorial Materials ### Primary Language: Jupyter Notebook ### Questionnaire: https://forms.gle/RhUaU5437sx5dsmAA ### Slides: 1. Lecture: [Lecture](https://drive.google.com/open?id=1YRwQ9S2PmRCp5WBcNufAhcqIKtmREkfn) 2. TTS demonstration: [TTS demonstration](https://nbviewer.jupyter.org/format/slides/github/espnet/interspeech2019-tutorial/blob/master/notebooks/interspeech2019_tts/interspeech2019_tts.ip

Overview: EvalForge is a tool designed to evaluate RAG applications and AI assistants, providing portable evaluation evidence and policy gates for CI systems. It includes JSON, JUnit, SARIF, and Markdown reports, and supports offline evaluation with two direct dependencies. Key features: - Import promptfoo, Ragas, or supported DeepEval results. - Enforce versioned policies and publish JSON, JUnit, SARIF, and Markdown reports. - Generate a portable artifact and JSON Schemas for evaluation results. - Compare multi-dimensional deltas on the same dataset fingerprint. - Apply explicit quality, security, latency, and cost thresholds. - Emit a non-zero exit code on regression. Tech stack: Python 3.9+. Why it matters: EvalForge provides a missing, deliberate interoperability layer between AI evaluation frameworks and CI platforms, making it easier to share and understand evaluation results. Repository: jsdhwfmax/EvalForge GitHub stars: 194 GitHub description: Evaluator-neutral AI evaluation evidence, baseline regression gates, and JSON, JUnit, and SARIF reports for CI.

### Overview: This Cordova plugin provides a simple and efficient way to integrate speech recognition functionality into your mobile applications using the Cordova platform. It supports both Android and iOS platforms and allows developers to easily implement speech recognition features without the need for complex native code. ### Key Features: - **Speech Recognition Integration**: Easily add speech recognition capabilities to your Cordova app. - **Language Support**: Supports multiple languages, including English, Spanish, Chinese, and many more. - **Prompt and Popup**: Provides options to display prompts and customize the user interface. - **Platform-Specific Behavior**: Differentiates between Android and iOS platforms, ensuring compatibility across both. - **Permissions and Availability Check**: Checks for the availability of speech recognition and prompts the user for permission if necessary. ### Tech Stack: Java ### Why It Matters: - **Ease of Use**: Simplifies the process of integrating speech recognition into mobile apps. - **Cross-Platform Compatibility**: Ensures that the app can run seamlessly on both Android and iOS devices. - **Language Support**: Allows developers to cater to a wider audience by supporting multiple languages. - **User Experience**: Customizable prompts and popup windows enhance the user experience, making the app more intuitive and user-friendly. ### GitHub Stars: 194 This plugin has gained significant traction in the Cordova community, indicating its usefulness and popularity among developers.

Overview: Dim0 is an open-source, real-time collaborative AI canvas that integrates notes, mini-apps, and an AI agent onto a single, infinite board. The agent reads and writes back to the canvas, enabling a seamless real-time collaboration experience. Key features: - Infinite canvas with thousands of nodes and nested boards - Real whiteboard underneath with hand-drawn and geometric shapes, images, and media - Board-aware agent that reads the workspace and writes back results - Mini-apps as real, persistent, and editable React apps - Rich notes with Notion-style rich text, math, and code - Code and documents sandbox for running code in sandboxes - Bring-your-own model switchable with various AI models - Real-time multiplayer with live cursors, shared edits, and conflict-free sync - Present from the canvas with frames as a slideshow - Open-source and self-hostable with your data staying yours Tech stack: TypeScript Why it matters: Dim0 provides a unified workspace where AI and human collaboration can seamlessly interact, making it an invaluable tool for developers, designers, and anyone looking to enhance their productivity and creativity. Repository: vcmf/dim0 Primary language: TypeScript GitHub stars: 193 GitHub description: Open-source, real-time collaborative AI canvas - notes, mini-apps, and agents on one infinite board. README excerpt: <p align="center"> <img src="docs/images/dim0-icon.png" alt="Dim0" width="120" /> </p> <h

### Overview: BluePixelModel is an advanced native 4K image editing model built on Flux 2 Klein and DCAE, designed for efficient and effective image manipulation. This model supports 4096x4096 square output without cascaded super-resolution, making it suitable for high-resolution editing tasks. ### Key features: - **Native 4K Direct Output**: Supports 4096x4096 square output without cascaded super-resolution, ensuring non-square images retain their aspect ratio. - **4-Step Inference**: Utilizes source-anchored target projection (SATP) distilled from a 50-step teacher, providing a more efficient 4K training process. - **Efficient 4K Training**: Uses DCAE-F64 to keep the 4K DiT token count at the 1K level, completing initial 4K adaptation on 32 Kunlunxin XPUs in about 4 days. - **Continual Learning**: Absorbs long-tail vertical editing skills through ACE-LoRA + OPD, avoiding task interference and catastrophic forgetting. ### Tech stack: The model is built on Python, utilizing the Flux 2 Klein and DCAE architectures for efficient image editing. ### Why it matters: This model is designed to streamline the editing process by leveraging advanced neural network techniques, making it highly effective for high-resolution image manipulation tasks.

Overview: This project provides a simple DALL-E 3 API for developers, allowing them to download images based on their queries. Key features: - Downloading images from the web using the DALL-E 3 Unofficial API - Using a cookie to bypass automation detection - Implementing a simple version of Idea2Image for prompt enrichment Tech stack: Python Why it matters: This API simplifies the process of generating images based on textual descriptions, making it useful for developers looking to automate image generation tasks.

### Overview: This repository contains the official implementation of the paper "Uncovering the Redundancy in Transformers via a Unified Study of Layer Dropping" (TMLR 2026). The project aims to enhance the efficiency of transformer architectures by systematically removing redundant layers, leading to significant speedup and memory savings without requiring specialized hardware. ### Key Features: - **Unified Dropping Taxonomy**: Systematically dissects and compares Block Drop, Attention-Layer Drop, MLP-Layer Drop, and Joint Layer Drop under a standardized framework. - **Significant Speedup & Memory Savings**: Achieves up to 2.1x inference speedup and over 40% KV cache memory reduction. - **High Performance Retention**: Retains >95-98% of core reasoning and general language capabilities (MMLU, GSM8K, ARC-c, HellaSwag) through importance-aware layer selection. - **Orthogonal Quantization Synergy**: Easily pairs with post-training 4-bit quantization (AWQ/GPTQ) for compounding latency and memory benefits. - **Plug-and-Play Hugging Face Integration**: Output models use standard `auto_map` configurations for seamless loading. ### Tech Stack: Python ### Why it Matters: This project provides a comprehensive solution for improving the efficiency of transformer architectures, making them more suitable for resource-constrained environments. The unified study of layer dropping offers a systematic approach to reducing redundancy, leading to significant performance improvements without compromising on core functionality.

### Overview: OS-Genesis is an innovative framework designed for automating GUI agent trajectory construction without human supervision. Leveraging reverse task synthesis and a trajectory reward model, it enables effective end-to-end training of GUI agents. ### Key Features: - **Reverse Task Synthesis**: Automatically generates tasks based on the reverse of the desired trajectory. - **Trajectory Reward Model**: Encourages agents to learn optimal trajectories by rewarding them for following the generated tasks. - **GUI Agent Trajectory Data Construction**: Automatically builds high-quality and diverse trajectory data for GUI agents. ### Tech Stack: Jupyter Notebook ### Why it Matters: OS-Genesis simplifies the process of creating high-quality GUI agent trajectory data, making it easier to train and improve GUI agents without the need for extensive human intervention or predefined tasks. This approach enhances the efficiency and effectiveness of GUI agent training, leading to more robust and adaptable AI systems.

### Overview: ReSim is a reliable world simulation model designed for autonomous driving, focusing on action-conditioned future prediction and training from diverse data sources. ### Key features: - **Reliable Action Control**: Supports expert, free-driving, commanded, and hazardous non-expert behaviors. - **Heterogeneous Training Data**: Trained on a mixture of web driving videos, real driving datasets, and simulated data. - **High-Fidelity Prediction**: Generates realistic future driving videos while improving controllability for expert and non-expert actions. ### Tech stack: Python, PyTorch, CUDA, SAT, DeepSpeed. ### Why it matters: ReSim provides a robust simulation environment for autonomous driving, ensuring reliable and realistic future driving scenarios.

### Overview: This project automates the creation of story-based videos by utilizing Ollama LLM to generate text ideas based on user prompts. It converts the generated stories into audio, creates captions from the audio, and merges them into videos. ### Key Features: - Text to Text: AI generates stories based on user prompts. - Text to Speech: Converts stories into audio. - Caption Generation: Captions are generated from audio. - Video Creation: Captions are merged into videos. ### Tech Stack: Go ### Why It Matters: This project streamlines the creation of engaging video content, making it easier for users to generate compelling narratives with minimal effort.

Overview: This repository implements Denoising Diffusion Probabilistic Models (DDPM) using Pytorch, focusing on the training and sampling methods of DDPM and the Unet architecture used in the Stable Diffusion Unet from the Hugging Face diffusers library. Key features: - DDPM Implementation - Unet Architecture - Training and Sampling Methods Tech stack: Pytorch Why it matters: This project provides a practical implementation of DDPM, which is a powerful technique for generating images from noise, and showcases the Unet architecture used in the Hugging Face diffusers library, making it a valuable resource for developers interested in image generation and diffusion models.

### Overview: A collection of prompts designed for use with GPT-4 via ChatGPT, OpenAI API, and a Gradio frontend, aimed at showcasing the influence of natural language structure on model outputs. ### Key Features: - **Prompts in Pseudocode Format**: Structured in a standardized pseudocode-like format for easy integration. - **Gradio Frontend**: A simple frontend for interacting with the model. - **Collection of Natural Language Programs**: Prompts designed to act upon concepts and data represented as text. ### Tech Stack: Jupyter Notebook ### Why It Matters: This project aims to demonstrate how the quality and presentation of natural language fed to a model significantly influence its outputs, providing insights into the capabilities of large language models like GPT-4 and ChatGPT.