Projects

### Overview: Trace MCP is a framework-aware code intelligence server designed to improve the efficiency of code review processes. It integrates with 88 frameworks, supports 81 languages, and reduces the input tokens needed to review a pull request by 72.7%, resulting in a median of 60 merged PRs in 6 open-source repositories that are not owned by the project. ### Key Features: - Framework Integration: Supports 88 frameworks. - Language Support: 81 languages. - Token Reduction: 72.7% fewer input tokens needed for code reviews. - Improved Review Efficiency: 13,595 tokens without Trace MCP, 3,291 with it — 72.7% less, median over 60 merged PRs in 6 open-source repos. ### Tech Stack: TypeScript ### Why It Matters: Trace MCP is crucial for developers who need to review code changes in multiple frameworks and languages. By reducing the amount of input tokens required, it speeds up the review process, making it more efficient and effective. This tool helps in maintaining a high standard of code quality while reducing the time and effort spent on code reviews.

### Overview: Smart Coding MCP is an extensible Model Context Protocol (MCP) server designed to enhance AI coding assistants by providing intelligent semantic code search. Built with local AI models, it leverages Matryoshka Representation Learning (MRL) for flexible embedding dimensions (64-768d), making it ideal for exploring unfamiliar codebases and finding related code even with slight variations in terminology. ### Key Features: - **`a_semantic_search`**: The primary tool for codebase exploration, converting natural language queries into AI embeddings to find code by meaning. - **`d_check_last_version`**: Fetches the latest version of any package from official registries, supporting 20+ ecosystems. - **`b_index_codebase`**: Triggers a full reindex of your codebase, generating new embeddings and updating the SQLite cache. - **`c_clear_cache`**: Deletes the embeddings cache entirely, forcing a complete reindex on the next search. - **`e_set_workspace`**: Changes the workspace path at runtime without restarting the server, switching projects without restarting the server. ### Tech Stack: JavaScript ### Why It Matters: Smart Coding MCP simplifies the process of finding relevant code for AI assistants, especially in unfamiliar or large codebases. By using AI embeddings, it can quickly and accurately locate code based on meaning rather than exact keywords, making it a valuable tool for developers and AI assistants alike.

Overview: This project is a pure-Python context management layer designed for LLM systems, integrating retrieval, re-ranking, memory decay, and token-budget enforcement into a single pipeline. Key features: - ContextPacket generation - Memory decay - TokenBudget enforcement - Compressor for document compression Tech stack: Pure-Python with the Python standard library and numpy. Why it matters: This library simplifies the complex process of context management in LLM systems by handling the retrieval, re-ranking, and memory decay steps, ensuring that the model sees the most relevant and up-to-date information efficiently. Repository: Emmimal/context-engine GitHub stars: 198 GitHub description: A pure-Python context management layer for LLM systems — retrieval, re-ranking, memory decay, and token-budget enforcement in one pipeline.

Overview: 本项目旨在设计并实现一个基于 Adaptive-RAG 的电商客服系统，以解决传统客服系统长尾问题、人工介入成本高、多跳推理能力差和幻觉问题严重等问题。 Key features: - 按问题复杂度自动分流 - 简单问题直接回答 - 复杂问题经过多跳检索并自我纠错 - 查单、退改通过工具调用完成 - 高风险写操作必须先经人工确认 Tech stack: Python Why it matters: 该项目通过明确的路由策略、混合检索、多跳推理、忠实度检查和人工确认等技术手段，有效解决了电商客服系统中的痛点问题，提高了客服效率和客户满意度，具有较高的工程化和可观测性。 Repository: JEen-wang/Adaptive-rag Primary language: Python GitHub stars: 186 GitHub description: 按问题复杂度自动分流的电商客服 RAG：简单问题直接回答；复杂问题经过多跳检索并自我纠错；查单、退改通过工具调用完成，且高风险写操作必须先经人工确认。基于 FastAPI 和 LangGraph，混合检索并附带引用，内置离线评测与 Prometheus 指标。涵盖简历写法和项目面试拷打

### Overview: A sentiment analyzer package for financial assets and securities using GPT models, enabling sentiment analysis of news articles related to specific assets. ### Key Features: - Fetches news articles and content. - Analyzes the sentiment of news articles. - Provides insights for investment and trading decisions. - Supports multiple assets and securities. ### Tech Stack: Python ### Why It Matters: This package is designed to help investors and traders analyze market sentiment through sentiment analysis of news articles, providing valuable insights for informed trading decisions.

Overview: Gaussian Mixture Flow Matching Models (GMFlow) is an official PyTorch implementation of the ICML 2025 paper. It extends diffusion/flow matching models by predicting a Gaussian mixture (GM) distribution of flow velocity, enabling precise few-step sampling and improved image generation quality without the need for classifier-free guidance. Key features: - Gaussian Mixture Output: Expands the network's output layer to predict a GM distribution. - Novel GM-SDE and GM-ODE Solvers: Leverage analytic denoising distributions and velocity fields for precise few-step sampling. - Improved CFG: Mitigates over-saturation issues and enhances image generation quality. - Efficiency: Maintains similar training and inference costs to standard diffusion/flow matching models. Tech stack: Python, PyTorch, CUDA Toolkit 11.8+, PyTorch 2.1+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11.8+, NVIDIA CUDA Toolkit 11

### Overview: Whistle speech-to-text runs on an ESP32-S3 microcontroller, allowing users to control LEDs by simply saying a color in seven languages. This project leverages a custom C99 implementation of the Whistle speech recognition model, which is completely self-contained and runs without network or cloud services. ### Key Features: - **Speech Recognition**: Transcribes spoken commands into specific LED actions. - **Seven Languages Support**: Capable of understanding commands in seven different languages. - **No Network Required**: Runs entirely on the ESP32-S3 without the need for internet or cloud services. - **Custom Model Implementation**: Uses an independent C99 implementation of the Whistle model, checked against the official engine. ### Tech Stack: - **Microcontroller**: ESP32-S3 - **Programming Language**: C99 - **Model**: Custom implementation of Whistle model ### Why It Matters: This project showcases the capabilities of ESP32-S3 in executing complex tasks without the need for external dependencies, making it a valuable tool for developers looking to create projects with low power consumption and minimal resource requirements.

Overview: This project builds a scalable attention-based speech recognition platform using Keras/Tensorflow for inference on the Nvidia Jetson Embedded Computing Platform. The final production model consists of a deep neural network with 3 dilated convolutional layers, 7 bidirectional recurrent layers, an attention layer, and time distributed dense layers. The model was trained on an Nvidia GTX1070 GPU for 30 epochs, achieving an overall cosine similarity of 78% on the test set. Key features: - Scalable attention-based speech recognition platform - Keras/Tensorflow inference on Nvidia Jetson - Deep neural network architecture with dilated convolutions, GRU cells, attention, and time distributed dense layers - Cosine similarity of 78% on test set - 30 epochs training time of 6.5 days Tech stack: Keras/Tensorflow, Nvidia Jetson, Nvidia GTX1070 GPU Why it matters: This project demonstrates the potential of real-time speech recognition on edge devices, with applications in mental health therapy, mobile, robotics, and other areas where cloud-based deep learning is not desirable.

### Overview: This project is a local-first tool designed to turn your notes, web pages, PDFs, and audio/video into a personal knowledge graph, ensuring that when you need to recall something, you get back the exact right piece with its original source, context, and reasoning intact. ### Key Features: - **Multi-format Ingestion**: Automatically extracts and links notes, web pages, Markdown, PDFs, audio/video into the graph. - **Topic-Specific Recall**: Retrieves facts scoped by topic, avoiding irrelevant domains. - **Single-Question Scoring**: Scores retrieved facts for confidence, ensuring reliable answers. - **Decision Tracking**: Logs decisions and their outcomes, closing the feedback loop over time. ### Tech Stack: The project uses Kuzu for the graph database, LanceDB for the vector store, Python 3.11/3.12 for the backend, and React with Node 18+ for the frontend. ### Why It Matters: The system refutes the notion that AI tools should optimize for always having an answer. Instead, it focuses on ensuring that when you need to recall something, you get back the exact right piece with its original source, context, and reasoning intact. This approach makes every answer it does give something you can actually trust, as the refusal rate is one of the quality metrics.

Overview: This project introduces a zero-shot method for transferring motion across objects and scenes using a pre-trained text-to-video diffusion model. The method synthesizes a video that complies with an input text prompt, maintaining the input video's motion and scene layout while transferring the target object's shape and fine-grained motion characteristics. Key features: - Text-driven motion transfer - Transfer motion across drastically different objects and scenes - Utilizes a pre-trained text-to-video diffusion model - Generates videos of 24 frames - Handles out-of-distribution cases - Runs on a single NVIDIA A40 GPU with approximately 32GB of video memory Tech stack: Python Why it matters: This method provides a new approach for zero-shot motion transfer, enabling the synthesis of videos that faithfully represent the input text prompt, even when the target and source objects are significantly different in shape and motion characteristics. This is particularly useful for applications where domain-specific training is not feasible, such as transferring motion across different species or objects in diverse environments.

Overview: **MEt3R** is a differentiable metric designed to evaluate the consistency between generated multi-view images, addressing the limitations of traditional reconstruction metrics in generative models. Key features: - **DUSt3R** backbone for dense 3D point map generation - **DINO** feature comparison in pixel space - Multiple metrics (PSNR, SSIM, LPIPS, RMSE, MSE) for evaluation - Support for nearest, bilinear, and bicubic upsampling methods Tech stack: Python, PyTorch, CUDA, PyTorch3D, FeatUp Why it matters: **MEt3R** provides a novel and effective way to measure the consistency between generated images, independent of the specific scene, making it valuable for evaluating the quality of generative models like multi-view latent diffusion.

Overview: Live2Diff is a pipeline that processes live video streams by a uni-directional video diffusion model. It aims to translate live streams into high-quality video content. Key features: - Uni-directional Temporal Attention with Warmup Mechanism - Multitimestep KV-Cache for Temporal Attention during Inference - Depth Prior for Better Structure Consistency - Compatible with DreamBooth and LoRA for Various Styles - TensorRT Supported Tech stack: Python Why it matters: Live2Diff leverages advanced video diffusion models to create high-quality video translations from live streams, offering a significant improvement in video quality and realism compared to traditional methods. The inclusion of warmup mechanisms and multitimestep KV-cache enhances the model's performance, making it suitable for applications requiring real-time processing and high accuracy. Repository: open-mmlab/Live2Diff Primary language: Python GitHub stars: 199 GitHub description: Live2Diff: A Pipeline that processes live video streams by a uni-directional video Diffusion model.

Overview: This project introduces "Golden Noise for Diffusion Models: A Learning Framework," focusing on the development of a novel learning framework for diffusion models. The work aims to enhance the performance and stability of diffusion models by incorporating golden noise, a technique that improves the quality and reliability of the generated images. Key features: - Implementation of a learning framework that incorporates golden noise for diffusion models. - Availability of inference code and model weights for deployment. - Comprehensive documentation and reproducibility through the use of Hugging Face Model Hub. - Data collection, model training, and evaluation scripts for researchers and developers. Tech stack: The project is built using Python, with dependencies including PyTorch, PIL, numpy, timm, argparse, and einops. Why it matters: This framework addresses the need for improved diffusion models by introducing a method that enhances the quality and reliability of generated images, making it valuable for applications requiring high image fidelity and stability. Repository: xie-lab-ml/Golden-Noise-for-Diffusion-Models GitHub stars: 199 GitHub description: [ICCV2025] The code of our work "Golden Noise for Diffusion Models: A Learning Framework".

### Overview: Distill is a context intelligence layer designed for LLM agents. It provides persistent, deduplicated, and sensitive memory that persists across sessions, preventing re-learning and conflicts, and compressing aging context for better performance. ### Key Features: - **Persistent Memory**: Ensures that the agent retains its knowledge across sessions without re-learning. - **Write-Time Deduplication**: Clusters semantically similar chunks and retains only one representative from each cluster. - **Conflict Detection**: Prevents conflicting information from sitting side by side, ensuring the agent surfaces the most relevant version. - **Hierarchical Decay**: Compresses context over time, keeping only the most relevant information. - **No LLM Calls**: Provides deterministic, deterministic, and fully auditable results. ### Tech Stack: Go ### Why It Matters: Distill is essential for agents that need to remember what matters, preventing non-deterministic outputs and maintaining a clear signal. It optimizes performance by reducing redundant context and preventing re-learning, making it a crucial component for building robust and reliable AI agents. Repository: Siddhant-K-code/distill Primary Language: Go GitHub Stars: 198 GitHub Description: Context intelligence layer for LLM agents: persistent memory with write-time dedup, sensitivity tagging, conflict detection, and hierarchical decay. ~12ms. No LLM calls. MIT.

### Overview: This repository contains a comprehensive survey of large language models (LLMs) applied to protein understanding and prediction. It covers various methods, including sequence-based deep representation learning, protein structure prediction, and evolutionary scale modeling. ### Key Features: - **Unified Rational Protein Engineering with Sequence-Based Deep Representation Learning**: A paper that uses deep learning to rationalize protein engineering. - **Learning Protein Sequence Embeddings Using Information from Structure**: An approach that leverages information from protein structure to embed sequences. - **Mutation Effect Estimation on Protein–Protein Interactions Using Deep Contextualized Representation Learning**: A method for estimating mutation effects on protein interactions. - **Prottrans: Toward Understanding the Language of Life through Self-Supervised Learning**: A paper that uses self-supervised learning to understand the language of life. - **Modeling Protein Using Large-Scale Pretrained Language Model**: A survey of large-scale pretrained language models applied to protein modeling. - **Single-Sequence Protein Structure Prediction Using a Language Model and Deep Learning**: A method for predicting protein structures using language models. - **Bertology Meets Biology: Interpreting Attention in Protein Language Models**: A paper that interprets attention in protein language models. - **Learning the Language of Viral Evolution and Escape**: A study that interprets attention in protein language models. - **TCR-BERT: Learning the Grammar of T-Cell Receptors for Flexible Antigen-Xbinding Analyses**: A method for predicting T-cell receptor binding. - **xTrimoPGLM: Unified 100B-Scale

### Overview: 将普通文本转化为具有真实手写体效果的图片：通过选择手写字体、信纸背景以及随机扰动（如字距、行距、字号、笔画位移、笔画旋转）来实现每个字、每一页的独特性。 ### Key features: - **图形界面（GUI）**：支持输入文本、选择字体/背景、调整排版参数，实时预览渲染效果，支持多页翻页与一键导出 PNG/PDF。 - **命令行（CLI）**：纯文本或 docx 输入，批量生成手写图片，所有排版、扰动、颜色参数均可通过命令行指定。 - **随机扰动**：字距、行距、字号、笔画位移、笔画旋转等参数均采用随机扰动，确保每个字、每一页的独一无二。 - **性能优化**：基于 `numpy` + `scipy` 的高性能渲染引擎，预览渲染页约 0.15 秒，实时交互流畅。 ### Tech stack: - **语言**：Python 3.14 - **依赖管理**：uv（`pyproject.toml` + `uv.lock`） - **界面**：PyQt6（纯 Qt 控件 + 自动布局，无背景图片依赖） - **渲染引擎**：numpy + scipy（FastEngine，默认）；handright 8.2.0（可选经典后端） - **图像处理**：Pillow - **PDF

Overview: Diff-UNet is a novel method that applies diffusion models to address 3D medical image segmentation, achieving superior accuracy compared to existing techniques. Key features: Diffusion Embedded Network, 3D Medical Image Segmentation, BraTS2020 and BTCV datasets. Tech stack: Python. Why it matters: The method demonstrates significant improvements in accuracy for 3D medical image segmentation tasks, making it a valuable addition to the field.

### Overview: The Meilisearch MCP Server is a Model Context Protocol (MCP) server designed to integrate with Meilisearch, enabling any MCP-compatible client to interact with the search engine through natural language conversation. This server simplifies the process of managing search indices and performing searches, making it easier for AI assistants to leverage Meilisearch's capabilities without needing to learn its API. ### Key Features: - 📊 Index & Document Management: Create, update, and manage search indices. - 🔍 Smart Search: Search across single or multiple indices with advanced filtering. - ⚙️ Settings Configuration: Fine-tune search relevancy and performance. - 📈 Task Monitoring: Track indexing progress and system operations. - 🔐 API Key Management: Secure access control. - 🏥 Health Monitoring: Keep tabs on your Meilisearch instance. - 🧱 Structured Tool Output: Every data-returning tool provides `structuredContent` alongside text, per the latest MCP specification. - 🔀 Modern MCP Stack: Built on the official MCP Python SDK v2 (spec 2026-07-28), backwards compatible with older MCP clients. - 🐍 Python Implementation: [TypeScript version also available](https://github.com/devlimelabs/meilisearch-ts-mcp). ### Why it Matters: - 🤖 Universal Compatibility: Works with any MCP client, not just Claude. - 🗣️ Natural Language Control: Manage Meilisearch through conversation with any LLM. - 🚀 Zero

Overview: Scratchpad is a modular reasoning framework designed for sophisticated AI assistants and power users, focusing on user-context and user-intent. It deconstructs ideas logically from A to B to Z, providing deeper and more reliable insights through structured cognitive operations. Key features: - **Modular Architecture**: Choose from standard reasoning templates, purpose-built domain tools, or persona-based frameworks. - **Focus on User Intent and Context**: Breaks down user intent and context to provide more relevant and accurate responses. - **Dynamic Contextual Adaptation**: Adapts to changes in user context dynamically, ensuring ongoing relevance in responses. - **Transparent Reasoning "Blackbox"**: Establishes a visible framework for the model's reasoning, enabling users to verify logic and confirm understanding of user context. - **Multi-Step Reasoning**: Identifies key elements, analyzes user perspectives, and applies cognitive operations like abstraction, comparison, and inference. - **Calibration of User Intent**: Aims to calibrate the model to align closely with the user's intent and context, enhancing relevance and accuracy. - **Advanced Logic**: Utilizes advanced reasoning to connect the dots and provide insightful responses. - **Error Detection and Correction**: Detects and corrects potential errors in reasoning, enhancing response reliability. - **Persona & Domain Specialization**: Specialized frameworks for specific fields (game development, research, creative writing) and reasoning personas. Tech stack: Scratchpad is built on a modular architecture and utilizes advanced reasoning techniques to provide deeper insights. Why it matters: Scratchpad is essential for AI assistants and power users seeking transparent

Overview: This repository provides a PyTorch implementation of common image generation metrics, including Inception Score and Fréchet Inception Distance. Key features: - Supports Inception Score and Fréchet Inception Distance metrics. - Integrates computation procedures to avoid multiple forward passes. - Handles large-scale image computations efficiently. - Supports GPU computation for faster performance. - Ready to use with pre-calculated FID reference for CIFAR-10 dataset. Tech stack: Python Why it matters: This package simplifies the process of evaluating image generation metrics, making it easier for developers to assess the quality of their models and compare results with official implementations.