Projects

### Overview: Welcome to Awesome-Deep-Research! This repository serves as your comprehensive guide to the cutting-edge world of Agentic Deep Research. We've meticulously curated a collection of resources for researchers, developers, and enthusiasts. ### Key Features: - 📖 Latest research papers and breakthroughs - 🚀 Open-source implementations and tools - 🤝 Contributing and Citations ### Tech Stack: Python, JavaScript, Markdown, and various open-source tools. ### Why it Matters: This repository is essential for anyone looking to explore the intersection of artificial intelligence and autonomous agents. It provides a wealth of resources and insights into the latest research and practical applications in the field.

Overview: Crabwalk is a real-time companion monitor designed to provide insights into the activities of OpenClaw agents across various messaging platforms, including WhatsApp, Telegram, Discord, and Slack. Key features: - Live activity graph visualization - Multi-platform monitoring capability - Real-time streaming from OpenClaw gateway - Action tracing for detailed tool inspection - Session filtering for targeted monitoring Tech stack: TypeScript Why it matters: Crabwalk simplifies the monitoring of OpenClaw agents by providing a comprehensive and intuitive interface, allowing developers to quickly identify issues and improve the performance of their AI applications. Repository: crabwise-ai/crabwalk GitHub stars: 873 GitHub description: 🦀 Crabwalk 🦀 Real-time companion monitor for OpenClaw agents.

### Overview: uniem 是一个致力于创建中文最好的通用文本嵌入模型的项目。它提供了一系列模型的训练、微调和评测代码，旨在简化和加速文本处理任务。 ### Key features: - FineTuner 支持多种模型微调，包括 M3E、sentence_transformers 和 SGPT。 - 实现了原生支持模型微调的 `FineTuner`，几行代码即可适配。 - 支持多种数据集和任务，如文本分类和文本检索。 - 提供了中文文本分类和文本检索的自动评测标准 MTEB-zh。 ### Tech stack: Python ### Why it matters: uniem 项目为中文文本处理任务提供了高效、统一的解决方案，简化了模型微调和评测过程，使得开发者能够专注于业务逻辑，无需担心复杂的模型训练和评估问题。

### Overview `Open Long-Tailed Recognition (OLTR)` is a PyTorch implementation of the long-tail recognizer presented in the paper "[Large-Scale Long-Tailed Recognition in an Open World](https://arxiv.org/abs/1904.05160)" by Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. The project was presented at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2019, receiving an **Oral Presentation** award. ### Key Features - **Modified Attention Module**: The attention module in the second stage of the Places-LT dataset has been updated to avoid issues with freezed weights. - **Updated Configuration Files**: The configuration files for the Places-LT dataset have been updated to improve the results. - **Fixed Bugs**: Bugs in the `utils.py` and `run_network.py` files have been fixed to ensure proper model training and reproducibility. - **Reproduced Results**: The results for ImageNet-LT and Places-LT have been updated to match the reported numbers in the paper. ### Tech Stack - **Python**: The project is implemented using Python 3. - **PyTorch**: The code is based on the PyTorch framework. ### Why It Matters This project provides a practical implementation of the long-tail recognition technique, which is crucial for applications in computer vision and image recognition tasks where the dataset is highly imbalanced. The

Overview: MedicalTorch is an open-source medical imaging framework built with PyTorch, designed to handle various datasets and preprocessing steps for medical imaging tasks. Key features: - Extensive set of loaders and pre-processors for medical imaging - Support for NIFTI datasets - Documentation available at [readthedocs](http://medicaltorch.readthedocs.io) Tech stack: Python Why it matters: MedicalTorch provides a comprehensive toolkit for researchers and developers working on medical imaging projects, streamlining the process of data preparation and model training.

Overview: This repository contains an implementation of our research paper "Baking Gaussian Splatting into Diffusion Denoiser for Fast and Scalable Single-stage Image-to-3D Generation and Reconstruction" at the 2025 International Conference on Computer Vision (ICCV). The code and checkpoints are a re-implementation and re-training of the original version developed at Adobe, offering a single-stage approach that does not require a depth estimator and can generate single-view 3D object models in under 6 seconds. Key features: - Single-stage image-to-3D generation - Mesh exportation - Scene reconstruction - Fast inference time (~6 seconds) Tech stack: Python Why it matters: This method provides a significant improvement in speed and scalability compared to existing state-of-the-art techniques, making it highly useful for applications requiring real-time 3D generation from single images.

### Overview SGPT: GPT Sentence Embeddings for Semantic Search is a repository containing code, results, and pre-trained models for the paper "SGPT: GPT Sentence Embeddings for Semantic Search" (arXiv:2202.08904). This repository provides a unified approach to applying GPT models as Bi-Encoders or Cross-Encoders for symmetric or asymmetric semantic search, with improved performance compared to existing methods. ### Key Features - **SGPT-BE (Symmetric Semantic Search BE)**: Produces semantically meaningful sentence embeddings by contrastive fine-tuning of bias tensors and position-weighted mean pooling. - **SGPT-CE (Symmetric Semantic Search CE)**: Uses log probabilities from GPT models without any fine-tuning. - **SGPT-BE (Asymmetric Semantic Search BE)**: Similar to SGPT-BE but designed for asymmetric search. - **SGPT-CE (Asymmetric Semantic Search CE)**: Uses log probabilities from GPT models without any fine-tuning. ### Tech Stack The repository uses Jupyter Notebook for code implementation and experimentation. ### Why It Matters SGPT provides a more efficient and effective way to apply GPT models for semantic search, offering improved performance compared to existing methods. This makes it a valuable tool for researchers and developers looking to enhance their semantic search capabilities. ### GitHub Stars The repository has been starred by 870 users, indicating its popularity and utility in the community. ### Structure The repository is organized into several directories: - **biencoder

Overview: This repository contains notes and implementations for the Reinforcement Learning course by David Silver, focusing on various algorithms discussed in the course along with their implementations using Keras and OpenAI's Gym framework. Key features: - Implementation of different reinforcement learning algorithms - Notes on Reinforcement Learning course by David Silver - Use of Keras and OpenAI's Gym framework - Implementation of algorithms in both TensorFlow and Python Tech stack: Python, Keras, TensorFlow, Gym, Numpy Why it matters: This repository serves as a comprehensive resource for students and researchers interested in Reinforcement Learning, providing practical implementations of theoretical concepts discussed in the course.

Overview: This toolkit provides a config-based CLI tool for fine-tuning, ablating, and unit-testing open-source language models (LLMs). It simplifies the process of launching a series of experiments with customizable prompts, LLMs, optimization strategies, and quality assurance. Key features: - Config-based CLI for launching fine-tuning experiments - Support for various LLMs from Hugging Face and other open-source models - Customizable prompts and data ingestion - Ability to ablate and test LLMs independently - One YAML configuration file for managing all experiment details Tech stack: Python Why it matters: This toolkit streamlines the process of fine-tuning and testing LLMs, making it easier for developers to experiment with different models and configurations without the complexity of managing multiple scripts or environments.

### Overview: CLODEx is a local-first, zero-trust agentic IDE designed for long-running engineering work. It provides a durable, task-focused workspace where code, Git, terminal, browser, models, and MCP tools are kept inside a single, secure desktop environment. ### Key Features: - **Durable Workspace**: Keeps task history and allows recovery after restarts. - **Integrated Tools**: Seamless navigation between files, diffs, Git, terminal, browser, and MCP without leaving the task. - **Community Observed 21**: Available for macOS, Windows, and Linux, with a verified source and technical preview. ### Tech Stack: TypeScript ### Why It Matters: CLODEx offers a durable, task-focused workspace that is essential for long-running engineering work, ensuring that code, Git, terminal, browser, models, and MCP tools are kept inside a single, secure desktop environment. It provides a seamless integration experience where files, Git, terminal, browser, and MCP share one workspace, making it perfect for understanding repositories, planning changes, editing multiple files, running commands, inspecting applications, and reviewing diffs.

Overview: This repository contains a collection of AI projects across various areas of machine learning, including convolutional neural networks, transfer learning, cloud-scale text classification, data generation, dimensionality reduction, distributed training, and fraud detection. The projects are accompanied by articles on the blog, and the primary language is Jupyter Notebook. Key features: - Introduction to Convolutional Neural Networks - Introduction to Transfer Learning - Cloud-Scale Text Classification With Convolutional Neural Networks - Introduction to Data Generation - Introduction to Dimensionality Reduction with t-SNE - Introduction to Distributed Training with DeepSpeed - Introduction to Fraud Detection Tech stack: Jupyter Notebook Why it matters: These projects provide practical examples and explanations of various machine learning techniques, making them valuable resources for developers looking to deepen their understanding of AI and machine learning.

### Overview: KiwiFS is a Markdown filesystem designed for agents and teams. It provides searchable, structured, and versioned markdown files, eliminating the need for databases, read-only retrieval layers, or ephemeral sandboxes. KiwiFS is a single binary, zero-config, and supports multiple access protocols, including REST, MCP, NFS, S3, WebDAV, and FUSE. ### Key Features: - **Searchable**: Files are indexed for quick and efficient searching. - **Structured**: The files are the source of truth, with a structured index for easy querying. - **Versioned**: Every write is an atomic commit, ensuring data integrity and history. - **Git Versioning**: Every write is an atomic commit, providing a clear audit trail. - **Full-text + Vector Search**: Uses BM25 via SQLite FTS5 for full-text search and OpenAI, Ollama, ONNX, and other vector search engines for more advanced queries. - **DQL Queries**: SQL-like queries over frontmatter for data manipulation. - **6 Access Protocols**: Supports REST, MCP, NFS, S3, WebDAV, and FUSE. - **19 Data Importers**: Supports Postgres, MySQL, MongoDB, Notion, CSV, Obsidian, and more for easy data ingestion. ### Tech Stack: Go 1.27 ### Why It Matters: KiwiFS makes markdown files writable, searchable, queryable, versioned, and human-readable. It eliminates the need for databases, read-only retrieval layers,

Overview: PPASR是一款基于PaddlePaddle实现的自动语音识别框架，旨在提供简单实用的语音识别项目，支持多种模型和预处理方法。 Key features: - 支持多个语音识别模型，包括DeepSpeech2、Conformer、Squeezeformer等。 - 支持多种解码器，如ctc_greedy_search、ctc_prefix_beam_search、attention_rescoring等。 - 支持多种预处理方法，如fbank、mfcc等。 - 支持多种数据增强方法，如噪声增强、混响增强等。 - 支持多种推理方法，如短音频推理、长音频推理、流式推理等。 - 更多特点等待发现。 Tech stack: Python, PaddlePaddle 2.6.1 Why it matters: 本项目提供了一个简单实用的语音识别框架，适合部署在服务器、Nvidia Jetson设备以及未来支持的Android等移动设备。项目结构优化，降低了使用难度，预处理库改用kaldi_native_fbank，支持多种语言的处理和混合训练。此外，模型下载资源丰富，包括预训练模型和测试集。 Repository: yeyupiaoling/PPASR GitHub stars: 869 GitHub description: 基于PaddlePaddle实现端到端中文语音识别，从入门到实战，超简单的入门案例，超实用的企业项目。支持当前最流行的DeepSpeech2、Conformer、Squeezeformer模型。

### Overview: PoinTr: Diverse Point Cloud Completion with Geometry-Aware Transformers is a transformer-based model designed for point cloud completion. This repository contains the PyTorch implementation for the paper "PoinTr: Diverse Point Cloud Completion with Geometry-Aware Transformers" (ICCV 2021 Oral Presentation). The model converts point clouds into sequences of point proxies and uses a transformer architecture for generation, making it particularly useful for diverse point cloud completion tasks. ### Key Features: - **Transformer Architecture**: The model employs a transformer-based encoder-decoder architecture for point cloud completion. - **Geometry-Aware Transformers**: It incorporates geometry-aware transformers to enhance the model's ability to handle diverse and incomplete point clouds. - **Diverse Incomplete Point Clouds**: The repository includes two challenging benchmarks, ShapeNet-55 and ShapeNet-34, with more diverse incomplete point clouds to better reflect real-world scenarios. ### Tech Stack: - **Python**: The implementation is written in Python. ### Why It Matters: PoinTr is significant because it addresses the challenge of point cloud completion in a diverse and complex manner, making it a valuable tool for researchers and practitioners in the field of point cloud processing and computer vision. The inclusion of pretrained models and benchmarks makes it accessible and useful for further research and development.

### Overview: OpenSwarm is an agentic CI/CD gate designed to review, fix, and verify changes in pull requests and local commits. It leverages Claude Code CLI, OpenAI Codex/GPT, and local models like Ollama and LM Studio for efficient and deterministic verification. ### Key Features: - **Review Gate**: Reviews commits or working-tree diffs, fixes confirmed findings, and re-enters the review loop. - **Linear Integration**: Supports integration with Linear for seamless workflow. - **Cognitive Memory**: Stores and recalls task outcomes for future prompts, enhancing decision-making. ### Tech Stack: - TypeScript ### Why It Matters: OpenSwarm is a powerful tool for CI/CD gatekeeping, ensuring that only verified changes are merged into the repository. It simplifies the review process, reduces manual intervention, and automates the verification phase, making the development pipeline more efficient and reliable. ### Repository: Intrect-io/OpenSwarm - GitHub Stars: 867 - Primary Language: TypeScript - GitHub Description: OpenSwarm — Autonomous AI dev team orchestrator powered by Claude Code CLI. Discord control, Linear integration, cognitive memory. - GitHub Sponsorship: Pledged by Atlas Cloud for enterprise AI infrastructure support. ### Quick Start: ```bash npm install -g @intrect/openswarm export OPENROUTER_API_KEY=… # or use an authenticated codex/claude/local provider openswarm review --path . --read-only # review the current diff without mutation openswarm review --max --fix

Overview: Neum AI is a powerful framework designed to manage and synchronize vector embeddings at massive scales, making it an essential tool for developers looking to leverage their data for contextualized Large Language Models. Key features: - High throughput distributed architecture for handling billions of data points. - Built-in data connectors for common data sources and embedding services. - Real-time synchronization to ensure data sources are always up-to-date. - Customizable data pre-processing for efficient data management. - Cohesive data management with metadata tracking for rich retrieval experiences. Tech stack: Python Why it matters: Neum AI simplifies the integration and management of vector embeddings, making it a crucial tool for developers working with large language models, ensuring efficient data management and context-aware retrieval.

### Overview: An Eigen-based, lightweight C++ interface to Nonlinear Programming solvers (Ipopt, Snopt). ### Key Features: - **Solver Independence**: Combines the advantages of Ipopt and Snopt with modern, intuitive formulations using Eigen. - **Automatic Index Management**: Formulates variables and constraints independently in separate classes, ensuring efficient implementation. - **Integration**: Pure cmake `find_package(ifopt)` or catkin/ros integration. - **Lightweight**: ~2k lines of code, making it easy to use and extend. ### Tech Stack: C++ ### Why It Matters: Modern, intuitive formulations of vectors and matrices in Eigen, combined with high-quality solvers for nonlinear optimization, makes `ifopt` a powerful tool for solving complex optimization problems efficiently.

Overview: KeyForge3D is an app that converts a photo of a key into a 3D-printable STL file, ideal for locksmiths and hobbyists. Key features: - Key Shape Extraction: Upload a photo of a key to extract its 2D profile. - Bitting Analysis: Automatically detect and analyze the key's bitting pattern. - 3D Model Generation: Convert the 2D key profile into a 3D model with accurate cuts and dimensions. - STL Export: Export the 3D model as an STL file for 3D printing. - User-Friendly GUI: Simple interface to upload images and generate models with a single click. - Scalability: Scale the key to real-world dimensions (default: 1 pixel = 0.1 mm; can be adjusted with a reference object). Tech stack: Python Why it matters: KeyForge3D is a valuable tool for locksmiths and hobbyists, providing a quick and accurate way to replicate keys using 3D printing technology. Repository: https://github.com/makalin/KeyForge3D GitHub stars: 867 GitHub description: KeyForge3D is an app that extracts the shape of a physical key from a photo, generates a 3D model of the key, and exports it as an STL file for 3D printing.

### Overview: ProjectMem is an open-source tool designed to enhance the memory and judgment of AI coding agents. It records issues, attempts, fixes, and decisions, providing a local-first memory layer that warns the agent before repeating failed approaches. ProjectMem is optimized for performance, running on a native MCP server without cloud resources or telemetry, ensuring 100% local operation. ### Key Features: - **Memory Recording**: Tracks past issues, attempts, fixes, and decisions to prevent re-reading and re-trying failed strategies. - **Local Operation**: Runs on a native MCP server, ensuring local memory and judgment without cloud resources or telemetry. - **100% Local**: No reliance on cloud services, ensuring data privacy and security. ### Tech Stack: - Python ### Why It Matters: ProjectMem is crucial for developers and AI engineers who want to improve the efficiency and effectiveness of their coding processes. By leveraging ProjectMem's memory and judgment layers, teams can avoid repeating mistakes and make informed decisions, ultimately leading to better project outcomes and faster development cycles. ### Repository: - **Repository**: [GitHub Repository](https://github.com/riponcm/projectmem) - **Stars**: 866 - **Description**: Open-source coding agents memory. Records issues, attempts, fixes, and decisions, then warns your agent before it repeats an approach that already failed. Native MCP server for Claude Code, Cursor, Antigravity, and Codex. 100% local, no cloud, no telemetry. MIT License. ### Primary Language: - **Primary Language

Overview: Statespace is a powerful tool for transforming CLI tools and files into shareable data apps for AI agents, allowing them to discover and use your data easily. Key features: - Any CLI tool can be made shareable - Safe by default with regex constraints - Self-describing Markdown pages - Composable across multiple pages - Shareable via URL, MCP server, or direct agent access Tech stack: Rust Why it matters: Statespace simplifies the process of making your data accessible to AI agents, making it easier for them to discover and use your data without needing to know the specific tools or commands you use.