Projects

### Overview: DocsAgent is a powerful MCP server designed to give AI agents instant, private access to your personal knowledge base. It supports Zotero, Obsidian, and Apple Notes, allowing agents like Claude, Cursor, Cline, and Qwen Code to search, read, and write your library directly on your machine. ### Key Features: - **Local-first & private**: The engine reads your Zotero library, Obsidian vault, and Apple Notes store directly on your machine, ensuring your notes and PDFs never leave your device. - **Three sources, one query**: Zotero items/annotations/notes, Obsidian vaults, and Apple Notes (macOS only); mixed search fuses all three by reciprocal rank. - **Native C++ search core**: Inverted-index BM25 + passage ranking, millisecond lookup, low memory footprint (160–227 MB for a 1,500-paper library). - **8 MCP tools**: 5 read + 3 write, with JSON-schema validated arguments, token budgets, result dedup, and a three-layer write safety gate. - **Two transports**: stdio for local MCP clients, Streamable HTTP for remote deployment (origin checks, API-key / OAuth 2.0 token introspection, per-request RBAC, `/health` probe). - **Two shells, one core**: This TypeScript package and a feature-equal Python wrapper ship the same tools over the same JSON-RPC contract. ### Tech Stack: TypeScript ### Why It Matters: DocsAgent simplifies access to your

### Overview: Actionformer is a Transformer-based model designed for temporal action localization, achieving state-of-the-art performance on various datasets including THUMOS14, ActivityNet, and EPIC-Kitchens 100. It simplifies the design by removing unnecessary bells and whistles and focuses on minimalistic yet effective solutions. ### Key features: - **Transformer-based Model**: Uses Transformers for temporal context modeling. - **Minimalistic Design**: Simplified architecture without bells and whistles. - **High Performance**: Achieves 71.0% mAP on THUMOS14, 36.56% on ActivityNet, and 42.54% on EPIC-Kitchens 100. - **Single Shot Localization**: Localizes moments of actions in a single shot without action proposals or pre-defined anchor windows. ### Tech stack: Python ### Why it matters: Actionformer is a significant advancement in the field of temporal action localization, offering a minimalistic yet effective solution that significantly outperforms previous models. Its simplicity and effectiveness make it a valuable tool for researchers and practitioners looking to improve action localization in videos.

### Overview: This repository contains fine-tuning projects and notes on common practical techniques for large language models (LLMs). It includes detailed Jupyter Notebook code for various fine-tuning tasks, such as LLaMA, CodeLLaMA, Mistral, Falcon, and Phi-1.5, as well as fine-tuning on custom datasets and using YouTube videos for explanations. ### Key features: - LLaMA-3-8B fine-tuning with ORPO - LLaMA-3-8B fine-tuning on custom dataset - LLaMA-3-8B fine-tuning with unsloth - LLaMA-3-8B fine-tuning with PEFT and QLoRA - LLaMA-3-8B fine-tuning with QLoRA - LLaMA-3-8B fine-tuning with QLoRA and ORPO - LLaMA-3-8B fine-tuning with QLoRA and ORPO and unsloth - LLaMA-3-8B fine-tuning with QLoRA and ORPO and unsloth and unsloth - LLaMA-3-8B fine-tuning with QLoRA and ORPO and unsloth and unsloth and unsloth - LLaMA-3-8B fine-tuning with QLoRA and ORPO and unsloth and unsloth and unsloth and unsloth - LLaMA-3-8B fine-tuning with QLoRA and ORPO and unsloth and uns

Overview: This project is a code repository for the AAAI 2018 publication "SEE: Towards Semi-Supervised End-to-End Scene Text Recognition". It includes a preprint on Arxiv and detailed installation instructions for running the project on a local machine or Docker container. Key features: - Installation guide for both direct installation and Docker container usage - General training hints for adjusting or deleting specific lines in the code - SVHN dataset experiments Tech stack: Python Why it matters: The project provides a comprehensive solution for semi-supervised scene text recognition, demonstrating advancements in deep learning techniques for computer vision tasks. Repository: Bartzi/see GitHub stars: 577 GitHub description: Code for the AAAI 2018 publication "SEE: Towards Semi-Supervised End-to-End Scene Text Recognition".

Overview: Platform for Situated Intelligence (\\psi) is an open-source framework designed for the development and research of multimodal, integrative-AI systems, enabling applications that process streaming sensor data and operate under latency constraints. Key features: - A modern, performant infrastructure for working with multimodal, temporally streaming data. - A set of tools for multimodal data visualization, annotation, and processing. - An ecosystem of components for various sensors, processing technologies, and effectors. Tech stack: C# Why it matters: \\psi provides a modern, extensible framework for developing and researching multimodal, integrative-AI systems, making it a valuable tool for researchers and developers in fields such as robotics, mixed-reality, and smart spaces.

Overview: Sportsipy is a free sports API written in Python that provides access to comprehensive statistics and information from major sports leagues in North America and around the world. Key features: - Access to statistics and information from major sports leagues in North America, including MLB, NBA, College Football and Basketball, NFL, and NHL. - Support for Professional Football (soccer) for thousands of teams from leagues around the world. - A plethora of sports information, including team lists, game dates, and detailed metrics. Tech stack: Python Why it matters: Sportsipy simplifies the process of integrating sports data into Python-based applications, particularly for data analytics and machine learning projects. It provides a convenient way to access detailed statistics and information from various sports leagues, making it easier to analyze and visualize sports data.

Overview: This repository provides a comprehensive, easily extendable pipeline for testing and implementing various deep metric learning methods. It includes implementations and dataloaders for loss functions, sampling methods, datasets, and architectures. Key features: - Loss Functions: Triplet Loss, Margin Loss, ProxyNCA, N-Pair Loss - Sampling Methods: Random Sampling, Soft Hard Sampling, Semihard Sampling, Distance Sampling, N-Pair Sampling - Datasets: CUB200-2011, CARS196, Stanford Online Products, In-Shop Clothes, PKU Vehicle-ID (optional) - Architectures: ResNet50, GoogLeNet Tech stack: Python Why it matters: This repository offers a thorough comparison and evaluation of deep metric learning, providing more criteria, metrics, and logging options compared to the original Revisiting Deep Metric Learning paper. Repository: Confusezius/Deep-Metric-Learning-Baselines GitHub stars: 576

Overview: Compartment is a fully offline, encrypted memory system designed for AI agents, providing persistent storage, recall, and integration with various MCP clients. Key features: - Encrypted, fully offline memory - One-click install via Cursor, VS Code, or Kiro - Persistent memory stored on your own computer - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with MCP clients - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients - Supports integration with Claude Code, Claude Desktop, Hermes Agent, OpenClaw, Cursor, and Codex - Supports integration with various MCP clients -

### Overview: Adaptive Classifier is a powerful PyTorch-based library for dynamic text classification that offers continuous learning, zero-downtime updates, and advanced defense mechanisms against adversarial inputs and strategic manipulation. ### Key Features: - **🚀 Universal Compatibility**: Works with any HuggingFace transformer model. - **⚡ Optimized Inference**: Built-in ONNX Runtime for 2-4x faster CPU predictions. - **📈 Continuous Learning**: Adds new examples without catastrophic forgetting. - **🔄 Dynamic Classes**: Adds new classes at runtime without retraining. - **⏱️ Zero Downtime**: Updates models in production without service interruption. - **✅ Knows What It Doesn't Know**: Calibrated confidence, prediction sets, out-of-distribution detection, and abstention. - **🧹 Fixable**: Forget a wrong example or a whole class without retraining. - **🎣 Active Learning & Drift**: Suggests which texts to label next and detects when traffic has shifted. - **🔌 Easy to Adopt**: Scikit-learn interface, HTTP server, async methods, and ONNX inference outside Python. ### Why It Matters: Adaptive Classifier is crucial for enterprises looking for robust and scalable text classification solutions. Its continuous learning capabilities ensure that models stay up-to-date with evolving data, while its advanced defense mechanisms protect against adversarial attacks and strategic manipulation, making it a valuable tool in dynamic and competitive environments. ### Tech Stack: Python ### GitHub Stars: 576

Overview: One-DM is an innovative model designed for generating stylized handwritten text, specifically for the ECCV 2024 paper. It requires only a single reference sample to imitate the writing style, showcasing remarkable performance even with a single reference. Key features: - One-Shot Diffusion Mimicker (One-DM) - Supports English, Chinese, and Japanese datasets - Utilizes high-frequency components from the reference sample to enhance handwriting style extraction - Outperforms previous methods by 15x with fewer references Tech stack: Python Why it matters: One-DM provides a significant improvement in stylized text generation, particularly in scenarios where multiple references are not feasible, making it a valuable tool in various applications requiring high-quality handwritten text synthesis. Repository: dailenson/One-DM Primary language: Python GitHub stars: 575 GitHub description: Official Code for ECCV 2024 paper — One-Shot Diffusion Mimicker for Handwritten Text Generation

### Overview: This repository provides curated datasets and resources for Natural Language Processing (NLP) in Bahasa Indonesia, including Named Entity Recognition, POS tagging, question and answering, paraphrasing, text summarization, hate speech detection, word analogy, formal-informal classification, multilingual parallel text, and unsupervised corpora. ### Key features: - Named Entity Recognition (NER) datasets - Indonesian Part-of-Speech (POS) Tagging corpus - Question and Answering resources - Paraphrasing techniques - Text Summarization tools - Hate Speech detection models - Word Analogy challenges - Formal-Informal classification models - Multilingual Parallel text corpora - Unsupervised Corpus datasets ### Tech stack: Python, Git, Docker, and various NLP libraries and tools. ### Why it matters: This repository serves as a comprehensive resource for researchers, developers, and language enthusiasts looking to enhance their NLP skills in Bahasa Indonesia. It provides a wide range of datasets and tools that can be used for various NLP tasks, making it an essential tool for anyone working on language processing in the Indonesian language.

### Overview: CycleISP is a novel framework that addresses the limitations of synthetic data in image denoising by leveraging the real camera imaging pipeline. This paper presents a method to generate realistic image pairs for denoising tasks, achieving state-of-the-art performance on real camera benchmark datasets. ### Key Features: - **Real Data Synthesis**: CycleISP generates realistic image pairs by converting sRGB images to RAW data and back, allowing for the training of image denoising networks on real camera data. - **Improved Data Synthesis**: The framework models the camera imaging pipeline in both forward and reverse directions, enabling the creation of diverse and realistic image pairs. - **Parameter Efficiency**: The model parameters are significantly reduced compared to previous methods, making it more efficient and practical for real-world applications. ### Tech Stack: - **Python**: The core implementation is written in Python, utilizing PyTorch for deep learning tasks. - **CUDA**: The framework is optimized for NVIDIA GPUs, leveraging CUDA for faster computations. ### Why It Matters: - **Realistic Data**: By using real camera data, CycleISP ensures that the denoising algorithms perform better on real-world images, as they are more representative of typical camera noise. - **Efficiency**: The framework reduces the computational complexity and parameter size, making it more practical and efficient for real applications. - **Versatility**: The method can be extended beyond image denoising, demonstrating its potential in other fields like color matching in stereoscopic cinema. ### Repository: - **Repository**: `swz30/CycleISP` - **GitHub Stars**:

### Overview: This repository is dedicated to collecting tools, datasets, and papers related to simultaneous translation research. It serves as a comprehensive resource for researchers and practitioners interested in the field of simultaneous translation. ### Key Features: - **Simultaneous Translation Tools**: Includes Fairseq, a sequence modeling toolkit, and SimulEval, a general evaluation framework for simultaneous translation. - **Datasets**: Covers conventional text-to-text translation datasets like IWSLT15 English-Vietnamese, WMT15 German-English, and WMT14 English-French. Also includes datasets for speech-to-text translation, speech-to-speech translation, and simultaneous interpretation. - **Tutorials and Talks**: Contains presentations and talks from notable conferences and workshops, such as PACLIC 2016 and EMNLP 2020. ### Tech Stack: - Python and PyTorch are used for implementing sequence models and machine translation tools. ### Why It Matters: This repository is a valuable resource for researchers and practitioners looking to stay updated on the latest developments in simultaneous translation. It provides a wealth of information and tools that can be directly applied to improve translation models and techniques. Repository: zhangshaolei1998/Awesome-Simultaneous-Translation GitHub Stars: 575

Overview: DynSLAM is a dense Simultaneous Localization and Mapping system designed for dynamic environments. It reconstructs both the static environment and individual dynamic objects like cars from stereo input and video sequences. Key features: - Support for stereo input - Outdoor operation capability - Voxel garbage collection for efficient map storage - Separate reconstruction of static and dynamic objects Tech stack: C++ with Jupyter Notebook Why it matters: DynSLAM is a significant contribution to the field of computer vision and robotics, offering robust solutions for dense mapping in large-scale dynamic environments, which is crucial for applications in robotics, autonomous vehicles, and augmented reality. Repository: AndreiBarsan/DynSLAM Primary language: Jupyter Notebook GitHub stars: 575 GitHub description: Master's Thesis on Simultaneous Localization and Mapping in dynamic environments. Separately reconstructs both the static environment and the dynamic objects from it, such as cars. README excerpt: DynSLAM: Simultaneous Localization and Mapping in Dynamic Environments This is a dense SLAM system written in C++. It builds on [InfiniTAM](https://github.com/victorprad/InfiniTAM), adding support for stereo input, outdoor operation, voxel garbage collection, and separate dynamic object (e.g., car) reconstruction. Developed as part of my Master's Thesis, it is accepted to ICRA 2018 accompanying the paper "Robust Dense Mapping for Large-Scale Dynamic Environments" by Andrei Bârsan, Peidong Liu, Marc

Overview: R-Net is a Tensorflow implementation of R-NET, a model designed for machine reading comprehension specifically for the SQuAD dataset. Key features: - Scaled multiplicative attention - Variational dropout - Concatenation of outputs from each layer - Bucketing method for training efficiency - Search method for prediction - Pretrained GloVe character embedding - Fasttext embedding Tech stack: Python Why it matters: R-Net is a significant contribution to the field of machine reading comprehension, offering improved performance and efficiency over the original R-NET paper.

### Overview: This project is a side project of YuzuMarker, focusing on the development of a first-ever CJK (Chinese, Japanese, Korean) font recognition and style extraction model. The project aims to automate the generation of a dataset of scene text images with different fonts, which can be used for various applications such as text recognition and style extraction. ### Key Features: - **Font Recognition and Style Extraction**: The project includes a model and implementation for recognizing and extracting styles from CJK fonts. - **Dataset Generation**: The repository also contains data for automatically generating a dataset of scene text images with different fonts, using the CJK font pack by VCB-Studio and thousands of background images from pixiv.net. - **Text Corpus**: The dataset includes randomly generated Chinese, Japanese, and Korean text mixed with English text to simulate real-world data. ### Tech Stack: - **Python**: The primary programming language used for the project. ### Why It Matters: This project is significant because it provides a practical tool for developers and researchers to handle and analyze CJK fonts, which is crucial in various applications such as OCR (Optical Character Recognition) and style extraction in text analysis.

### Overview: This project provides a comprehensive implementation of Dense Object Nets, a method for learning dense visual descriptors that can be used for robotic manipulation tasks. The code is based on recent developments in self-supervised dense descriptor learning and is designed to be task-agnostic and applicable to both rigid and non-rigid objects. ### Key features: - Dense Object Nets: A method for learning dense visual descriptors for previously unseen and potentially deformable objects. - Multi-object descriptor learning: Novel contributions to enable multi-object descriptor learning. - Generalization across classes of objects: Demonstrates descriptors that generalize across classes of objects. - Distinct descriptors for each object instance: Shows descriptors that are distinct for each object instance. - Robotic manipulation tasks: Demonstrates grasping of specific points on an object across potentially deformed object configurations. - Class-specific grasps: Demonstrates using class general descriptors to transfer specific grasps across objects in a class. - PyTorch implementation: The code is written in Python and uses PyTorch for training and evaluation. ### Tech stack: Python ### Why it matters: This project is significant because it provides a robust and task-agnostic method for visual understanding and manipulation of objects, which can be applied to a wide variety of robotic manipulation tasks. The ability to learn and generalize across classes of objects and specific points on objects is a valuable contribution that can significantly enhance the performance and applicability of robotic manipulation systems.

### Overview: AnglE is a powerful library for training and inferring sentence embeddings, optimized for state-of-the-art BERT/LLM-based models. It achieves remarkable performance on various benchmarks, including STS and MTEB, with a single line of code. ### Key Features: - **Loss Functions**: AnglE offers ACL24, Contrastive, CoSENT, and Espresso loss functions. - **Backbones**: Supports BERT, LLM, and Bi-directional LLM models. - **Training**: Supports single-GPU and multi-GPU training. ### Tech Stack: Python ### Why it Matters: AnglE simplifies the process of training and inferring sentence embeddings, making it an essential tool for researchers and developers looking to enhance their models with improved sentence embeddings.

### Overview: Meta Agents Research Environments (ARE) is a comprehensive platform designed to evaluate AI agents in dynamic, realistic scenarios. Unlike static benchmarks, this research platform introduces evolving environments where agents must adapt their strategies as new information becomes available, mirroring real-world challenges. ### Key Features: - **Dynamic Environments**: Scenarios that evolve over time with new information and changing conditions. - **Multi-Step Reasoning**: Complex tasks requiring 10+ steps and several minutes to complete. - **Real-World Focus**: Grounded situations that mirror actual real-world challenges. - **Comprehensive Evaluation**: The Gaia2 benchmark with 800 scenarios across multiple domains. ### Tech Stack: Python ### Why It Matters: Meta Agents Research Environments (ARE) provides a comprehensive platform for evaluating AI agents in dynamic, realistic scenarios. Unlike static benchmarks, ARE introduces evolving environments where agents must adapt their strategies as new information becomes available, mirroring real-world challenges. This makes it an essential tool for researchers and developers looking to evaluate the capabilities of AI agents in complex, multi-step reasoning tasks.

Overview: This repository contains the implementation of Enformer, a deep learning model developed by DeepMind for predicting gene expression, specifically designed for use with Pytorch. Key features: - Enformer model for gene expression prediction - Fine-tuning capabilities for downstream tasks - Pretrained weights available on Hugging Face Tech stack: Pytorch Why it matters: Enformer is a powerful tool for gene expression prediction, offering a significant improvement in accuracy compared to traditional methods. Its implementation in Pytorch makes it accessible and easy to integrate into various research and development projects. Repository: lucidrains/enformer-pytorch GitHub stars: 575