Projects

### Overview: Kolmogorov–Arnold Transformer (KAT) is a PyTorch implementation of the paper "Kolmogorov–Arnold Transformer" presented at ICLR 2025. This model replaces the MLP layers in transformers with KAN layers, aiming to improve performance in large-scale training scenarios. ### Key features: - **GPU Implementation**: The KAT model is designed for GPU usage, leveraging CUDA for efficient computation. - **Base Function**: The base function replaces B-spline with CUDA-implemented Rational KANs, enhancing the model's efficiency and performance. - **Group KAN**: Shares weights among groups of edges for improved efficiency and reduced memory usage. - **Initialization**: Maintains activation magnitudes across layers to ensure consistent performance. ### Tech stack: The KAT model is built using PyTorch, with support for both PyTorch 1.x and 2.x versions. ### Why it matters: KAT addresses the performance limitations of vanilla ViT models by replacing MLP layers with KAN layers, resulting in significant improvements in large-scale training scenarios such as ImageNet. This model is particularly effective in replacing MLP layers in transformers, leading to better performance and efficiency.

### Overview: Follow me on LinkedIn: https://www.linkedin.com/in/ashishpatel2604/ In this repository, you'll find 365 days of computer vision learning, presented as LinkedIn posts. Each day, a new topic is covered, from EfficientDet to Residual Attention Network. ### Key Features: - **EfficientDet** - A lightweight object detection model. - **YOLO Series** - Object detection algorithms. - **Detr** - A detector that uses transformer-based models. - **Vision Transformer** - A model that leverages vision transformers. - **Dynamic RCNN** - A dynamic object detection model. - **DeiT: (Data-efficient image Transformer)** - A model that is data-efficient. - **Yolov5** - A popular real-time object detection model. - **DropBlock** - A technique for handling class imbalance. - **FCN** - A fully convolutional network. - **Unet** - A U-Net-based model for image segmentation. - **RetinaNet** - A model for object detection with a single anchor box. - **SegNet** - A model for semantic segmentation. - **CAM** - A technique for visual attention. - **R-FCN** - A model for object detection with region proposals. - **RepVGG** - A model that uses residual learning. - **Graph Convolution Network** - A model that uses graph convolution. - **DeconvNet** - A model that uses deconvolution. - **ENet** - A model

### Overview: `s3` is a framework designed to train search agents in retrieval-augmented generation (RAG) without requiring substantial data. It focuses solely on the search component, achieving strong performance with minimal data usage compared to previous methods. The framework is modular, efficient, and compatible with any black-box language model, making it a versatile tool for improving search capabilities in various applications. ### Key Features: - **Efficient Data Usage**: s3 trains search agents using a fraction of the data required by existing methods. - **Modular Architecture**: The framework is designed to be easily integrated with different language models, including black-box models. - **Search Component Focus**: By isolating the search component, s3 ensures that the generator remains unchanged, maintaining the flexibility of the original model. - **Performance Benchmark**: The performance of s3 is demonstrated through a performance overview, showcasing strong results across various QA tasks. ### Tech Stack: - Python - PyTorch - VLLM - Flash Attention - PySerini for efficient retrieval and evaluation ### Why It Matters: `s3` provides a solution for training search agents using minimal data, making it a valuable tool for applications that require efficient and effective search capabilities without the need for extensive training data.

Overview: Confucius4-TTS is a multilingual and cross-lingual zero-shot text-to-speech engine designed for high-quality speech synthesis across 14 languages. Key features: - 14 Languages Supported: Chinese, English, Japanese, Korean, German, French, Spanish, Indonesian, Italian, Thai, Portuguese, Russian, Malay, and Vietnamese (more languages coming soon) - Unconstrained Voice Cloning: No reference transcript required - Cross-Lingual Voice Transfer: Unaccented speech synthesis across 14 languages - Zero-Shot Voice Transfer: Clone voices without additional training - Seamless Emotion Transfer: Clone the feeling, not just the voice - Robust Generalization: Stable performance in real-world multilingual scenarios Why it matters: Confucius4-TTS simplifies the process of switching languages while maintaining the same voice, delivering fluent, natural, and expressive speech. Tech stack: Python GitHub stars: 842

Overview: This repository contains code and data related to the paper "Automated Hate Speech Detection and the Problem of Offensive Language," presented at ICWSM 2017. The project focuses on developing a classifier for detecting hate speech and offensive language. Key features: - Python 2.7 code for replicating analyses from the paper - Labeled data in pickle and CSV formats - Lexicon for more accurate classification of hate speech - Scripts and instructions for running the classifier on new data Tech stack: Python 2.7 Why it matters: This project provides a framework for researchers and developers interested in automated language analysis and hate speech detection. The data and code can be used to replicate the findings and contribute to the field of natural language processing and sentiment analysis.

### Overview: AutoCoder is a model designed for code generation, surpassing GPT-4 Turbo in accuracy and offering a feature to automatically install required packages and run code without user intervention. ### Key Features: - **Code Interpreter with Package Installation**: AutoCoder can automatically install required packages and run code without user intervention. - **Improved Accuracy**: Achieves a test accuracy of 90.9% on the HumanEval base dataset, surpassing GPT-4 Turbo (90.2%). - **Comparison with OpenSource Models**: Offers a new feature of automatically installing packages, expanding the scope of code interpreter's application. ### Tech Stack: Python ### Why it Matters: AutoCoder significantly enhances the usability and accuracy of code generation models, making it a valuable tool for developers and researchers in the field of artificial intelligence and machine learning.

Overview: AIVA is a general-purpose virtual assistant designed for developers, providing a bot-generalization experience across multiple platforms and languages. Key features: - Cross-platform deployment on Slack, Telegram, Facebook, and any hubot adapters. - Multi-language support for Node.js, Python, Ruby, and other languages using SocketIO. - Built-in AI tools (Tensorflow, SkFlow, Scikit, Pandas, Indico.ml, spaCy, Watson, Google APIs). - Hackable, extending Hubot with custom modules. - Powerful and easy-to-use, with detailed setup and feature documentation. Tech stack: JavaScript. Why it matters: AIVA simplifies the process of building a virtual assistant, allowing developers to focus on their primary goals rather than the complexities of bot development. It provides a robust, multi-language, cross-platform solution that leverages AI tools, making it ideal for developers looking to integrate AI capabilities into their applications quickly and efficiently.

Overview: Open Agent Connector is an open-source, self-hosted alternative to Composio, Nango, Merge.dev, and Arcade.dev for AI agent integrations. It allows developers to inspect, run, and extend their own agent integration layer. Key features: - MCP gateway - Scoped runtime tokens - OAuth - Action discovery - Direct MCP actions - REST API - Connections & credentials - Local web console - Outbound HTTP safeguards - Idempotent execution - Run records - Unified catalog Tech stack: Python Why it matters: Open Agent Connector provides a compact, self-hosted implementation for developers who want to inspect, run, and extend their own agent integration layer. It is an open-source option to evaluate alongside Composio, Nango, Merge.dev, and Arcade.dev, prioritizing a compact self-hosted implementation that can be modified. Repository: Anil-matcha/open-agent-connector GitHub stars: 845

### Overview: IncarnaMind is a free, open-source desktop app for macOS, Windows, and Linux that allows users to write with AI and ask questions about their own documents. It checks every sentence and figure in a draft against the original documents, ensuring every answer cites the exact page it comes from. Compatible with ChatGPT, Claude, or Ollama, IncarnaMind is a powerful tool for analysts, lawyers, and engineers. ### Key Features: - **Source Verification**: Every answer shows the page it came from, and clicking on the quote lights it up. - **Document Comparison**: Checks every sentence and figure in a draft against the original documents. - **AI Integration**: Works with ChatGPT, Claude, or Ollama, or can be integrated with your own API key. - **Desktop App**: Runs on macOS, Windows, and Linux, ensuring compatibility across different operating systems. ### Tech Stack: TypeScript ### Why It Matters: IncarnaMind is a valuable tool for anyone who needs to verify the accuracy and integrity of their AI-generated content. It ensures that every answer is backed by the original source, making it a reliable and trustworthy solution for professionals who need to stand behind their work.

### Overview: Codync is an open-source Grok Bot alternative that allows you to message and delegate coding agents such as Claude Code, Codex, Cursor, Pi, OpenCode, Grok Build, Gemini, Copilot, and many more. These bots can be integrated into your computer, iPhone, Mac, Windows, Linux, or terminal, making them accessible from anywhere. Codync is designed to be a convenient and efficient way to collaborate with your coding teammates. ### Key Features: - **Integrated Agents:** Codync integrates popular coding agents into your computer, allowing you to delegate tasks directly from your favorite apps. - **Persistent Bots:** These bots can be set up to run in the background, ensuring they are always ready to assist you. - **Cross-Platform Compatibility:** Codync works seamlessly across macOS, Windows, Linux, and terminals, making it accessible from anywhere. - **No Subscriptions:** Codync is free and open-source, with an Apache 2.0 license, ensuring it is available to all users without any limitations. ### Tech Stack: - **Rust:** The primary language used to build Codync, providing robust and efficient code execution. - **Swift:** Used for iOS development, ensuring a smooth and intuitive user experience on iPhones. - **Electron:** Used for desktop applications, allowing Codync to run on any device with a compatible terminal. ### Why It Matters: - **Convenience:** Codync makes it easy to delegate tasks to your coding agents, eliminating the need to switch between different applications. - **Efficiency:** By integrating agents

### Overview: The **Linear Attention Transformer** is a transformer model that introduces a linear complexity variant of attention, making it highly efficient for long-range language modeling. This model combines local attention with a scaled global attention mechanism, significantly reducing memory usage and computational complexity. ### Key Features: - **Linear Complexity Attention**: The model scales linearly with respect to the sequence length, making it highly efficient for long sequences. - **Local Attention**: Utilizes local attention heads to focus on local context, reducing the need for global attention. - **Global Attention**: Scales linearly with the sequence length, ensuring efficient handling of long-range dependencies. - **Efficient Memory Usage**: The model is designed to be memory-efficient, reducing the need for large memory buffers. - **Autoregressive or Causal**: The model can be used in both autoregressive and causal settings, depending on the application requirements. ### Tech Stack: - **Python**: The primary programming language used for implementing the model. ### Why It Matters: - **Efficiency**: The linear complexity attention mechanism makes the model highly efficient, especially for long sequences, reducing memory usage and computational costs. - **Memory Savings**: The model is designed to be memory-efficient, making it suitable for applications where memory is a constraint. - **Versatility**: The model can be used in both autoregressive and causal settings, providing flexibility in different applications. ### Repository: - **Repository**: lucidrains/linear-attention-transformer - **GitHub Stars**: 845 - **GitHub Description**: A fully featured Transformer that mixes

Overview: Veritas Kanban is a lightweight task management board designed for AI agents, offering a visual Kanban board with optional AI agent orchestration layers. Key features: - Local-first task management board - Optional AI agent orchestration layers (CLI, MCP, OpenClaw, Squad Chat webhooks, workflows, governance) Tech stack: TypeScript Why it matters: Veritas Kanban provides a simple and efficient way to manage tasks and workflows, with optional advanced features for AI-driven project management. Repository: BradGroux/veritas-kanban GitHub stars: 844 GitHub description: Lightweight orchestration harness built for your AI agents. The unfiltered truth about where your project stands.

### Overview: AlpacaFarm is a simulation framework designed to facilitate research and development in learning from human feedback, particularly focusing on methods that can be trained without the need for human data. This includes the simulation of pairwise preferences from language models, automated evaluation of instruction-following models, and reference implementations of baseline methods. ### Key Features: - **Simulating Pairwise Preference**: AlpacaFarm allows researchers to simulate pairwise preferences from advanced language models like GPT-4, enabling the development of instruction-following models without the need for human data. - **Automated Evaluation**: The framework provides automated evaluations for instruction-following models, facilitating the development of more efficient and effective methods. - **Reference Implementations**: It includes validated reference implementations of baseline methods such as PPO and best-of-n, allowing for comparison and modification of learning algorithms. ### Tech Stack: Python ### Why it Matters: AlpacaFarm simplifies the process of developing and researching instruction-following models, making it accessible and cost-effective for research without the need for extensive human data collection. This framework is particularly useful in the field of reinforcement learning from human feedback (RLHF) and related methodologies. Repository: tatsu-lab/alpaca_farm GitHub Stars: 844

### Overview: This repository is a PyTorch implementation of multi-task learning architectures, including MTI-Net (ECCV 2020). It aims to complement existing works on multi-task learning and provide a comprehensive set of models and training strategies. ### Key Features: - Multi-Task Learning (MTL) models - Efficient multi-scale task interaction networks (MTI-Net) - Support for various datasets and tasks - Pre-trained HRNet and ResNet models ### Tech Stack: PyTorch ### Why It Matters: This repository provides a practical implementation of multi-task learning techniques, making it a valuable resource for researchers and practitioners in the field of computer vision and machine learning.

Overview: This project is a scanner designed to read bar codes printed on the edges of cards, revealing the location of each card in a deck. The device uses an IR camera to detect invisible marks, ensuring the marks remain hidden from view. Key features: - IR camera for detecting invisible marks - Raspberry Pi Zero W with NoIR camera module - Abra iOS client for displaying scanned images - Testbed with visualization tools for understanding the system's operation Tech stack: Swift, Python, Raspberry Pi Zero W, NoIR camera module, Abra iOS client Why it matters: This project is crucial for developing a practical, invisible card magic trick that can be performed without the audience seeing the marks, ensuring the illusion's effectiveness.

### Overview: DeepResearch Bench is a comprehensive benchmark for deep research agents designed to evaluate and compare the performance of various AI models across different domains and tasks. ### Key Features: - **Model Evaluation**: Benchmarks models in multiple domains including RACE, FACT, and others. - **Scoring Metrics**: Utilizes a variety of metrics including Overall, PAR, OPC, FAP, and FAS to assess model performance. - **Community Submission**: Allows users to submit their models for evaluation and display their results on the leaderboard. - **Leaderboard Migration**: Plans to migrate the leaderboard to a new system by June 1, 2026, ensuring compatibility and comparability. ### Tech Stack: - Python ### Why It Matters: DeepResearch Bench provides a standardized and comprehensive framework for evaluating AI models, facilitating fair and accurate comparisons across different research areas.

Overview: AutoSchemaKG is a novel framework for automatic knowledge graph construction that combines schema generation via conceptualization, designed to address the challenges of constructing high-quality knowledge graphs from unstructured text. Key features: - Knowledge Graph Triple Extraction - Schema Induction - ATLAS Knowledge Graphs: 900+ million nodes, 5.9 billion edges, built without predefined schemas or manual intervention Tech stack: Python Why it matters: AutoSchemaKG achieves state-of-the-art performance on multiple benchmarks and demonstrates strong generalization capabilities across different domains, making it a valuable tool for knowledge graph construction and semantic bridging.

Overview: FastALPR is a high-performance Automatic License Plate Recognition (ALPR) system that offers fast and efficient ONNX models by default, with the option to switch to custom models if needed. Key features: - Uses advanced models for precise license plate detection and OCR. - Easily switch out detection and OCR models. - Quick setup with a simple API. - Ready-to-use detection and OCR models included. - Optimized with ONNX Runtime for fast performance. Tech stack: Python Why it matters: FastALPR is designed for developers looking for a reliable and efficient solution for license plate recognition, offering both pre-configured models and the flexibility to use custom models for enhanced performance and customization.

Overview: This is a React hook that converts speech from the microphone to text, making it accessible to React components. Key features: - Integrates with the Web Speech API for speech recognition. - Provides a hook that allows components to access the transcript of speech. - Supports React 16.8 and requires version 3.x of the library for TypeScript compatibility. Tech stack: JavaScript, Web Speech API. Why it matters: This library enables the creation of voice-enabled applications, enhancing user interaction and accessibility, particularly in modern browsers.

### Overview Golf is a **framework** designed to streamline the creation of MCP server applications. It allows developers to define server's capabilities—*tools*, *prompts*, and *resources*—as simple Python files within a conventional directory structure. Golf then automatically discovers, parses, and compiles these components into a runnable MCP server, minimizing boilerplate and accelerating development. ### Key Features - **Enterprise-Grade Authentication**: Includes JWT, OAuth Server, and development tokens for secure authentication. - **Built-in Utilities for LLM Interactions**: Simplifies interactions with large language models. - **Automatic Telemetry Integration**: Integrates telemetry to monitor and optimize server performance. ### Tech Stack - Python 3.10 or newer ### Why It Matters Golf is designed to make it easier and faster to build MCP server applications, focusing on the core logic of your agent while automating the complexities of authentication, telemetry, and server infrastructure. This results in a more efficient and secure development process, especially for projects that require real-world MCPs and AI agents.