Projects

### Overview: itr-wala is a terminal-based application designed to simplify the filing of Indian Income Tax Returns (ITRs). It provides a deterministic tax engine that computes tax math accurately without relying on artificial intelligence, ensuring compliance with both Income Tax Act 1961 and 2025-26. The application is available in Python and comes with a skills.sh plugin for easy integration into tax filing workflows. ### Key Features: - **Terminal-based Interface**: Runs directly from the terminal, eliminating the need for CA or portal. - **No CA or Fee**: No need for a Chartered Accountant, no 3-hour portal wait time. - **Comprehensive Tax Calculation**: Accurately computes tax math using tested code, not relying on language models. - **New Feature**: Support for advance tax filing for FY 2026-27, including 15 December 2026 and 15 March 2027. - **Statutory Compliance**: Ensures compliance with statutory dates and legal year labels. - **User-Friendly**: Intuitive interface and easy integration with skills.sh for tax filing workflows. ### Tech Stack: Python ### Why It Matters: - **Compliance**: Ensures compliance with Indian Income Tax Act without the need for CA or portal. - **Accuracy**: Accurate tax computation using tested code, eliminating the risk of errors. - **User-Friendly**: Easy to use and integrate with existing tax filing workflows. - **Comprehensive**: Supports both FY 2025-26 and FY

Overview: The CERLAB UAV Autonomy Framework is a comprehensive and modular framework designed for autonomous unmanned aerial vehicles (UAVs). It includes components such as autonomous flight, global waypoint planning, 3D mapping, dynamic obstacle detection, remote control, optimal trajectory time allocation, trajectory tracking, and a lightweight simulator. Key features: - Autonomous flight package integrating all other modules - Global waypoint planner library - 3D mapping library - Dynamic obstacle detection and tracking algorithm - Rviz configuration and launch files for easy visualization - Optimal trajectory time allocation library - Trajectory planning library - Lightweight Gazebo/ROS-based simulator Tech stack: C++ and ROS (ROS Melodic and ROS Noetic) Why it matters: This framework is essential for researchers and developers looking to create autonomous UAV systems, offering a versatile and modular approach to achieving autonomous navigation, exploration, and inspection tasks. GitHub stars: 879 Repository: Zhefan-Xu/CERLAB-UAV-Autonomy

Overview: This list focuses on tools and workflows where AI plays a central role in the development process, emphasizing the use of prompts, iterations, and exploration. Key features: - Web-Based Builders: Bolt.new, Lovable, Vercel’s v0, Capacity, CHAI.new, Replit, Create, Trickle, Tempo, Softgen, Lazy AI, HeyBoss, Creatr, Rork, Firebase Studio, Napkins, HeroUI Chat, Rocket.new, embedible.io, FlyPloy, MyVibe, Rosebud AI, Forge. - Editors and IDEs: Not specified. - Mobile Tools: Not specified. - Extensions & Plugins: Not specified. - Desktop & Local Apps: Not specified. - CLI Tools: Not specified. - AI-Driven Task Management: Not specified. - Monitoring & Cost Tracking: Not specified. - Project Documentation: Not specified. - Articles & Updates: Not specified. - Contributing: Not specified. Tech stack: HTML. Why it matters: This list aims to provide a comprehensive collection of tools and resources for developers working with AI, focusing on the AI-first development approach known as "vibe coding." Repository: awesome-vibe-coding/awesome-vibe-coding GitHub stars: 878 GitHub description: A hand-picked collection of tools and references for building software with the help of AI—through prompts, iterations, and exploration.

Overview: This repository offers a comprehensive learning experience for Python courses, focusing on modern AI applications. Key features: - Type Hints for better code readability and maintainability - Practical examples demonstrating AI integration with Python - Hands-on projects that simulate real-world AI tasks Tech stack: Jupyter Notebook Why it matters: This repository is essential for developers looking to enhance their Python skills with modern AI techniques, ensuring they can effectively integrate AI into their projects.

### Overview: This project is a significant challenge in autonomous driving focused on predicting 3D occupancy in complex scenes. The goal is to estimate the 3D space occupancy based on surround-view images, which is crucial for understanding the 3D surroundings in autonomous vehicles. ### Key features: - **Large-Scale Benchmark**: The project provides a large-scale occupancy benchmark based on the nuScenes dataset. - **Voxelized Representation**: The occupancy is represented in a voxelized 3D space, where the occupancy state and semantics are jointly estimated. - **Comprehensive Task**: The challenge involves predicting the 3D occupancy of the scene, which is challenging due to the complex geometrical shapes of objects and the absence of background information. ### Tech stack: - **Python**: The project is primarily developed and run using Python. ### Why it matters: This project is significant because it addresses a key challenge in autonomous driving by providing a comprehensive and large-scale benchmark for predicting 3D occupancy, which is essential for scene perception and autonomous driving.

### Overview: FridayOS 1.0 is an open-source project designed to serve as a local-first AI second brain for knowledge work. It utilizes the Six-Brain Method, providing a local Markdown blueprint, an Obsidian example vault, and bilingual guides for building an AI second brain. Inspired by the human brain's division of functions, it organizes sources, knowledge, tasks, and reusable methods into cooperating regions. ### Key Features: - **Local Knowledge Management Blueprint**: A Markdown-based local knowledge management blueprint. - **Obsidian Example Vault**: An example vault for building an AI second brain. - **Bilingual Guides**: Guides available in English and Chinese. ### Tech Stack: Python ### Why It Matters: FridayOS 1.0 is designed to support effective knowledge work by organizing sources, knowledge, tasks, and reusable methods into cooperating regions inspired by the human brain's division of functions. This project aims to help users manage their knowledge effectively, ensuring that they can access, reuse, and improve their knowledge in a structured manner. ### Repository: Repository: Robinson-AI-Lab/FridayOS1.0 Primary Language: Python GitHub Stars: 878 GitHub Description: Local-first AI second brain & personal knowledge management: Six-Brain Markdown blueprint, Obsidian example vault, offline deployment and bilingual guides.

### Overview: Gesture Controlled Virtual Mouse simplifies computer interaction by allowing users to control their computer using hand gestures and voice commands. This project leverages the MediaPipe and OpenCV libraries to recognize gestures and voice commands, making it accessible without the need for additional hardware. ### Key Features: - **Neutral Gesture**: Halts/Stop execution of current gesture. - **Move Cursor**: Assigns cursor to midpoint of index and middle fingertips, moving the cursor to desired location. - **Left Click**: Single left click gesture. - **Right Click**: Single right click gesture. - **Double Click**: Double click gesture. - **Scrolling**: Horizontal and vertical scroll gestures with proportional speed. - **Drag and Drop**: Drag and drop functionality for file transfers. - **Multiple Item Selection**: Multiple item selection for easy file management. ### Tech Stack: The project uses Python 3.8.5, MediaPipe, and OpenCV libraries to recognize hand gestures and voice commands. ### Why It Matters: This project enhances user interaction by making computer control intuitive and accessible, reducing the need for direct physical contact and enhancing the overall user experience.

### Overview: RETRO-pytorch is an implementation of Deepmind's Retrieval-based Attention Network (RETRO) in PyTorch. This project aims to leverage rotary embeddings for relative positional encoding and Faiss library for efficient nearest neighbor search, aiming to reach GPT-3 performance with significantly fewer parameters. ### Key Features: - Rotary embeddings for relative positional encoding - Faiss library for efficient nearest neighbor search - Causal chunked cross attention - 1000-layer retriever transformer - Training wrapper for processing text documents into necessary memmapped arrays ### Tech Stack: PyTorch ### Why it Matters: - Reduces model parameters by 10x compared to state-of-the-art models - Demonstrates the potential of retrieval-based transformers for large-scale language models - Provides a practical implementation for further research and scaling ### GitHub Stars: 878 ### Repository: lucidrains/RETRO-pytorch ### Primary Language: Python ### Installation: ```bash $ pip install retro-pytorch ``` ### Usage: ```python import torch from retro_pytorch import RETRO # Instantiate and configure RETRO retro = RETRO( chunk_size = 64, # chunk size for indexing and retrieval max_seq_len = 2048, # max sequence length enc_dim = 896, # encoder model dimension enc_depth = 3, # encoder depth dec_dim =

### Overview TerraTorch is a PyTorch toolkit designed for fine-tuning Geospatial Foundation Models (GFMs). It leverages the power of PyTorch Lightning and the TorchGeo library to provide a flexible and efficient framework for training and deploying models on geospatial data. ### Key Features - **Model Training Tools**: TerraTorch offers a set of convenient tools for fine-tuning Image Segmentation, Classification, and Pixel Wise Regression tasks. - **Model Factories**: Easy-to-use model factories that allow combining different backbones and decoders for various applications. - **Data Modules**: Ready-to-use datasets and datamodules that require minimal configuration, making it straightforward to work with geospatial data. - **CLI and Jupyter Notebooks**: Launch fine-tuning tasks through command-line interface or Jupyter notebooks for easy configuration. - **Pre-trained Models**: Access to open-source pre-trained Geospatial Foundation Models such as Prithvi, TerraMind, SatMAE, and ScaleMAE. ### Tech Stack TerraTorch is built on top of PyTorch Lightning and TorchGeo, utilizing the latest advancements in PyTorch and geospatial data processing. ### Why It Matters TerraTorch simplifies the process of fine-tuning and deploying Geospatial Foundation Models, making it accessible to developers and researchers alike. Its modular architecture and comprehensive set of tools make it an ideal choice for anyone working with geospatial data and AI. ### GitHub Repository The repository for TerraTorch is hosted on GitHub at [

Overview: Skywork-UniPic is an open-source repository for the unified multimodal model series, featuring three distinct modeling paradigms: UniPic-3, UniPic-2, and UniPic-1. These models are designed for single-image editing, multi-image composition, and text-to-image generation, with a focus on performance and efficiency. Key features: - UniPic-3: Open-source SOTA multi-image editing model, supports 1-6 input images, 8-step inference with 12.5x speedup via CM + DMD distillation. - UniPic-2: SD3.5M-Kontext and MetaQuery variants based on Efficient Architectures with Diffusion Post-Training, delivering state-of-the-art performance in text-to-image generation, fine-grained image editing, and multimodal reasoning. - UniPic-1: Unified Autoregressive Modeling for joint visual understanding and generation, enabling a single transformer to handle both perception and synthesis tasks. Tech stack: Python Why it matters: Skywork-UniPic provides a unified framework for image editing, generation, and understanding, offering high-fidelity synthesis, seamless image editing, and robust visual perception capabilities, making it a valuable tool for developers and researchers in the field of image processing and AI.

Overview: Eleven is an agent-first reverse-engineering app designed for macOS, Windows, and Linux. It allows any Agent Client Protocol (ACP) agent to read, rename, and patch binaries, Android apps, and GPU kernels while you watch and approve in the same window. Key features: - Bring your own agent - Agents drive the window - Ask about any code - You stay in control - Mach-O, ELF, and PE - Android apps - Source files - Rewrite in any language - GPU kernels - Shaders, models, and GPU profiles - Projects - Patching - Diffing Tech stack: TypeScript Why it matters: Eleven is a powerful tool for reverse-engineering and debugging applications, providing a unified interface for interacting with various types of code and binaries across different platforms.

Overview: This repository contains a PyTorch implementation of Sparsely-Gated Mixture of Experts (Mixture of Experts), which is a technique for increasing the parameter count of language models without significantly increasing computational resources. Key features: - Sparsely Gated Mixture of Experts - Pytorch Implementation - Supports both single-level and hierarchical mixture of experts - Includes examples for single machine and hierarchical models - Provides a TensorFlow implementation reference Tech stack: Pytorch Why it matters: This implementation allows for the efficient training of large language models by increasing the parameter count without a proportional increase in computational requirements, making it suitable for applications where memory and computational resources are constrained. GitHub stars: 877 Repository: lucidrains/mixture-of-experts

Overview: This project provides Python bindings for libpostal, a fast statistical parser/normalizer for international street addresses. Key features: - Python bindings for libpostal - Fast address parsing and normalization - Supports Python 3.5+ and CPython - Compatible with libpostal on Linux, macOS, and Windows Tech stack: C Why it matters: This project simplifies the process of parsing and normalizing international street addresses, making it easier for developers to integrate address handling into their applications.

### Overview: JarvisArt is a project that aims to liberate human artistic creativity by utilizing an intelligent photo retouching agent. This project was accepted by CVPR 2025 and is part of the NeurIPS 2025 conference. The agent, JarvisArt, is designed to enhance and transform photos in a way that respects human artistic expression, making it easier for anyone to create stunning images. ### Key Features: - **Intelligent Photo Retouching**: JarvisArt uses an intelligent algorithm to enhance and retouch photos, making them look more artistic without compromising the original quality. - **Collaborative Learning**: The agent learns from both human evaluators and other AI models to improve its performance continuously. - **Versatile Applications**: It can be used for various applications such as enhancing portraits, fixing imperfections, and creating unique art pieces. ### Tech Stack: - **Programming Languages**: Python - **Frameworks**: TensorFlow, PyTorch - **Databases**: None specified ### Why It Matters: JarvisArt is significant because it combines the power of AI with human creativity, making it easier for artists to enhance their work without the need for advanced technical skills. This project not only showcases the potential of AI in the arts but also opens up new possibilities for democratizing creativity. ### Repository: - **Repository**: LYL1015/JarvisArt - **Stars**: 876 - **Description**: NeurIPS' 2025 project that aims to liberate human artistic creativity via an intelligent photo ret

### Overview: Flatkey CLI is a powerful tool for generating images, videos, audio, and text through a single Flatkey credit balance. It offers one CLI for all media types and allows for programmatic access without the need for expensive premium models or long wait times. ### Key Features: - **CLI for All Media Types**: Generate images, videos, audio, and text using a single Flatkey credit balance. - **One Credit Balance**: Use one `FLATKEY_API_KEY` for all media types, making it agent-friendly and efficient. - **JSON Mode**: Output in clean JSON format for agent-friendly scripts. - **Local Outputs**: Generate outputs locally for easy integration. - **Cross-Platform**: Available as a npm package for macOS, Linux, and Windows. - **Progress Animation**: Visual progress indicator for human runs, disabled in `--json` mode. ### Tech Stack: JavaScript ### Why It Matters: - **Efficient Cost**: Flatkey offers a token-based pricing model that is significantly cheaper compared to retail routes, especially for image generation. - **Programmatic Access**: Designed for agents, scripts, and batch testing, making it ideal for media teams building repeatable generation pipelines. - **Fast Retries and Local Outputs**: Built for fast CLI retries and local outputs, making it suitable for creative workflows. - **One Credit Balance**: One `FLATKEY_API_KEY` for all media types, reducing the need for expensive premium models. - **JSON Mode**: Clean JSON output for easy integration with scripts and agents. - **Cross-Platform**:

### Overview: The Oxylabs Agent Skills repository provides a collection of structured instructions for AI agents to interact with Oxylabs products. These skills include residential, mobile, datacenter, and dedicated datacenter proxies, a web blocker, a web scraper API, a remote agent browser, and video data extraction and downloads. ### Key Features: - Residential, Mobile, Datacenter, and Dedicated Datacenter proxies with geo-targeting, IP rotation, and session persistence. - Web blocker that handles fingerprinting, JavaScript rendering, and CAPTCHAs. - Production-grade web scraping with structured JSON parsing for 40+ targets and geo-targeting. - Remote agent browsers via CDP (Playwright/Puppeteer) with built-in request handling and residential proxies. - Video data extraction, search, and high-bandwidth proxy video downloads. ### Tech Stack: JavaScript ### Why it Matters: These skills are crucial for automating interactions with Oxylabs products, enhancing security, and improving the reliability of web scraping and video data extraction tasks. They streamline the process of accessing and processing data from various Oxylabs services efficiently and securely. ### GitHub Stars: 876 The repository has been starred by numerous developers, indicating its usefulness and popularity in the community.

### Overview: Fine-tuning Large Language Models with PyTorch and Hugging Face is an intermediate-level book that focuses on key concepts for understanding and fine-tuning large language models. The book covers quantization, low-rank adapters, and formatting templates, which are essential for mastering the current landscape and future developments in the field. ### Key Features: - **Quantization**: Techniques for reducing the size of large language models. - **Low-Rank Adapters (LoRA)**: Simplified versions of large models for faster inference. - **Formatting Templates**: Pre-built datasets and templates for efficient training. ### Tech Stack: - **PyTorch**: The primary deep learning framework used for fine-tuning models. - **Hugging Face**: The ecosystem for working with large language models, providing comprehensive tools and configurations. ### Why It Matters: - **Stability and Future-Proofing**: The concepts discussed in the book are likely to remain central to the fine-tuning process for the foreseeable future. - **Practical Applications**: These concepts can be applied to various large models, not just language models. - **Toolset for Data Scientists**: Essential tools for any data scientist's toolkit, making it easier to train or fine-tune a variety of large models. ### Repository: - **GitHub Repository**: dvgodoy/FineTuningLLMs - **Primary Language**: Jupyter Notebook - **GitHub Stars**: 874 - **GitHub Description**: Official repository of my book "A Hands-On Guide to Fine-Tuning LLMs with PyTorch and

### Overview: VectorCode is a code repository indexing tool designed to enhance the performance of your LLMs by providing contextual information about the projects you're working on. This project is currently in beta and is actively being improved. ### Key Features: - **Code Repository Indexing**: Automatically index and provide information about your code repositories. - **Neovim Plugin**: Offers APIs and integrations for building or enhancing AI plugins. - **Neovim Documentation**: Detailed documentation for both command-line tool and Neovim plugin usage. ### Tech Stack: Python ### Why It Matters: This tool significantly improves the quality of LLM outputs by providing relevant context from the project, thereby reducing hallucination and enhancing the overall performance of the model. ### GitHub Stars: 874 ### Star History: The project has seen a steady increase in stars, reflecting growing interest and adoption. ### Why VectorCode? LLMs often struggle with understanding close-source projects or cutting-edge developments. VectorCode helps by injecting task-relevant context from the project into the prompt, significantly improving the quality of the model's output and reducing hallucination. ### Documentation - **CLI Documentation**: For command-line tool usage. - **Neovim Plugin Documentation**: Detailed instructions for Neovim users. - **Wiki**: Extra tricks and tips for maximum usage. - **Discussions**: General questions and cool usages. - **Pull Requests**: Contributions and feature development. ### About Versioning - **0.x.x**: Early stage, features may change. - **1.0.x**: Major changes, features will

### Overview: The makerskills plugin offers a suite of AI agent skills designed to streamline personal and professional tasks. It integrates with Claude Code, Codex, and other Agent Skills hosts, providing decision-making, research, and meta-skills to enhance productivity and creativity. ### Key features: - **Decision Making**: Facilitates quick and structured decision-making with the 37signals framework. - **Research**: Helps in gathering and organizing information through deep research and citation tracking. - **Second-Brain**: Aims to improve personal knowledge by capturing and summarizing thoughts. - **Content Rotation**: Automates the process of rotating content for better engagement. - **Scenario Modeling**: Assists in creating realistic scenarios for problem-solving. - **CFO Cadence**: Supports financial planning and analysis. - **Domain Hunts**: Helps in finding available domain names for new projects. - **Meta-Skills**: Enhances authoring capabilities for writing more effectively. ### Tech stack: Python ### Why it matters: The makerskills plugin simplifies complex tasks by automating repetitive processes, making it easier for founders and indie operators to focus on higher-value activities. Its integration with Claude Code and other AI hosts ensures that users can leverage cutting-edge AI to enhance their personal and professional skills. Repository: coreyhaines31/makerskills Primary language: Python GitHub stars: 872 GitHub description: AI agent skills for the personal operator's craft — decisions, research, second-brain, content rotation, scenario modeling, CFO cadence, domain hunts, and meta-skills to

Overview: Epsilla is an open-source vector database designed to provide high performance, cost-effectiveness, and scalability for large language models. Leveraging parallel graph computing, it bridges the gap between information retrieval and memory retention in LLMs. Key features: - High performance and production-scale similarity search for embedding vectors. - Full fledged database management system with familiar database, table, and field concepts. - Metadata filtering. - Hybrid search with a fusion of dense and sparse vectors. - Built-in embedding support, with natural language in natural language out search experience. - Cloud native architecture with compute storage separation, serverless, and multi-tenancy. - Rich ecosystem integrations including LangChain and LlamaIndex. - Python/JavaScript/Ruby clients, and REST API interface. Tech stack: Epsilla's core is written in C++ and leverages advanced academic parallel graph traversal techniques for vector indexing, achieving 10 times faster vector search than HNSW while maintaining precision levels of over 99.9%. Why it matters: Epsilla is crucial for ensuring scalability, high performance, and cost-effectiveness of vector search in large language models, making it an essential tool for developers and researchers in the field of AI and NLP. GitHub stars: 873 Repository: epsilla-cloud/vectordb