Projects

Overview: WeldFusionNet is a multimodal deep learning model designed for automated weld defect classification using sensor, audio, and video data. It achieves an F1 score of 0.9567 on a test set, demonstrating its effectiveness in industrial applications. Key features: - Multimodal deep learning approach - Uses sensor, audio, and video data - Achieves a high F1 score of 0.9677 for binary classification - Type Macro F1 score of 0.9401 for multi-class classification Tech stack: Python, PyTorch, Streamlit Why it matters: WeldFusionNet provides a robust solution for automated weld defect classification, significantly improving the efficiency and accuracy of quality inspection processes in industrial settings.

### Overview: This project involves the development of a Korean Speech-to-Text (Korean STT) model using the Open Source tool kit Kospeech. The project utilizes data from foreign speakers to create a Korean speech recognition model and proposes a business idea for a hackathon. The model was trained on 300,000 training data and tested on 10,000 test data with a CER score of 0.085 and WER score of 0.1919. The project received 29th place in the performance category and 2nd place in the final evaluation. ### Key Features: - **Model Development**: A Korean STT model was developed using Kospeech. - **Data Utilization**: The project utilized data from foreign speakers for training. - **Performance Metrics**: The model achieved a CER score of 0.085 and WER score of 0.1919. - **Hackathon Participation**: The project was a part of a hackathon where the team proposed a business idea. - **Error Analysis**: The project included an analysis of errors and bugs in the original open-source code, with the results being detailed in the blog post at [this link](https://mingchin.tistory.com/152). ### Tech Stack: - Python 3.8 - Kospeech ### Why It Matters: The project showcases the effectiveness of using open-source tools for speech recognition, demonstrating how data from foreign speakers can be utilized to create a high-performing Korean speech recognition model

### Overview: Seedream 5.0 Pro API 한국어 가이드: 모델 ID, 장당 $0.0146 (1K-layer) 요금, 파라미터, cURL·Python 예제와 OpenAI 호환 엔드포인트 설명. ### Key features: - **모델 ID 및 요금**: 1K-layer 기준 장당 $0.0146 (1달러부터 시작). - **파라미터**: 1K 레이어 기준으로 최소 1.46센트. - **OpenAI 호환 엔드포인트**: cURL·Python 예제와 설명. - **API 키 발급**: Seedream 5.0 Pro API를 호출하는 데 필요한 API 키 발급 방법. ### Tech stack: API 호출 및 파라미터 설정. ### Why it matters: - **고정 가격**: 장당 $0.0146 (1K-layer)의 종량제로 구매가 가능. - **소비 확인**: 모든 호출에 대한 비용이 응답에 포함되어 있습니다. - **비동기 작업**: 제출 후 `task_id`를 받고 `GET /v1/tasks/{id}`를 폴링하여 배치 및 재시도를 처리합니다. - **소모 예측**: 1K 레이어 기준으로 1.46센트로 계산됩니다. ### Repository

### Overview: SmartTextileGlove is a pioneering project that leverages machine learning to capture complex hand movements and interactions with objects using stretchable smart textile gloves. This innovative approach has the potential to revolutionize the way we interact with technology, making it more intuitive and responsive. ### Key Features: - **Machine Learning for Hand Movements:** The glove uses advanced machine learning algorithms to recognize and interpret complex hand movements. - **Stretchable Smart Textiles:** The glove is made from stretchable smart textiles that can conform to the user's hand, providing a natural and comfortable fit. - **Data Acquisition and Analysis:** The project includes a data acquisition software written in C# and a Python code for analyzing the collected data. ### Tech Stack: - **C# for iOS Software:** Developed using Microsoft Visual Studio C# 2017. - **Swift for iOS Software:** Used for the graphical user interface. - **C for Firmware Codes:** Developed using C to ensure compatibility with various devices. - **Unity for Unity Demo Software:** Utilized for interactive demonstrations. - **Python for Data Analysis:** Used for processing and analyzing the collected data. ### Why It Matters: SmartTextileGlove represents a significant advancement in the field of wearable technology, offering a more intuitive and responsive user experience. By combining machine learning with stretchable smart textiles, it opens up new possibilities for technology integration into everyday wearables, potentially enhancing the quality of life for users.

Overview: KaiROS AI is a powerful local AI assistant designed for Windows and Android, offering a suite of features including local LLM inference, model catalog, chat interface, and more. Key features: - Local LLM Inference - Model Catalog (40+ models) - Chat Interface - Chat History & Sessions - System Prompt Editing - Custom Model Import - Markdown Rendering - Vision Models (Multimodal) - RAG (Document Chat) - RAG-as-a-Service (RaaS) - Web Search Integration - Dynamic Context Sizing - Pre-flight RAM Check - Export (Markdown/JSON/Text) - Dark/Light Theme Tech stack: C# and .NET 9 Why it matters: KaiROS AI provides a privacy-first solution for local AI assistant capabilities, allowing users to run large language models directly on their devices without the need for cloud services or external dependencies. Repository: avikeid2007/KaiROS-AI GitHub stars: 27

Overview: Lora Merge is a C# project designed for merging LoRaWAN messages, enabling seamless integration and data consolidation in IoT applications. Key features: - Efficient message merging - Support for multiple LoRaWAN protocols - Automatic error detection and correction - Scalable architecture for large datasets Tech stack: C# Why it matters: Lora Merge simplifies the process of handling and managing large volumes of LoRaWAN messages, ensuring reliable and consistent data flow in IoT environments.

### Overview: Pithos is a dependency-free C inference engine designed for Liquid AI's LFM2, LFM2.5, and LFM3 models. It provides a terminal chat, OpenAI-compatible server, web agent, and terminal agent all in a single executable. The engine runs on an ordinary laptop and can handle large models efficiently. ### Key Features: - **Terminal Chat**: A web-based interface for real-time conversations. - **OpenAI-Compatible Server**: Supports prompt reuse and HTTP server with prompt reuse across turns. - **Web Agent**: Offers a variety of functionalities including web search, file management, calculator, and MCP servers. - **Terminal Agent**: Executes tasks with shell commands, background processes, files, and the web, in both safe and autonomous modes. ### Tech Stack: - **C11**: The engine is written in C11 with no external dependencies. - **Windows, Linux, macOS**: The engine runs on these operating systems. - **Xbox One in Developer Mode**: The server runs on an Xbox One in Developer Mode, with a status and chat screen on the TV. - **GPU Acceleration**: Optional acceleration using OpenCL, Direct3D 12, and Vulkan for GPU performance. ### Why It Matters: - **Performance**: Runs 13.5 GB models on 16 GB of RAM, achieving high speeds. - **Compatibility**: Compatible with OpenAI, allowing seamless integration with existing infrastructure. - **Versatility**: Supports multiple platforms and devices, including Xbox One. ### Tech Stack: - **C1

Overview: This free, open-source guide is designed for AI engineers and Forward Deployed Engineers (FDEs) to prepare for interviews. It covers LLM, RAG, AI agents, evals, and system design questions, providing detailed answers and insights into how companies run their interview loops. Key features: - Comprehensive coverage of LLM, RAG, AI agents, evals, and system design questions. - Detailed answers and expanded explanations for each question. - Insights into the FDE market and salary expectations. - Interactive practice through a live coding environment. Tech stack: The guide is written in Markdown and uses GitHub for version control. Why it matters: This handbook is essential for anyone preparing for interviews in the AI and FDE fields, providing a clear understanding of the types of questions they will face and how to prepare effectively. Repository: bicced/ai-engineer-interview-handbook GitHub stars: 27 GitHub description: Free, open-source AI engineer & Forward Deployed Engineer (FDE) interview handbook — LLM, RAG, AI agents & MCP, evals, AI system design, and company-by-company interview loops (2026).

Overview: BridgeRust is a monorepo that provides two AI-infrastructure products— Embex and BridgeTime— and the BridgeRust framework that powers them. The framework is designed to be a unified system for building cross-language Rust libraries that can be deployed to Python and Node.js. Key features: - Embex: A universal vector database client that abstracts fragmentation into a single, production-ready API, allowing seamless switching between different vector databases like LanceDB, Qdrant, and Pinecone. - BridgeTime: A Rust-powered datetime toolkit that provides a robust and efficient datetime handling system. - BridgeRust Framework: A unified system for building cross-language Rust libraries that can be deployed to Python and Node.js. Tech stack: BridgeRust is built using Rust, with bindings for Python, Node.js, and WebAssembly. Why it matters: BridgeRust simplifies the process of building AI infrastructure libraries that can be easily integrated into Python and Node.js applications, making it a valuable tool for developers looking to enhance their projects with AI capabilities.

Overview: Speech Kit is a comprehensive speech and language toolkit designed specifically for Obsidian, a popular note-taking application. It simplifies the process of dictation, transcription, translation, and listening directly within the editor, eliminating the need for multiple plugins and separate tools. Key features: - Dictate live text and audio files. - Transcribe audio and video files. - Import YouTube captions into notes. - Translate text locally. - Listen to notes. - Refine with AI presets. Tech stack: TypeScript Why it matters: Speech Kit consolidates fragmented tools into a single, unified workflow within Obsidian, making it easier to manage and integrate with existing note-taking practices.

Overview: 本项目旨在分享高质量的AI项目，帮助开发者了解和实践相关技术。 Key features: - 收集和分享高质量的AI项目 - 每周在公众号汇总和分享项目 - 实践和解读项目内容 - 欢迎感兴趣的同学加入 Tech stack: Python Why it matters: 通过分享高质量的AI项目，帮助开发者快速入门和实践相关技术，促进知识的共享和应用。 Repository: bupt-ai-club/awesomeProject GitHub stars: 27

Overview: Beyond Time Shifts is a reference-free audio-visual synchronization evaluator built on Qwen2.5-Omni-3B, utilizing the SynthSync dataset and R-GRPO training method. It is designed to evaluate generative audio-visual models and is compatible with the SyncBench benchmark. Key features: - SynthSync dataset of authentic generative sync failures - Continuous Omni-LLM evaluator with a `[SCORE]`-token projection head - R-GRPO training method for internalizing global causal structure Tech stack: Python Why it matters: Beyond Time Shifts provides a practical solution for evaluating generative audio-visual models by addressing the limitations of traditional metrics, offering a reference-free and absolute metric that humans can easily judge. Repository: chenhaoqcdyq/BeyondTimeShifts Primary language: Python GitHub stars: 27 GitHub description: Beyond Time Shifts (ECCV 2026): a reference-free audio-visual synchronization evaluator built on Qwen2.5-Omni-3B, with SynthSync data, R-GRPO training, and the SyncBench benchmark.

### Overview: Muse Glimmer is an open-weight 30B agentic multimodal model designed for always-on local agent workflows. It runs on consumer hardware after 4-bit quantization, with tool calling, failure recovery, controllable reasoning, and a 2B perception encoder for images and video frames. ### Key features: - **Open-Weight Model**: 30-billion-parameter agentic multimodal model. - **Always-On Local Agent Workflows**: Runs on consumer hardware after 4-bit quantization. - **Tool Calling**: Supports function calling and reasoning. - **Failure Recovery**: Handles failures and recovers from them. - **2B Perception Encoder**: For images and video frames. ### Tech stack: JavaScript ### Why it matters: An open agentic model that runs on your device — Apache 2.0 weights, dense 30B (2B perception + 28B decoder), ~4-bit under ~20 GB, trained for tool loops, failure recovery, and controllable reasoning. ### Repository: [cobusgreyling/Muse-Glimmer](https://github.com/cobusgreyling/Muse-Glimmer) GitHub stars: 27 GitHub description: Introducing Muse Glimmer: open-weight 30B agentic multimodal model that runs on your device (Meta). Interactive local agent lab + guide. Apache 2.0 · on-device AI · function calling ### Tech Stack: JavaScript

### Overview: Bitune is an implementation of Bidirectional Instruction-Tuning, a method designed to enhance the performance of language models by incorporating bidirectional instruction tuning. This project aims to explore and experiment with different ideas related to the paper, providing a clean and reproducible implementation. ### Key Features: - **Bidirectional Instruction-Tuning**: A method for enhancing language model performance. - **Modified HuggingFace Model Classes**: The project modifies HuggingFace model classes to include KV-cache, trainable mixing modules, and a custom _Bitune wrapper_. - **Mixing Modules**: The project includes a mixing module that combines two sets of features from the model, allowing for bidirectional tuning. - **Ablations**: The project provides ablation scripts to explore different aspects of the method. - **Scripts**: Contains scripts for instruction tuning setup, downstream task training, and ablation experiments. ### Tech Stack: - **Python**: The primary language used for the project. - **HuggingFace Models**: Modified and utilized for the project's requirements. - **Wandb**: Used for logging and tracking experiments. ### Why It Matters: This project is a practical implementation of Bidirectional Instruction-Tuning, providing a clean and reproducible way to explore and experiment with the method. The modifications and ablations offer insights into the effectiveness of different approaches in the context of language model training and performance enhancement. ### Repository: The repository is named `dkopi/Bitune` and has been starred by 27 contributors, indicating its popularity and usefulness in the community. ### Primary Language: The primary language

### Overview: The self-bot project is a lightweight WhatsApp bot built using TypeScript. It responds only to self-messages and utilizes the Baileys library for its functionality. The bot is designed to be simple and efficient, focusing on text-to-sticker conversions and other basic features. ### Key Features: - **Text to Sticker Conversion**: Converts text messages into stickers. - **Search Pinterest Images**: Searches for images from Pinterest. - **Sticker Creator**: Utilizes the WA-Sticker-Formatter library to create stickers. - **Note Database**: Stores notes in a SQLite database using Sequelize. - **Image Captioning**: Adds captions to images and converts them into stickers. ### Tech Stack: The project is built using TypeScript, with Baileys as the primary library for interacting with WhatsApp. It also incorporates SQLite for database storage and Sequelize for ORM functionality. ### Why It Matters: This bot serves as a simple yet effective tool for automating tasks related to text-to-sticker conversions and image processing. It's designed to be lightweight and easy to use, making it a useful addition to any developer's toolkit for automating repetitive tasks or creating basic bots for personal use.

### Overview: AirControl is a gesture-based and voice-controlled system designed for Windows, allowing users to control presentations, mouse cursor, and screen annotations without the need for a keyboard or mouse. It leverages MediaPipe for gesture recognition and voice recognition for commands. ### Key Features: - **Presentation Mode**: Wave to change slides, start/stop playback. - **Mouse Mode**: Air mouse, pinch to click, scissor to scroll. - **Drawing Mode**: Finger writing, fist to clear, shape correction. - **Voice Assistant**: Offline keyword spotting, offline dictation, mode-aware command sets, and spoken confirmations. ### Tech Stack: - Python - MediaPipe - SenseVoice-Small - Sherpa-ONNX - LLM (optional) ### Why It Matters: - **Efficiency**: Enables hands-free control, enhancing productivity and reducing reliance on traditional input methods. - **Accessibility**: Supports gesture and voice commands, making the system accessible to a wide range of users. - **Versatility**: Integrates with various applications and presentations, making it a versatile tool for remote control and presentation management. ### Repository: - **GitHub Repository**: enixz/AirControl - **Stars**: 26 - **Primary Language**: Python - **GitHub Description**: 🎯 AirControl - Gesture + Voice Control System for Windows - **GitHub Release**: v1.4.0 - **GitHub License**: Apache 2.0 - **GitHub Stars**: 26 ### Tech Stack: - Python ### Why It Matters:

Overview: This project is a private AI image generation website powered by the Gemini API, built with Flask. It includes a user system, point system, chat interface, and management backend. The website is open-source and available for use. Key features: - Text-to-Image and Image-to-Image capabilities - Support for multiple turns in chat - User management, point system, and backend configuration Tech stack: Flask Why it matters: This project leverages the Gemini API for AI image generation, providing a private, open-source solution for image creation and management. It includes robust features for user management, point system, and backend configuration, making it suitable for both personal and professional use. Repository: gbmomo/gemini-image-webapp GitHub stars: 27 GitHub description: Private AI Image Generation Website Source Code powered by Gemini API | Ready-to-use | Text-to-Image/Image-to-Image/Multi-turn Chat | User System + Credits Billing + Redemption Codes | Built with Flask.

Overview: This repository contains scripts for training a Kaldi model for German speech recognition (ASR). Key features: - Scripts for preparing data, language model, and lexicon - Training using Docker for easy execution - Results reported in word error rates for Megs v2 Tech stack: Shell Why it matters: This tool provides a comprehensive set of scripts for training a Kaldi model for German ASR, making it a valuable resource for researchers and developers in the field of speech recognition.

Overview: This MATLAB implementation of CHiME4 baseline Beamformit is designed for speaker diarization tasks in meetings, offering high accuracy and practical performance. Key features: - Beamformit.m: MATLAB script for real-time beamforming. - beamformit_step_by_step.mlx: MATLAB script for step-by-step implementation details. - Supports MATLAB and OCTAVE with signal package. Tech stack: MATLAB. Why it matters: This implementation provides a practical and efficient solution for beamforming tasks in acoustic scene analysis, particularly useful for speaker diarization in noisy environments. Repository: gogyzzz/beamformit_matlab Primary language: HTML GitHub stars: 27 GitHub description: A MATLAB implementation of CHiME4 baseline Beamformit

### Overview: ToK is a fully local, high-quality, and extensible Retrieval-Augmented Generation (RAG) application that leverages AI models and indices to query documents and generate contextually accurate responses. It is designed for users to upload documents and receive more informed answers tailored to their queries. Additionally, the app automatically stores chats, allowing users to revisit and continue conversations at any time. ### Key Features: - **Fully Local, Secure, and Privacy-Focused:** Ensures all operations are performed locally, keeping data private and secure. - **Versatile Document Uploading:** Supports uploading of various types of text documents and folders for creating an index. - **Optional Metadata Assignment:** Enhances document uploads with custom metadata for better organization and accessibility. - **Comprehensive UI Settings:** Access and modify important settings like chunking options, temperature, context window size, and chat mode through the user interface. - **Custom Prompts:** Allows adding and using personalized prompts for both the Large Language Model (LLM) and the chat engine. - **Persistent Chats with Titles:** Chats are saved with customizable titles, enabling revisit and continuation of conversations later. - **Persistent Chat History:** Access and review past conversations at any time. - **Model Selection from Ollama:** Allows choosing and pulling AI models directly from Ollama for specific tasks and preferences. - **Streaming Responses with Smart Snippets:** Provides real-time responses with smart output formatting, especially for code snippets. - **Extensible:** Easily extend ToK's functionality with additional features or integrations. ### Tech Stack: