Projects

Overview: Haas-ASR is an inference code for Haas speech recognition models, designed for fast and accurate Japanese speech recognition. The model achieves a weighted CER strict of 6.12% on the asr-ja-bench v1.1, making it competitive with commercial APIs. Key features: - Fast and accurate Japanese speech recognition - Compatible with Linux with NVIDIA GPU support - Supports Python 3.11+ and NeMo 2.7 - Transcribes audio in real-time Tech stack: Python Why it matters: Haas-ASR is a valuable tool for developers looking to integrate fast and accurate Japanese speech recognition into their projects, offering comparable performance to commercial APIs while being more cost-effective. GitHub stars: 32 Repository: third-intelligence/Haas-ASR

### Overview: Cooking AGI is a comprehensive resource for anyone looking to learn about Artificial Intelligence (AI) and Machine Learning (ML). This bilingual (English / Chinese) guide includes worked examples, step-by-step figures, and coding practice, making it an ideal tool for both beginners and advanced learners. The site covers foundational concepts, interview preparation, engineering practice, and tech history, providing a holistic view of AI and ML. ### Key Features: - **Foundations:** How models work, learn, and are evaluated. - **Interview Prep:** ML / LLM review, ML coding, Python, and algorithm patterns. - **Engineering Practice:** Recommendation, RAG, and post-training systems, end to end. - **Tech History:** Power, chips, networks, software, and AI, with original sources. - **Career Notes:** What I tried in internship and new-grad searches. ### Tech Stack: Python ### Why It Matters: Free to read, this resource offers a deep dive into AI and ML, covering everything from foundational concepts to advanced engineering practices. Whether you're a beginner looking to learn the basics or an experienced developer seeking to enhance your skills, Cooking AGI provides valuable insights and practical examples.

### Overview: Hark is an Android application designed for hearing-impaired users, providing an on-device wake word detection feature. It uses PocketSphinx to detect when a configured word is spoken, triggering a heads-up notification immediately, even when the screen is off or the device is in airplane mode. ### Key Features: - **Offline Wake-Word Detection**: Detects wake words on the device without any network calls. - **No Network Calls**: The application runs entirely on the device, ensuring privacy and security. - **Accessibility Feature**: Alerts users when they are addressed, particularly useful for hearing-impaired individuals. ### Tech Stack: - **Language**: Kotlin - **UI**: Jetpack Compose + Material 3 - **Architecture**: MVVM, `ViewModel` + `StateFlow` - **DI**: Hilt - **Persistence**: DataStore ### Why It Matters: Hark is a significant improvement over cloud-based wake-word detection, offering a more reliable and accessible solution for users who rely on hearing aids or cochlear implants. It ensures that users can be alerted without the need for an internet connection, enhancing their independence and quality of life. ### GitHub Stars: 35 ### Repository: tuckercr/wakewordapp ### Primary Language: Kotlin ### Tech Stack: Kotlin, Jetpack Compose + Material 3, MVVM, Hilt, DataStore

### Overview: This repository contains an end-to-end Automatic Speech Recognition project developed by Victor369Basu. The project utilizes PyTorch for the neural network model and MLflow to manage the ML lifecycle, including experimentation, reproducibility, deployment, and a central model registry. ### Key Features: - **Neural Network Model:** The project employs a neural network architecture consisting of Residual-CNN blocks, BidirectionalGRU blocks, and fully connected layers. - **MLflow Integration:** MLflow is used to track and compare model performance, manage experiments, and deploy models. - **LibriSpeech Dataset:** The project utilizes the LibriSpeech dataset for training and validation, including a subset of 2000 files for faster training and validation. - **Feature Extraction:** Feature extraction is performed using Mel-frequency cepstrum coefficients (MFCC) and other techniques like MelSpectogram and Frequency Masking. - **Acoustic Model:** The model architecture includes Residual-CNN blocks, BidirectionalGRU blocks, and fully connected layers for classification. - **Loss Function:** The CTC (Connectionist Temporal Classification) loss function is used as the base loss function. - **Decoding:** Greedy Decoder is used for text decoding. - **ML Lifecycle Pipeline:** The project follows an ML lifecycle pipeline that includes initializing the MLflow server, creating experiments, starting runs, saving models, and registering them for deployment. ### Tech Stack: - **MLflow:** Used for managing the ML lifecycle, tracking, and comparing model performance. - **Pytorch:**

### Overview: AnySD is an official repository for the implementation and benchmark evaluation of AnyEdit, a unified high-quality image editing model that can be applied to various tasks. ### Key features: - AnySD: The official implementation and benchmark evaluation repository for AnyEdit. - AnyBench: A guide for evaluating AnyEdit using the AnyBench tool. - AnySD: The model implementation and benchmark evaluation repository for AnyEdit. ### Tech stack: Python ### Why it matters: AnySD serves as a comprehensive resource for developers and researchers interested in image editing and model benchmarking, providing both the model implementation and evaluation tools.

### Overview: 基于 Gitee AI 的免费 2K 图片生成器，提供 18000+ AI 提示词库和 12 种预设风格，每天可生成 100 张免费 2K 高清图片。内置快速生成功能、完全免费、支持多尺寸适配和提示词库搜索。 ### Key features: - 极速生成：基于 Z-Image-Turbo 模型，8 步快速生成 2K 高清图片。 - 完全免费：每天 100 张免费生成额度，2K 分辨率。 - 风格预设：12 种预设风格（公众号/小红书配图）。 - 多尺寸支持：1:1 / 4:3 / 3:4 / 3:2 / 2:3 / 16:9 / 9:16 七种比例，适配公众号、小红书等平台。 - 提示词库：内置 18000+ 精选 AI 提示词，按用途 / 风格 / 主体分类，支持关键词搜索。 - 历史记录：自动保存生成记录，支持单条删除、一键重新下载。 - 图床上传：一键把生成图上传到 GitHub 图床，外链自动复制，换设备不丢图。 - 隐私安全：API Key 仅保存在本地浏览器，不会上传

### Overview: ReNeg is a reward-guided approach that directly learns negative embeddings for text-to-image and text-to-video models, enhancing their visual output with strong generalization capabilities. ### Key Features: - **Re**ward-guided learning for negative embeddings. - **Neg**ative embeddings exhibit strong generalization capabilities. - Seamlessly adaptable to text-to-image and text-to-video models. - Highly effective yet simple, enhancing the visual appeal of outputs from base Stable Diffusion models. ### Tech Stack: Python ### Why it Matters: ReNeg amplifies the visual appeal of outputs from base Stable Diffusion models, making them more engaging and visually appealing.

### Overview: This project is a native inference provider plugin for the Hermes Agent, integrating Google Antigravity and Cloud Code Assist, enabling state-of-the-art Gemini and Claude models with full streaming support and native tool calling. ### Key features: - **Google Antigravity / Cloud Code Assist Integration**: Seamless integration of Google's Antigravity and Cloud Code Assist into the Hermes Agent. - **Streaming Support**: Full streaming support for Gemini and Claude models. - **Native Tool Calling**: Native tool calling capabilities. - **Reasoning/Thinking Tokens**: Includes reasoning and thinking tokens for enhanced functionality. ### Tech stack: Python ### Why it matters: This plugin solves the problem of authenticating with Google account via OAuth 2.0 PKCE, allowing subscription holders to use their quota without burning paid API tokens, thus unlocking the full potential of their subscription models in Hermes Agent.

Overview: This project is a baseline implementation for sentiment analysis during the COVID-19 pandemic, using BERT for end-to-end fine-tuning. The F1 score on the DataFountain platform is 0.716. Key features: - BERT-based sentiment analysis using end-to-end fine-tuning - Labeled data only, no unlabeled data used - No hyperparameter tuning - Data pre-processing script provided Tech stack: Python Why it matters: This project serves as a simple baseline for sentiment analysis during the pandemic, showcasing the effectiveness of BERT in handling labeled data alone. Repository: zhengyima/ncov_sentiment GitHub stars: 35

### Overview: CRAFT: Corpus Retrieval and Augmentation for Fine-Tuning is a project that introduces a novel method for generating synthetic datasets tailored to specific tasks. The repository contains code, datasets, and checkpoints for the paper "CRAFT Your Dataset: Task-Specific Synthetic Data Generation Through Corpus Retrieval and Augmentation," presented at TACL and EMNLP 2025. ### Key features: - **Synthetic Datasets**: The repository provides datasets for various tasks such as BioQA, CommonSenseQA, MedQA, RecipeGen, and Summarization. - **Corpus Retrieval and Augmentation**: The datasets are generated by retrieving relevant corpus and augmenting them with new examples. - **Few-Shot Learning**: The project includes few-shot learning experiments that use the first 8 lines of each dataset file. - **Performance Metrics**: Models trained on CRAFT datasets match the performance of general instruction-tuned LLMs and even outperform training on human-curated data for tasks like summarization. - **Robustness**: The synthetic datasets have a lower 5-gram overlap with test sets compared to in-domain training sets, making them more robust against distribution shifts. ### Tech stack: Python ### Why it matters: CRAFT's synthetic datasets offer a more efficient and effective alternative to human-curated data, providing a strong performance boost across various tasks, including summarization and few-shot learning.

### Overview: An AI assistant for Dota 2 that analyzes game situations in real-time and provides tactical advice through an overlay. The system uses RAG (Retrieval-Augmented Generation) for more relevant and well-grounded responses. ### Key features: - 🎮 Dota 2 Integration via GSI - 🤖 AI Provider Support: Google Gemini and OpenRouter - 🔍 RAG System: Searches relevant information from knowledge base before generating responses - 💬 Automatic Advice every N seconds - ❓ Interactive Questions directly in-game - ⌨️ Hotkey Management - 📚 Built-in Knowledge Base with information about heroes, abilities, items ### Tech stack: Go ### Why it matters: The AI assistant provides valuable insights and tactical advice to Dota 2 players, enhancing their gameplay experience and strategy.

### Overview: OpenCode vision support automatically saves pasted images and guides non-vision models to recognize them via tool calling. It supports both single and multiple images. ### Key features: - Automatically saves pasted images - Guides non-vision models to recognize them using a tool - Supports single and multiple images ### Tech stack: TypeScript ### Why it matters: This feature enhances the functionality of OpenCode models by enabling them to understand and recognize images, making them more versatile and useful in various applications.

Overview: AutoRef is an innovative project that aims to improve image generation by optimizing the harness around a frozen image generator and reasoning model. It achieves significant improvements in performance across various benchmarks, including MultiBanana tasks. Key features: - AutoRef-Harness: A code-based harness that optimizes the image generator and reasoning model. - Multi-reference image generation: Combines multiple reference images into a single image following an instruction. - FLUX.2 [klein] 4B: Achieves a performance increase of 7.37 from 5.72. - AutoRef-Harness transfers unchanged to unseen reference counts and other generators. - AutoRef-Harness uses a reasoning model (GPT-5.5) to guide the generator and drafts three images per task. Tech stack: Python Why it matters: AutoRef-Harness significantly improves image generation performance, making it a valuable tool for developers looking to enhance their image generation models. Repository: KuOnoda/AutoRef GitHub stars: 34 GitHub description: Official implementation of "Harness Optimization for Agentic Multi-Reference Image Generation"

Overview: MoDES is an efficient and accurate framework for accelerating Mixture-of-Experts (MoE) models in multimodal large language models (LLMs). It achieves significant performance improvements and faster inference times by dynamically skipping experts. Key features: - Efficiently skipping experts to enhance MoE MLLM inference. - Computationally lightweight while maintaining high performance. - Demonstrated performance boost of up to 10.67% on Qwen3-VL-MoE-30B-A3B-Instruct. - Improves inference speed with 2.16x prefilling and 1.26x decoding speedup. Tech stack: Python Why it matters: MoDES offers a practical solution for accelerating MoE MLLMs, making them more efficient and faster to train and infer, which is crucial for applications requiring high performance and low latency in large language models.

### Overview: Generative Media Skills for AI Agents is a comprehensive toolkit designed for AI agents to generate, edit, and display professional-grade images, videos, and audio. Powered by MuAPI.ai, this project offers a high-performance, schema-driven architecture for Claude Code, Cursor, Gemini CLI, and OpenCode. The ultimate multimodal toolset for creating, editing, and displaying professional-grade media content, this project is essential for anyone looking to leverage AI for creative and professional applications. ### Key Features: - **Image Generation**: Create high-quality images with advanced AI algorithms. - **Video Generation**: Generate professional-grade videos with realistic effects and animations. - **Audio Generation**: Produce high-fidelity audio clips and recordings. - **Schema-Driven Architecture**: A schema-driven approach ensures consistent and high-quality results across all media types. ### Tech Stack: - Shell ### Why It Matters: This project is crucial for anyone looking to integrate AI into their creative workflows. Whether you're an artist, designer, or professional, the ability to generate professional-grade media content with ease is invaluable. The schema-driven architecture ensures consistency and quality, making it a go-to solution for AI-driven media creation. ### GitHub Stars: 31 ### Related Projects: - **minimax-music-3-api**: Python SDK for MiniMax Music 3.0 text-to-music generation. - **awesome-minimax-music-3-prompts**: Curated song prompts and lyrics-formatting guide for MiniMax Music 3.0. - **miniMax-H3-API**: Python SDK for MiniMax H3

Overview: This project provides non-blocking Asterisk modules for accessing VoiceKit services, including speech recognition and speech synthesis, which can be used individually or in conjunction with other third-party repositories. Key features: - Speech Recognition - Speech Synthesis Tech stack: Shell Why it matters: This project offers a convenient and efficient way to integrate VoiceKit services into Asterisk, making it easier for developers to implement speech-based functionalities without the need for extensive Asterisk configuration. GitHub stars: 34 Repository: Tinkoff/asterisk-voicekit-modules

### Overview: ChemEval is a comprehensive benchmark designed for evaluating large language models and multimodal large language models in chemistry. It comprises 62 tasks across four levels, covering textbook-level chemical knowledge, literature understanding, molecular understanding, and scientific knowledge deduction. ### Key Features: - **4 Progressive Levels**: From advanced knowledge question answering to molecular understanding. - **13 Capability Dimensions**: Probe objective and subjective QA, information extraction, inductive generation, molecular name recognition, property prediction, retrosynthesis, reaction condition recommendation, outcome prediction, and mechanism analysis. - **Textual and Multimodal Coverage**: Includes text-only tasks and chemistry-related visual inputs such as molecular structure diagrams, statistical charts, tables, reaction profiles, and spectra. - **Expert-Validated Construction**: Combines curated open-source datasets with expert-authored and expert-validated materials, including in-house tasks designed for practical chemical research scenarios. ### Tech Stack: Python ### Why It Matters: ChemEval provides a robust and comprehensive benchmark for evaluating LLMs and MLLMs in chemistry, ensuring that models are evaluated on a wide range of tasks and dimensions, from basic knowledge to advanced scientific reasoning.

Overview: This project introduces a fine-tuned Video Generator that leverages pretrained Stable Diffusion and Disney models to convert text descriptions into video outputs. It includes both training and inference functionalities, making it versatile for various applications. Key features: - Fine-tuning of Stable Diffusion and Disney models for text-to-video generation - Personalized text-to-video generation using DreamBooth models - Training and inference capabilities with detailed instructions Tech stack: Python Why it matters: This project offers a practical solution for creating high-quality video content from textual descriptions, enhancing creativity and efficiency in video production. Repository: Yazdi9/Video-Generator-text-to-video Primary language: Python GitHub stars: 34

Overview: This project aims to build a speech assistant similar to Amazon Alexa or Google Home, allowing users to interact with their devices in real-time without the need for a voice API or network connection. Key features: - Real-time keyword spotting - Wake-up word detection - Speech/voice command recognition - End-point detection Tech stack: Keras Why it matters: This project offers a hands-on approach to building a speech assistant, providing a practical learning experience for developers interested in creating their own voice-controlled devices.

Overview: VecDB is a simple vector embedding database designed for learning and fun projects. It supports multiple storage engines and integrates with AI embeddings for enhanced functionality. Key features: - Supports multiple storage engines (Bolt, Neo4j, etc.) - Integrates with AI embeddings for improved search results - Configurable with environment variables for easy setup Tech stack: Go Why it matters: VecDB is a practical tool for anyone interested in databases and AI, offering a simple yet powerful way to manage and search for vector embeddings. Its modular design makes it versatile for both learning and production use. Repository: alash3al/vecdb GitHub stars: 34 GitHub description: a vector embedding database with multiple storage engines and AI embedding integrations