Projects

### Overview: CUPE is a PyTorch model designed for contextless phoneme prediction from speech audio. It processes 120ms frames independently, ensuring each frame's embeddings are acoustically pure. ### Key Features: - 📝 **Model Performance Metrics:** - 🏆 **English (New: Oct2025):** 0.24 (Phoneme Error Rate), 0.21 (Phoneme Group Error Rate) - 🌍 **Multilingual MLS (8 European):** 0.31 (Phoneme Error Rate), 0.26 (Phoneme Group Error Rate) - 🌐 **Multilingual MSWC (38 languages):** 0.49 (Phoneme Error Rate), 0.39 (Phoneme Group Error Rate) ### Tech Stack: PyTorch ### Why It Matters: CUPE excels in providing high-quality phoneme predictions for both English and multilingual speech recognition tasks. Its contextless processing ensures accurate phoneme classification across various languages and dialects.

Overview: GustANN is a high-throughput, cost-effective billion-scale vector search system designed for GPUs. It achieves impressive performance with a speed boost of 7.81x over DiskANN and requires minimal memory, making it ideal for large-scale applications. Key features: - High Throughput: Achieves ~250K QPS on billion-scale datasets. - Memory Efficient: Requires only ~40GB of memory for both GPU and CPU. - Flexible Interface: Supports multiple search modes and storage backends. Tech stack: Cuda Why it matters: GustANN is highly efficient and scalable, making it a valuable tool for developers looking to optimize vector search capabilities on GPUs.

### Overview: Gemma4.pas is a high-performance AI inference engine developed entirely in Delphi, utilizing Vulkan compute shaders to run Google's Gemma 4 model natively on Windows. Unlike most platforms, this project features zero external dependencies and accelerates complex mathematical operations directly on the user's GPU. ### Key features: - **Text Generation**: Streaming token output with configurable sampling (temperature, top-k, top-p) and a private reasoning channel. - **Tool Use**: The model can call functions via structured tool declarations and a complete Jinja chat template engine. - **Vision**: Describing images at variable detail levels (70 to 1120 soft tokens). VCL Graphics decode — no third-party image libraries. - **Audio**: Understanding speech and sound from WAV files (PCM16/F32, any sample rate). 12-layer conformer encoder. - **Video**: Sampling frames across a clip and describing temporal content. Windows Media Foundation frame extraction, reusing the vision tower. - **Embeddings and RAG**: Billed EmbeddingGemma-300m (24-layer bidirectional, 768-dim output) for semantic search, document retrieval, and memory. - **Single-File Deployment**: One VPK archive holds both models (Gemma 4 E4B + EmbeddingGemma-300m), ready to memory-map and run. - **4-bit Quantization**: Q4_0 decoder weights with DP4A integer dot products; vision/audio/embedding encoders stay at full

### Overview: partway is a macOS voice control application that acts partway through your sentence, integrating seamlessly with your existing workflow. It uses TypeSafe's System One model to provide typed decisions and probabilities, allowing for immediate actions mid-sentence. ### Key features: - **Voice Control**: Actively listens to your speech and provides typed responses. - **Jev Integration**: Seamlessly integrates with TypeSafe's System One model for quick and accurate decisions. - **Real-Time Actions**: Mid-sentence actions are performed based on the most likely outcomes. ### Tech stack: Swift ### Why it matters: partway simplifies voice control by providing immediate actions mid-sentence, enhancing productivity and efficiency. Its integration with Jev ensures quick and accurate responses, making it a valuable tool for macOS users looking to streamline their workflow. ### Repository: - **Repository**: tostechbr/partway - **GitHub Stars**: 32 - **GitHub Description**: Voice control for macOS that acts partway through your sentence, built on Jev (TypeSafe) ### Install: 1. Download partway-macos-arm64.dmg and drag it to Applications. 2. Open partway in System Settings → Privacy & Security → Accessibility. 3. Allow the Microphone and Speech Recognition. 4. Turn partway on in System Settings → Privacy & Security → Accessibility. ### What you can say: - "open the notes app", "abre o Safari" - "create a new note", "cria uma nota nova" - "abre o linkedin no google", "open

### Overview: 《动手学世界模型》是一本旨在帮助开发者构建和训练世界模型的项目指南。本书通过从简单的九宫格世界开始，逐步引入图像、历史、遮挡、不确定性和三维空间等概念，逐步构建起一个完整的世界模型。通过学习如何从连续观察中推测看不见的状态、学习时间、行动对世界的影响，以及如何在想象中规划未来，读者可以掌握世界模型的核心概念和技术。 ### Key features: - **亲手训练自己的世界模型**：从一个可以逐格检查的九宫格世界出发，每一步只引入一个新问题：图像、记忆、预测、规划。 - **公式旁边就是代码**：每学一个公式，就在 Notebook 里找到对应的 PyTorch 实现——运行它、修改它、观察模型如何变化。 - **五条路线，任选一条走通**：共同基础完成后，从空间、视频、决策、JEPA、机器人中选择一条路线，完整训练并检验一个小模型。 - **十九份实验，全部可以直接运行**：从基础实验一路排到路线收尾，每份 Notebook 只解决一个问题，开箱即跑。 - **系统学习模型评价**：从单步误差到多步预测、动作响应与陌生场景测试，讲清楚每一种检查能回答什么问题。 - **从复现经典走向独立研究**：复现经典方法、分析一次稳定的失败、完成对照实验，最终提出你自己的模型改进。 ### Tech stack: 使用 Python 编写

### Overview: 面向 DeepSeek Harness 的本地图像生成插件，通过本地 ComfyUI 调用 Qwen-Image 2.1 模型生成图像。插件的网络边界仅限于本地回环（`127.0.0.1`），所有出站请求统一走 `guardedFetch`，保证生成过程不依赖外部网络，无需任何 API 密钥，不产生调用费用。 ### Key features: - **本地回环网络限制**：所有出站请求统一走 `guardedFetch`，强制 `redirect: 'error'`，保证生成过程不依赖外部网络。 - **NVIDIA/Intel/AMD 显卡适配**：优化版支持 NVIDIA、Intel 和 AMD 显卡。 - **Windows/Linux 兼容**：插件在 Windows 和 Linux 系统上均能正常运行。 ### Tech stack: JavaScript ### Why it matters: 插件提供了一种本地生成高质量图像的方法，无需依赖外部 API 密钥或网络，适用于需要稳定性能和隐私保护的场景。

### Overview: 本项目是关于多模态情绪识别方法的研究，支持文本、音频和视频三种模态的情绪分类与情感分析任务。通过集成深度学习技术，实现高效、灵活且多数据集支持的多模态情绪识别套件。 ### Key features: - 多模态融合：支持文本、音频、视频三种模态的单独和组合使用。 - 多数据集支持：支持MELD、IEMOCAP、SIMS等主流情绪数据集。 - 高效训练：集成缓存机制、混合精度训练、知识蒸馏等优化技术。 - 灵活配置：基于TOML配置文件的模块化设计，支持各种实验配置。 - 完整流程：从数据预处理到模型训练、评估的完整工具链。 ### Tech stack: Python 3.12+ ### Why it matters: 本项目为多模态情绪识别提供了一套全面且高效的解决方案，适用于多种应用场景，有助于提高情绪识别的准确性和效率。

Overview: This project is a free AI porn girl generator, AI video creator, and NSFW AI generator. It utilizes the Wan 2.5 model and deepfake technology to produce uncensored and uncensored AI art, making it an essential tool for developers looking to create realistic and engaging content. Key features: - AI porn girl generator - AI video creator - NSFW AI generator - Wan 2.5 model - Deepfake generator - Free access to uncensored AI art Tech stack: C++ Why it matters: This project is crucial for developers who need to create realistic and engaging content for their applications, especially those targeting a mature audience. The Wan 2.5 model and deepfake technology ensure high-quality, uncensored AI art, making it a valuable tool in the development of AI-driven applications.

Overview: This project demonstrates the integration of SpringAI with Ollama to facilitate conversational interactions with library documents, utilizing a vector database for efficient querying and AI-generated responses. Key features: - Document retrieval from vector database - AI-based document embedding creation - Vector database query for nearest document - AI-generated answer based on document content - Spring Boot and Spring AI integration - Ollama support for AI/LLM models - Mcp Client and Server for integrating with external Mcp Servers - Image database search using Postgresql with PGVector extension - Relational database search using LLMs to generate embeddings Tech stack: Java, Spring Boot, Spring AI, Spring AI Ollama support, PostgreSQL, Liquibase, JPA, Angular, Angular CLI, Typescript, Gradle Why it matters: This project showcases the practical application of AI/LLM integration in a library setting, enabling users to query and converse with documents stored in a vector database, enhancing the library's digital capabilities.

### Overview: This repository contains the official codebase for "Quantile Reward Policy Optimization: Alignment with Pointwise Regression and Exact Partition Functions" (Matrenok et al. 2025). The code includes all the scripts used to train and produce the results presented in the paper, as well as infrastructure and experiment management code for running experiments at scale on a SLURM cluster. ### Key features: - All the scripts used to train and produce the results presented in the paper. - All of our infrastructure and experiment management code for running and managing experiments at scale on a SLURM cluster. - A reference implementation for a scalable code sandbox on SLURM clusters with container runtimes, which does not require elevated privileges. ### Tech stack: Python ### Why it matters: This repository is a reference implementation of the algorithm described in the paper "Quantile Reward Policy Optimization: Alignment with Pointwise Regression and Exact Partition Functions" (Matrenok et al. 2025). It provides a comprehensive set of scripts and infrastructure for training and evaluating the model, making it useful for researchers and developers looking to understand and replicate the results presented in the paper.

### Overview: DrawTalking is a prototype user experience concept designed to facilitate the creation of interactive worlds and storytelling through sketching and speaking. It is a Ph.D. thesis project and a notable ACM UIST publication, focusing on the integration of interactive elements into the storytelling process. ### Key Features: - **Sketching and Speaking**: Users can create interactive worlds by sketching and narrating their ideas. - **Programming-like Control**: The system provides a programming-like interface for controlling the interactive elements without requiring programming knowledge. - **Interactive Worlds**: The prototype allows for the building and interaction of worlds that can be explored and controlled through sketches and spoken commands. ### Tech Stack: - **Programming Language**: C ### Why It Matters: DrawTalking represents a novel approach to interactive storytelling and world-building, offering a unique blend of creativity and technical control. It has the potential to inspire future research in natural interfaces for creative exploration and authoring, emphasizing user control and flexibility in storytelling. ### Repository: KTRosenberg/DrawTalking - **GitHub Stars**: 31 - **GitHub Description**: A prototype for building interactive worlds and telling stories at the same time by sketching and speaking. It is a Ph.D. thesis project and a notable ACM UIST publication. - **Primary Language**: C - **Paper**: "DrawTalking: Building Interactive Worlds by Sketching and Speaking" - Paper: https://dl.acm.org/doi/10.1145/3654777.3676334 - **Website**: [

Overview: MemVectorDB is an in-memory VectorDB written in Rust, designed for fast and efficient data storage and retrieval. Key features: - **Fast:** Stores vectors in memory for quick insertion and retrieval. - **Vertical Scalability:** Scalable based on available system resources. - **Metadata Support:** Supports metadata for RAG applications. - **Persistence:** Logs can be restored for data recovery. Tech stack: Rust Why it matters: MemVectorDB is optimized for high-performance vector storage, making it suitable for applications requiring fast data access and efficient resource utilization. Repository: KevKibe/memvectordb GitHub stars: 31 GitHub description: An fast in-memory VectorDB in rust.

### Overview: This comprehensive guide is designed for anyone looking to transition into becoming an AI Engineer. It covers the foundational concepts, hands-on projects, and practical tips needed to build and deploy AI systems. ### Key Features: - **Prerequisites**: Covers Python basics, probability, and tools setup. - **Foundations**: Explains Generative AI, LLMs, and how they work. - **Prompt Engineering**: Teaches how to interact with models effectively. - **LLM APIs**: Guides connecting to OpenAI, Anthropic, and other APIs. - **RAG**: Provides long-term memory capabilities for models. - **Agents & Tools**: Builds autonomous AI systems. - **Fine-Tuning**: Customizes models for specific use cases. - **MLOps**: Covers Docker, CI/CD, and MLflow for AI deployment. - **Production**: Mentors in evaluating, deploying, monitoring, and optimizing AI systems. - **Projects**: Includes 15 real-world projects, from beginner to advanced levels. ### Tech Stack: Python ### Why It Matters: This guide is designed to help anyone, regardless of their background, understand and build AI systems. It provides a clear, practical approach that focuses on the "why" of AI, the "how" of implementation, and the "how to" of production. ### Repository: MuhammadIbtisam/ai-engineer-roadmap ### GitHub Stars: 31 ### Primary Language: Python ### GitHub Description: A complete, no-fluff, human-first guide to becoming an AI Engineer. Built for

Overview: TalentMatch is a streamlined platform designed to match companies with potential candidates using advanced NLP techniques and a user-friendly web interface built with Flask and MongoDB. Key features: - Easy navigation to different sections - Secure user registration and login - Resume and job description matching through advanced NLP - Fine-tuned base spacy English model for technical keyword extraction - Resume upload for candidates - Job description entry for companies - Resume matching based on skills overlap - Infrastructure: Flask, MongoDB, SpaCy, HTML, CSS3, GitHub Tech stack: Flask, MongoDB, SpaCy, HTML, CSS3, GitHub Why it matters: TalentMatch streamlines the process of resume and job description matching, making it an invaluable tool for both recruiters and job seekers in the tech industry, ensuring accurate and relevant matching through advanced NLP techniques.

Overview: This course, taught by AI genius Andrew NG, covers cutting-edge topics in generative AI, including how it works, common uses cases, the life cycle of genAI projects, advanced technologies such as RAG, fine-tuning, and pre-training, and implications for business and society. Key features: - How generative AI works - Common uses cases - Life cycle of genAI projects - Advanced technologies options - Implications for business and society Tech stack: Jupyter Notebook Why it matters: This course provides a comprehensive understanding of generative AI, its applications, and its impact on various industries, making it essential for developers looking to leverage this technology responsibly.

Overview: AI Interview Mocker is a platform that generates personalized mock interview questions using the Gemini API, helping users improve their interview skills and get job-ready. Key features: - Personalized AI Interviews - Voice-Enabled Interaction - Video & Audio Support - Smart Feedback Engine - Easy-to-Use Dashboard - Session History - Secure & Private Tech stack: Next.Js, TailwindCSS, ShadCN, Google Gemini API, Langchain, Clerk, Drizzle, Postgres. Why it matters: It provides a convenient and effective way to practice for job interviews, improving confidence and communication skills.

Overview: pdf.ai is an AI-powered PDF assistant that allows you to chat with your PDF documents using cutting-edge web technologies. Deployed using Vercel Edge Functions, it provides fast, AI-driven responses to your queries. Key features: - Chat with your PDF using AI - Fast responses via Edge Functions - Embedding + context retrieval for smarter answers - Local development runs smoothly Tech stack: Next.js, Vercel Edge Functions, OpenAI GPT, LangChain, Pinecone Vector DB Why it matters: pdf.ai is a powerful tool for anyone looking to interact with PDF documents in real time, enhancing productivity and information retrieval with AI-driven insights.

Overview: XHSWork 是一款基于人工智能的小红书内容创作平台，提供一键生成图文、AI封面设计、爆款笔记仿写和智能编辑器等功能。 Key features: - 一键图文生成 - AI封面设计 - 爆款笔记仿写 - 智能编辑器 - 违禁词检测 - 深度思考模式 - 参考图风格分析 - 一键发布 Tech stack: TypeScript Why it matters: XHSWork简化了小红书内容创作流程，提高了效率，同时提供了丰富的功能，帮助创作者快速生成高质量的内容。 Repository: WTM30124/XHSWork GitHub stars: 31 GitHub description: 基于 AI 的小红书内容创作工具网站，支持一键生成图文、AI 封面设计、爆款笔记仿写、智能编辑器等功能。

Overview: Memorose is a next-generation self-evolving multimodal memory brain designed for long-lived agents. It offers persistent memory, hybrid retrieval, graph context, shared knowledge, and controlled forgetting in a Rust-native stack. Key features: - Layered Memory - Scoped Memory Isolation - Facts and Procedures - Hybrid Retrieval - Memory Lifecycle - Multimodal Native Tech stack: Rust Why it matters: Memorose is a self-hosted memory runtime for agents that need durable memory and flexible memory infrastructure, built for agents that require reliable, continuously evolving, and highly flexible memory. Repository: ai-akashic/Memorose GitHub stars: 31 GitHub description: Next-generation self-evolving multimodal memory brain.

**Overview:** This repository contains LinkedIn posts about Generative AI knowledge sharing, learning resources, and research explanations. The content is contributed by the creator, Ashutosh1919, and can be expanded with additional high-quality posts from other users. **Key features:** - Posts about Generative AI knowledge sharing - Learning resources for AI and machine learning - Research explanations on AI and machine learning topics **Tech stack:** GitHub **Why it matters:** This repository serves as a valuable resource for anyone interested in Generative AI, providing insights and learning materials that can help in understanding and applying AI technologies effectively.