Projects

### Overview: This project aims to compile a comprehensive list of questions and answers for Generative AI interviews. The primary goal is to provide a foundational understanding of Generative AI, including key concepts, techniques, and best practices. The project is based on the "Mastering LLM" course and includes the author's own modifications and insights. ### Key Features: - **Generative AI Interview Questions and Answers**: A collection of questions and answers for interviews focused on Generative AI. - **LLM and Prompting Basics**: Explains the difference between predictive/discriminative and generative AI models. - **Large Language Models (LLM)**: Provides a detailed definition and overview of LLMs, including their training methods and applications. - **Retrieval Augmented Generation (RAG)**: Discusses how RAG models integrate retrieval with generative models for improved performance. - **Chunking Strategies**: Techniques for breaking down text into manageable chunks for better understanding and processing. - **Embedding Models for Retrieval**: Explains the role of embedding models in improving the accuracy of retrieval-based systems. - **Vector Retrieval, Databases, and Indexes**: Covers the basics of vector retrieval, databases, and indexes for efficient data access. - **Advanced Search Algorithms**: Discusses more advanced search techniques for improved query efficiency. - **Language Model Internal Workings**: Provides insights into the inner workings of language models, including the architecture and training methods. ### Tech Stack: - **GitHub Repository**: a-tabaza/genai_interview_questions - **Primary Language**: Unknown -

Overview: This project allows developers to generate materials and textures from text prompts in Unity, utilizing models like `chatgpt` and `dall-e`. Key features: - Supports URP and HDRP materials - Provides an easy-to-use interface for generating materials and textures - Utilizes `chatgpt` and `dall-e-3` models for generating textures Tech stack: C# Why it matters: This tool simplifies the process of creating materials and textures in Unity, making it easier for developers to create visually appealing assets without needing deep knowledge of shaders or advanced modeling techniques.

Overview: This project integrates vLLM with Ray Serve for fast and scalable LLM serving, providing additional features such as multi-server model deployment and autoscaling. Key features: - Multi-server model deployment - Autoscaling - Streaming output Tech stack: Python Why it matters: This integration allows for efficient and scalable inference of large language models, making it suitable for real-time applications and high-traffic environments.

### Overview: This workshop provides systematic approaches to evaluating Generative AI workloads for production use. You'll learn to build evaluation frameworks that go beyond basic metrics to ensure reliable model performance while optimizing cost and performance. ### Key features: - Operational Metrics: evaluate cost and performance. - Quality Metrics: tune and optimize the quality of results. - Understanding Failures: discover failure patterns by reading agent traces. - Agentic Metrics: evaluate agents and use agents for evaluation. ### Tech stack: Jupyter Notebook ### Why it matters: This workshop teaches you how to build evaluation frameworks that go beyond basic metrics to ensure reliable model performance while optimizing cost and performance.

### Overview: Kesha Voice Kit is a powerful tool designed for developers, providing local speech to text and text to speech capabilities. It is optimized for Apple Silicon and runs efficiently on Linux and Windows, making it ideal for both macOS and Windows environments. ### Key Features: - **Transcribe Locally**: Transcribes speech to text in 25 languages, up to 19x faster on Apple Silicon and 10x on an x86 CPU. - **Speak Back**: Supports text-to-speech in 9 languages. - **Plug into Agents**: Integrates seamlessly with agents and scripts, allowing for voice message responses. - **Small Rust Engine**: Runs on a single binary file, approximately 74 MB, without the need for ffmpeg, Python, or native Node addons. ### Tech Stack: The project uses TypeScript for the frontend and Rust for the engine, leveraging the SmallRust framework. ### Why It Matters: Kesha Voice Kit is designed to be the fastest speech-to-text and text-to-speech solution on Apple Silicon, with a significant performance advantage over Whisper. It is optimized for performance and efficiency, making it a valuable tool for developers looking to integrate voice functionality into their projects. ### Repository: The project is hosted on GitHub at [drakulavich/kesha-voice-kit](https://github.com/drakulavich/kesha-voice-kit). ### GitHub Stars: The repository has received 79 GitHub stars, indicating its popularity among the developer community. ### Tech Stack: TypeScript ### Why It Matters: Fast

### Overview: AI-powered literature review tools leverage machine learning to expedite and enhance the scholarly process of identifying, analyzing, and synthesizing relevant research. ### Key features: - **Top Computer Science Literature Databases**: Access prominent databases containing computer science literature. - **Generative AI Prompt Tools**: Assist in creating human-like text based on given prompts. - **Research Gaps**: Identify gaps in existing research. - **Paper Discovery**: Find scholarly articles related to specific topics. ### Tech stack: Unknown ### Why it matters: The AI-powered literature review tools streamline the scholarly process, ensuring a comprehensive grasp of existing knowledge and identifying underexplored domains and emerging inquiries within specific fields.

Overview: This is an offline voice recognition module designed for MagicMirror², allowing users to control the mirror's modules and modes through voice commands. Key features: - Supports MagicMirror² integration - Handles offline operation for privacy - Configurable keyword for user input - Configurable timeout for keyword recognition - Configurable microphone ID - Supports multiple modes and commands Tech stack: JavaScript, Node.js, SphinxBase, PocketSphinx, Sphinx-continuous, lmtool Why it matters: Ideal for users who prefer to keep their voice commands private and without the need for a server, making MagicMirror² more accessible and user-friendly. Repository: fewieden/MMM-voice GitHub stars: 79 GitHub description: Offline Voice Recognition Module for MagicMirror²

Overview: WeiClaw is a project that allows you to integrate any AI agent into the WeChat ecosystem with a single command. It supports full modal communication between WeChat and AI agents, including text, images, voice, video, and file sharing. Key features: - Zero-dependency integration via `npx` command - Supports any OpenAI-compatible API (Codex, Gemini, Claude, OpenCode, self-built) - Full modal support: text, images, voice, video, and file sharing - Supports multiple Agent modes - Agent can send messages actively - Agent can receive messages actively Tech stack: JavaScript Why it matters: WeiClaw simplifies the integration of AI agents into the WeChat ecosystem, making it easy to access cutting-edge AI capabilities without the need for complex configuration or dependencies. Repository: kellyvv/WeiClaw Primary language: JavaScript GitHub stars: 79 GitHub description: Connect WeChat to any AI Agent with one command README excerpt: <p align="center"> <img src="docs/banner.png" alt="weiclaw" /> </p> <h1 align="center">weiclaw</h1> <p align="center"> <a href="https://www.npmjs.com/package/weiclaw"><img src="https://img.shields.io/npm/v/weiclaw?style=flat-square&color=cb3837" alt="npm" /></a> <a href="https://github.com/kellyvv/weiclaw"><img src="https

### Overview: Generative AI Pharmacist is an animated avatar designed by Kenneth Leung Ty, showcasing the capabilities of generative AI in providing medication counseling. This project aims to demonstrate the potential of AI in healthcare through realistic and professional-sounding interactions. ### Key Features: - **Medication Counselling**: The avatar provides brief medication counseling on two commonly prescribed drugs. - **Animated Avatar**: Features a realistic animated pharmacist avatar. - **Realistic Voice Acting**: Uses AI-generated audio to make the interaction sound professional and realistic. - **Generative AI Tools**: Utilizes Midjourney for image generation, ChatGPT for script creation, ElevenLabs for voice synthesis, and D-ID for video editing. ### Tech Stack: - **Midjourney**: Image generation tool by OpenAI. - **ChatGPT**: AI tool for generating script. - **ElevenLabs**: Voice synthesis service. - **D-ID**: Creative reality studio for video editing. ### Why it Matters: This project showcases the potential of generative AI in healthcare by providing realistic and professional-sounding medication counseling. It highlights the speed and improvement of AI tools in the field of healthcare, making it a valuable demonstration for the development of AI in medicine.

### Overview: This library provides continuous speech recognition functionality for Android apps, allowing developers to integrate voice recognition into their applications with options to use Google Voice Input in a Dialog or perform offline speech recognition. It supports both online and offline modes, ensuring seamless integration with minimal impact on app performance. ### Key Features: - **Online Speech Recognition**: Uses Google Voice Input in a Dialog for real-time voice recognition. - **Offline Speech Recognition**: Enables offline speech recognition, providing offline translation capabilities. - **Permissions Handling**: Automatically requests and handles RECORD_AUDIO permission internally, eliminating the need to request permissions in the manifest. - **Listener Interface**: Implements `OnSpeechRecognitionListener` for event-driven callbacks, including `OnSpeechRecognitionStarted`, `OnSpeechRecognitionStopped`, `OnSpeechRecognitionFinalResult`, and `OnSpeechRecognitionError`. ### Tech Stack: Java ### Why it Matters: This library simplifies the process of integrating speech recognition into Android apps, making it easier for developers to implement voice-based functionalities without the hassle of handling permissions or implementing complex speech recognition logic. The ability to switch between online and offline modes ensures flexibility and compatibility across different environments, making it a valuable tool for developers working on voice-based applications.

### Overview: Meuxe is a companion app that remembers conversations, talks back, and stays on your machine. It uses your existing Claude Code agent and can optionally add cloud voices. ### Key Features: - **Remembers Conversations:** Keeps track of past conversations and remembers what matters to you. - **Personalized Agent:** Runs through the CLI agent you install, not a locked-in model API. - **Live2D and VRM Characters:** Expressive avatars with live captions and customizable expressions. - **Voice Interaction:** Speaks aloud with live captions and listens through your microphone. - **Mini Mode:** Transparent desktop widget with chat and mic controls. - **Keyboard Shortcuts:** Toggle mini mode, focus chat, and control the mic. - **Local Privacy:** Memories, persona, and relationship state stay on your machine. - **Editable Profiles:** Layered companion profiles as `.yaml` and `.md` files. ### Tech Stack: TypeScript ### Why It Matters: Meuxe is a companion app that remembers conversations, talks back, and stays on your machine. It uses your existing Claude Code agent and can optionally add cloud voices.

Overview: Scripty is a speech-to-text bot for Discord, rewritten in Rust. It offers a native binary for production testing and a guide for building from source. Key features: - Speech-to-text bot for Discord - Rust-based implementation Tech stack: Rust Why it matters: Scripty simplifies the process of integrating speech recognition into Discord bots, making it easier for developers to add this feature to their projects.

### Overview: Inaudio is an offline dictation and read-aloud application designed for the desktop. It allows users to dictate text into the application and listen to system audio, such as meetings or videos, with real-time transcription and read-aloud functionality. ### Key Features: - **Offline Dictation**: Dictate text into the application and have a searchable transcript history. - **Live Transcription**: Transcribe system audio, such as meetings or videos, as it is heard. - **Audio Source Selection**: Choose between system audio, system audio with microphone, or a specific audio device. - **Noise Reduction**: Clean up background noise on the device using a voice detector and optional speech cleanup model. - **Read Aloud**: Read text or your clipboard aloud with on-device speech synthesis. - **Model Management**: Manage speech models, microphone settings, and playback in one desktop app. ### Tech Stack: - TypeScript - Electron - React - Sherpa-Onnx ### Why It Matters: Inaudio simplifies the process of dictating and listening to audio, making it accessible even when internet connectivity is limited. The offline functionality ensures that users can capture and transcribe audio without an internet connection, enhancing productivity and accessibility. ### Repository: - **Repository**: shikki841/inaudio - **GitHub Stars**: 79 - **GitHub Description**: Offline dictation and read-aloud for the desktop. - **Primary Language**: TypeScript - **GitHub Actions**: CI status badge - **MIT License**: License badge - **Latest Release**: GitHub release badge

### Overview: OpenSakura is MekaHime's first AI companion, an open-source 3D VRoid avatar designed for web and Discord interactions. It features real-time voice and emotion state, powered by OpenRouter or Ollama, with a focus on adult-oriented interactions. ### Key Features: - **3D VRoid Avatar**: Acts out mood, memory, and voice. - **Real-Time Voice**: Communicates through text and voice. - **Emotion State**: Tracks and displays 26 emotions. - **Memory**: Stores and retrieves user-specific memories. - **OpenRouter or Ollama Brain**: Enables seamless communication with Discord and Ollama. ### Tech Stack: - **Python**: The primary programming language used for the backend and frontend. - **3D Graphics**: Utilizes three.js for rendering the avatar. - **WebRTC**: For real-time communication between the user and the avatar. ### Why It Matters: OpenSakura represents MekaHime's first attempt at creating an AI companion, offering a glimpse into the future of AI integration into everyday interactions. Its open-source nature allows for further development and improvement, making it a valuable tool for researchers and developers interested in AI and virtual assistants. ### Repository: zeekk0/OpenSakura - **Stars**: 65 - **Description**: OpenSakura: MekaHime's first AI companion, open source. 3D VRoid avatar, real-time voice, emotion state, OpenRouter or Ollama brain. - **GitHub URL**: https

### Overview: This repository contains the official source code for the paper "Diffusion-based Conditional ECG Generation with Structured State Space Models" by Juan Miguel Lopez Alcaraz and Nils Strodthoff. The paper proposes algorithms, primarily SSSD-ECG, for generating 12-lead ECG signals conditioned on disease labels. ### Key features: - SSSD-ECG: A structured state space model for generating ECG signals. - Conditional generation: The model generates ECG signals conditioned on disease labels. - Diverse algorithms: Primarily SSSD-ECG, with other algorithms as secondary. ### Tech stack: Python ### Why it matters: This repository provides a comprehensive solution for generating ECG signals conditioned on disease labels, which is crucial for medical research and applications in cardiology.

Overview: FlowDPS is an innovative framework that extends diffusion inverse solvers into the flow framework, providing a more general perspective on generative modeling and inverse problems. Key features: - Integrates diffusion inverse solvers into the flow framework - Seamlessly integrates into a latent flow model with a transformer architecture - Performs posterior sampling through flow-driven methods - Confirms superior performance across four linear inverse problems without additional training Tech stack: Python Why it matters: FlowDPS offers a more general approach to generative modeling and inverse problems, potentially leading to improved performance and broader applicability compared to existing methods.

Overview: This repository contains code to convert spoken words into text form using a speech recognition model hosted on IBM Cloud. Key features: - Converts spoken words into text form - Uses a pre-trained model hosted on IBM Cloud - Supports input in the form of a WAV file Tech stack: Python Why it matters: This tool can be useful for applications that require speech recognition, such as virtual assistants, transcription services, or educational tools. Repository: IBM/MAX-Speech-to-Text-Converter GitHub stars: 78 GitHub description: Converts spoken words into text form.

Overview: Lille is a powerful 130-million-parameter language model trained from scratch as part of a fully open-source deep learning stack. It includes a custom tokenizer, dataset, and optimizer, making it efficient and performant. Key features: - Lille-130M-Base: A foundational model pretrained on 4.27 billion tokens from the FineWeb-Edu dataset, with strong general knowledge and text completion abilities. - Lille-130M-Instruct: A model fine-tuned on the Kyoto-Corpus, excelling at instruction-based tasks such as following commands and engaging in chat. Tech stack: Python Why it matters: Lille is a significant contribution to the field of natural language processing, offering a compact yet powerful model that can be used for various applications, including chatbots, instruction-following, and more. Repository: Nikityyy/lille Primary language: Python GitHub stars: 78 GitHub description: A powerful 130-million-parameter model trained from scratch as part of a truly open-source stack, including a custom tokenizer, dataset, and optimizer. README excerpt: # Lille 130M  ## Table of Contents 1. [Model Summary] 2. [Evaluation] 3. [How to Use] 4. [Training and Finetuning] 5. [Training Details] 6. [Limitations] 7. [The Truly Open-Source Stack] 8. [License] 9. [Citation] ##

### Overview: QuarkStar is a self-contained inference engine optimized for **Qwen3.6-35B-A3B** and **Qwen3.8-27B**. It is designed to run on low-cost hardware, including a $150 machine with a BC-250 GPU, and can be used on Apple Silicon with Metal. The engine is built with Vulkan and Metal, and supports both Vulkan and Metal runtimes. It is lightweight and does not require a general GGUF runner, making it ideal for small, low-cost devices. ### Key Features: - **Vulkan and Metal Runtimes**: QuarkStar supports both Vulkan and Metal runtimes, offering flexibility in hardware compatibility. - **GPU Optimization**: The engine is optimized for the BC-250 GPU, leveraging its fast memory and unified memory capabilities. - **Model Support**: Supports **Qwen3.6-35B-A3B** and **Qwen3.8-27B** models, with Qwen3.6-35B-A3B being fast and tolerant of aggressive quantization. - **Lightweight and Efficient**: Designed to run on low-cost hardware, making it suitable for personal machines and even low-end devices like toasters. - **Compressed KV Caches**: Utilizes compressed KV caches for efficient context management, enabling long context runs. - **AI-Driven Development**: Built with assistance from AI coding agents, ensuring high-quality and efficient development. ### Tech Stack: - **Programming Languages**: C - **GitHub

### Overview: llm_bridge is a tiny, vendor-neutral library that provides a unified interface for interacting with various large language models (LLMs). It supports OpenAI, DeepSeek, Alibaba Qwen, AWS Bedrock, Google Gemini, and allows you to use your own backend by wrapping any function with the `callable` provider. ### Key features: - **Canonical Interface + Thin SDK Adapters:** A single interface (`LLMClient`) with pluggable adapters for different LLM providers. The SDKs handle the heavy lifting (retries, streaming, types), while llm_bridge normalizes the calls. - **Zero Required Dependencies:** The core (`mock`, `callable`) is pure standard library. Each vendor SDK is an optional extra you can install on demand. - **Bring Your Own Backend:** You can wrap any function with the `callable` provider, making it easy to integrate with your own backend. - **One Interface:** `chat(messages)` and `complete(prompt)` returning a normalised `LLMResponse` (content, token usage, latency). ### Tech Stack: Python ### Why it Matters: llm_bridge simplifies the process of integrating different LLMs into your application by providing a single, unified interface. It allows you to switch between different providers without changing your application code, making it highly flexible and easy to use.