Projects

### Overview: CrisperWhisper is a highly accurate speech-to-text system that provides both verbatim and intended transcripts. It offers precise word-level timings and can upgrade existing transcripts by identifying and correcting disfluencies and vocal events. ### Key features: - **Word-level timings**: Precise word timing of 30 ms on read speech and 41 ms on conversational speech, the most precise word timing of any system we benchmarked. - **Verbatize**: Transcribes audio to clean, intended transcripts, ensuring the most accurate representation of the speaker's intended meaning. - **Multilingual support**: Works across most languages, including Whisper. - **Seamless longform transcription**: Transcribes audio without the usual chunk-boundary artifacts, ensuring continuous transcriptions. - **Production inference**: A CTranslate2 runtime with speculative decoding and built-in mitigation of Whisper's looping-hallucination failure mode. ### Tech stack: Python ### Why it matters: CrisperWhisper 2.0 is the most accurate verbatim speech recognition system available, providing clean, intended transcripts that are essential for various applications such as TTS data, clinical speech analysis, and dataset construction.

### Overview: This repository is a curated collection of resources for machine learning, machine learning specializations, and data science. It includes a variety of materials such as books, courses, and online resources to help beginners and professionals alike deepen their understanding of machine learning and data science. ### Key Features: - **Books and Courses**: A wide range of books and courses on machine learning, including specialized courses in machine learning and data science. - **GitHub Repositories**: Access to open-source repositories and projects related to machine learning and data science. - **Online Courses (MOOC)**: A selection of free online courses on machine learning, including MOOCs from prestigious universities. - **Data Science Interview Questions**: A comprehensive list of interview questions for preparing for data science job interviews. ### Tech Stack: - **Programming Languages**: Python, R, and other relevant programming languages for data science. ### Why It Matters: Machine learning and data science are crucial fields in today's data-driven world. This repository serves as a valuable resource for anyone looking to enhance their skills in these areas, whether they are beginners or professionals seeking to deepen their knowledge.

Overview: NeuronBlocks is a NLP deep learning modeling toolkit designed to simplify the process of building end-to-end pipelines for neural network model training for NLP tasks. The toolkit includes two major components: Block Zoo for building neural network components and Model Zoo for pre-configured NLP models. It supports a wide range of NLP tasks, including sentence classification, sentiment analysis, and more. Key features: - Block Zoo: Provides commonly used neural network components as building blocks for model architecture design. - Model Zoo: Offers a suite of pre-configured NLP models in JSON configuration files. - Python-based toolkit with support for both Python 3.6 and above and PyTorch 0.4.1 and above. - Supports English and Chinese language support. - Supports a variety of NLP tasks such as sentence classification, sentiment analysis, and more. Tech stack: Python, PyTorch. Why it matters: NeuronBlocks simplifies the process of NLP model development by providing pre-configured models and a modular architecture, making it easier and faster to build and deploy NLP models. Repository: microsoft/NeuronBlocks GitHub stars: 1456 GitHub description: NLP DNN Toolkit - Building Your NLP DNN Models Like Playing Lego GitHub image: https://i.imgur.com/IanH6xI.png GitHub language: Python GitHub stars: 1456

Overview: PixelAnnotationTool is a developer-friendly tool designed to quickly annotate images in directories. Key features: - Manually and automatically annotate images - Uses OpenCV's watershed algorithm for segmentation - Provides a pseudo manual approach with brushes for corrections - Compatible with Linux, macOS, and Windows - Supports C++ and Qt Tech stack: C++ and Qt Why it matters: PixelAnnotationTool simplifies the annotation process, making it accessible to developers and enhancing their workflow efficiency.

Overview: ChunkHound is an open-source codebase intelligence tool that provides context for agents and teams across current code, git history, and technical web research. It aims to give agents and teams cited context to understand software better, ensuring that changes are safe and supported. Key features: - Contextual code research - Cited answers from current code, git history, and technical web research - Deep understanding for research before editing, large PRs, releases, and debugging - Trace bugs and incidents using external evidence - Reconcile code with external documentation Tech stack: ChunkHound is built using Python, with a focus on local-first development. It leverages APIs for embeddings and language models like Claude, Anthropic, and OpenAI. Why it matters: ChunkHound simplifies the context required for software development, making it safer and more effective. It helps teams avoid guesswork and ensures that changes are well-supported and aligned with the broader context of the software. Repository: chunkhound/chunkhound GitHub stars: 1455 GitHub description: Your entire engineering context, deeply understood.

Overview: WebPerf Snippets is a collection of JavaScript snippets designed to measure and debug web performance. These snippets are tested and run in CI, ensuring consistent results across different environments. Key features: - Measure and debug web performance - Fixed code snippets tested in CI - Runs in the console, CLI, DevTools extension, and Agent Skills - Returns structured data following a single contract Tech stack: JavaScript Why it matters: The snippets are deterministic and reliable, making them ideal for AI audits and performance assessments. They provide consistent results across different environments, ensuring accurate and actionable insights. Repository: nucliweb/webperf-snippets Primary language: JavaScript GitHub stars: 1454 GitHub description: Tested JavaScript snippets that measure and debug web performance. Your AI agent runs them for you, or you run them yourself in DevTools: the same code every time.

### Overview: ColliePWA is a self-hosted PWA mobile client for Herdr, enabling users to interact with Claude Code, Codex, and OpenCode from their phones. It features a status dashboard, push notifications, quick actions, and customizable keyboard layouts, all accessible via a browser on iOS or Android. ### Key Features: - **Status Dashboard**: Led by what needs your input, with each pane representing a workspace or agent. - **Push Notifications**: Alerts when an agent is blocked on user input. - **Quick Actions and Slash Commands**: Configured per agent. - **Customizable Keyboard Layouts**: `Esc`, `Ctrl+C`, arrows, modifier combinations, and more. - **Start from the Phone**: A "+" button opens a new page for selecting machines, agents, commands, and folders. - **Output Search and Full Conversation History**: Beyond standard terminal scrollback. - **Files Screen**: Shows agent's workspace files with syntax-highlighted diffs and last commit. - **File Attachments**: Images, markdown, text, and code files. - **Device Pairing**: Always on, requires a token for every request. - **Crews**: Multiple machines' Collies behind one URL, with failover. - **Walky Talky**: Short messages between agents, with receipts and logs. - **Durable Agents**: Long-running sessions on the phone. - **Twelve UI Languages**: Available in 12 languages. - **Herdr Session Switching**: Managed from the web interface. - **PWA Support**: Runs locally on loopback

### Overview: SpeechT5 is a unified-modal speech-text pre-training model designed for spoken language processing tasks. It offers a comprehensive approach to speech and text pre-training, enabling improved performance in various applications such as speech recognition, translation, and text-to-speech synthesis. ### Key Features: - **Unified-Modal Encoder-Decoder Pre-training**: SpeechT5 leverages a unified encoder-decoder architecture for both speech and text pre-training, enhancing the model's ability to handle diverse modalities. - **End-to-End ASR Model with Unpaired Speech Data**: The Speech2C model pre-trains a transformer decoder specifically for end-to-end Automatic Speech Recognition (ASR) using unpaired speech data. - **Speech-Text Pre-Training with Unpaired Textual Data**: SpeechLM enhances speech pre-training by incorporating unpaired textual data, improving the model's ability to generalize across different domains. - **Unified Codec Language Models for Speech Recognition, Synthesis, and Translation**: VioLA models bridge the gap between speech recognition, synthesis, and translation, providing a unified approach to these tasks. - **Expressive Neural TTS with Prosody-Aware Pre-training**: Prosody-SpeechT5 models incorporate prosody-aware pre-training for more expressive text-to-speech synthesis. ### Tech Stack: - **Python**: The primary programming language used for developing SpeechT5 models. ### Why It Matters: SpeechT5 and its variants have significantly advanced the field of spoken language processing by providing unified pre-training approaches that enhance the performance of various downstream tasks, including ASR, translation,

### Overview: Empryo is an AI coding harness designed for desktop, terminal, and CI environments. It maps your repository, edits code by symbol through AST and LSP, and runs checks. Empryo is available at [https://empryo.com](https://empryo.com) and is compatible with TypeScript. ### Key Features: - **Code Map and Analysis**: Empryo reads the map of your repository and edits code by name, ensuring precise replacements. - **Code Editing**: It edits code by name, replacing functions or classes exactly. - **Self-Checking**: Empryo checks its own work, including typechecking, linting, and testing, and fixes them in the same turn. - **Project Memory**: It remembers decisions and past bugs, providing a detailed project history. ### Tech Stack: - TypeScript ### Why It Matters: Empryo is designed to streamline the development process by providing a powerful tool for code management and quality assurance. It ensures that your code is always up-to-date and free of errors, making your projects more efficient and reliable.

Overview: BrowserAI is a project that allows developers to run production-ready Large Language Models (LLMs) directly in their browsers, offering 100% privacy, zero server costs, and offline capabilities. It simplifies the process of integrating AI into web applications. Key features: - 🔒 100% Private: All processing happens locally in the browser. - 🚀 WebGPU Accelerated: Near-native performance. - 💰 Zero Server Costs: No need for complex infrastructure. - 🌐 Offline Capable: Works without internet after initial download. - 🎯 Developer Friendly: Simple SDK with support for multiple engines. - 📦 Production Ready: Pre-optimized models. Tech stack: TypeScript. Why it matters: BrowserAI provides a seamless, privacy-conscious solution for integrating AI into web applications, making it ideal for web developers, companies, researchers, and hobbyists seeking to experiment with AI without infrastructure overhead. Repository: sauravpanda/BrowserAI Primary language: TypeScript GitHub stars: 1450 GitHub description: Run local LLMs like llama, deepseek-distill, kokoro and more inside your browser.

### Overview: This repository contains implementations of various state-of-the-art image classification models using PyTorch, including popular architectures like ResNet, DenseNet, and WRN. The models are optimized for performance and accuracy on datasets such as CIFAR-10, CIFAR-100, MNIST, and ImageNet. ### Key features: - Implementations of ResNet, ResNet-preact, WRN, DenseNet, PyramidNet, ResNeXt, and other advanced models. - Optimized for performance and accuracy on popular datasets. - Compatible with PyTorch and torchvision libraries. - Supports both CPU and GPU training. ### Tech stack: Python, PyTorch, torchvision, NVIDIA Apex. ### Why it matters: This repository provides a comprehensive set of PyTorch models for image classification, making it an excellent resource for developers looking to improve their model performance and accuracy.

Overview: MLX-Tune allows fine-tuning of language models, vision models, text-to-speech, speech-to-text, embeddings, and OCR models on Apple Silicon Macs using the MLX framework. Key features: - Fine-tune LLMs, VLMs, TTS, STT, Embeddings, and OCR models on your Mac. - Leverage unified memory up to 512GB on Mac Studio. - Use an Unsloth-compatible API for your existing training scripts. - Export models in HuggingFace format and GGUF for Ollama/llama.cpp. - Support for 5 TTS models (Orpheus, OuteTTS, Spark, Sesame, Qwen3-TTS) and 7 STT models (Whisper, Moonshine, Qwen3-ASR, NVIDIA Canary, Voxtral, Voxtral Realtime, NVIDIA Parakeet TDT). Tech stack: Python Why it matters: MLX-Tune simplifies the fine-tuning process for language models and related tasks on Apple Silicon Macs, making it easier and more convenient for developers to work with unsophisticated hardware. Repository: ARahim3/mlx-tune Primary language: Python GitHub stars: 1444 GitHub description: Fine-tune LLMs on your Mac with Apple Silicon. SFT, DPO, GRPO, Vision, TTS, STT, Embedding, and OCR fine-tuning — natively on MLX. Unsloth-compatible API. README excerpt

Overview: Agent Development Kit (ADK) is an open-source, code-first TypeScript toolkit designed for building, evaluating, and deploying sophisticated AI agents with flexibility and control. Key features: - Code-First TypeScript: Define agent logic, tools, and orchestration with full type safety. - Browser and Server: Ships ESM, CommonJS, and web bundles. Run agents in Node.js or directly in the browser. - Rich Tool Ecosystem: Built-in tools for Google Search, Google Maps, Vertex AI Search, and URL context. Connect MCP servers, wrap any function as a tool, or add code execution. - Multi-Agent Orchestration: Compose agents into sequential, parallel, loop, and routed workflows. Delegate to remote agents via the A2A protocol. - Dev Tools and CLI: Interactive dev UI for testing and debugging. CLI commands for scaffolding (`adk create`), local testing (`adk run`, `adk web`), and deployment (`adk deploy cloud_run`). Tech stack: TypeScript Why it matters: ADK provides a flexible and modular framework for building, deploying, and orchestrating AI agent workflows, enabling robust debugging, versioning, and deployment anywhere.

Overview: Redner is a differentiable renderer that computes correct rendering gradients without approximation, allowing for backpropagation from image to 3D scene parameters. It supports inverse rendering and fast local shading with physically-based lighting. Key features: - Correct rendering gradients - Physically-based mode for realistic lighting - Fast local shading with accurate gradient estimation - CUDA backend for GPU acceleration - TensorFlow support with ongoing improvements Tech stack: Python, PyTorch, TensorFlow Why it matters: Redner enables real-time optimization and accurate rendering of complex scenes, making it essential for applications in computer graphics, photogrammetry, and inverse rendering.

Overview: TensorFold is a language model inference engine designed for Apple Silicon and NVIDIA GPUs, serving models directly without the need for Python or MLX installation. Key features: - Direct serving of language models from a Zig binary. - Supports Metal and CUDA for optimal performance. - Built-in support for Apple Silicon and NVIDIA GPUs. Tech stack: Zig. Why it matters: TensorFold simplifies the deployment of large language models, making it accessible and efficient for both Apple Silicon and NVIDIA GPU environments. Repository: ashhart/TensorFold GitHub stars: 1382 GitHub description: LLM Inference Engine for Metal and CUDA.

### Overview: This repository contains a collection of data science projects completed by the user, focusing on various techniques including machine learning, natural language processing, and reinforcement learning. The projects are presented as Jupyter notebooks and are intended as a reference for those starting out in the field or looking to build their own portfolio. ### Key Features: - **Machine Learning**: - Predicting Boston Housing Prices - Finding Donors for CharityML - Creating Customer Segments - Training a Smartcab to Drive - Digit Sequence Recognition with a CNN - **Natural Language Processing**: - Disaster Message Classifier - 3-way Sentiment Analysis for Tweets - Cross Language Information Retrieval - Yelp Review Classification ### Tech Stack: Python 3.14 ### Why it Matters: These projects serve as a practical guide for aspiring data scientists, demonstrating various techniques and tools used in the field. The projects are designed to be run on a current Python 3.14 environment, ensuring compatibility with the latest tools and libraries.

### Overview: MOSS-Image is an official repository for a high-quality 5B-parameter DiT model designed for text rendering, layout generation, and efficient fine-tuning. It is compatible with various downstream tasks and platforms, including SGLang, vLLM-Omni, and ComfyUI. ### Key features: - **Text Rendering Control**: Fine-tune the placement of text and visual elements. - **Strong Text Rendering**: Generate clear, accurate text within images. - **Long-Prompt Support**: Accept text prompts up to 4K tokens. - **Versatile Visual Styles**: Create high-quality anime and photorealistic images. - **Easy Fine-Tuning**: Adapt the base model to custom tasks and domains. - **Compact and Efficient**: Efficient image generation with a 5B backbone. - **Heterogeneous GPU Training**: Trained on alternative GPU platforms. ### Tech stack: Python ### Why it matters: MOSS-Image provides a powerful tool for text rendering and layout generation, making it highly valuable for applications requiring high-quality image output. Its compatibility with various platforms and its efficient training on alternative GPU setups make it a versatile solution for developers and researchers alike.

### Overview: 这是一个汇总了日本语LLM（语言模型）资源的项目，包括日本语LLM的介绍、模型的详细信息、以及开发者的使用指南。该项目由日本语言模型研究开发中心维护，旨在为开发者提供一个全面的资源库。 ### Key features: - 日本语LLM的介绍 - 模型的详细信息，包括模型名称、模型大小、开发人员、以及使用的模型和数据集 - 模型的使用指南，包括模型的安装、配置、以及使用方法 ### Tech stack: 使用了GitHub作为项目存储库，TypeScript作为主要编程语言。 ### Why it matters: 这个项目为开发者提供了日本语LLM的全面信息，帮助他们选择合适的模型进行开发，同时也为研究和开发提供了宝贵的资源。

Overview: Paritok is a non-destructive compression gateway designed to cut token bills for AI coding agents by 25% on the first request and up to 85% for long or saturated sessions. It fits 3x more turns in the same context window and is powered by the first open-source 4B compression model specifically trained for coding agents. Key features: - Non-destructive compression of tool schemas, results, and file reads. - Compounding savings over a session. - Fits 3x more turns in the same context window. - Works with Claude Code, Cursor, Codex, and any agent that supports BASE_URL. Tech stack: Python. Why it matters: Paritok simplifies the token bill for AI coding agents by reducing the input token cost, making it more efficient and cost-effective to use AI in the development process.

### Overview: This repository is a curated list of resources focused on Offensive AI, covering various types of attacks, defenses, and tools used in ethical hacking and adversarial machine learning. ### Key Features: - **Adversarial Machine Learning**: A comprehensive guide to adversarial machine learning attacks, including extraction, inversion, poisoning, and evasion methods. - **Extraction**: Techniques for stealing model parameters and hyperparameters. - **White-box**: Solvable with linear equations. - **Black-box**: Requires training a substitute model. - **Defensive Actions**: Strategies to mitigate these attacks. - **Rounding output values**. - **Differential privacy**. - **Ensembles**. - **Tools**: Various tools and libraries used in the field. - **ART**. - **Cleverhans**. ### Tech Stack: HTML ### Why It Matters: This repository is essential for developers and security professionals looking to understand and implement offensive AI techniques. It provides a practical guide for ethical hacking and defense against adversarial machine learning attacks.