Projects

### Overview: This notebook provides a step-by-step implementation of the Google Veo 3 architecture from scratch, based on the technical report and model card published by Google. It covers data preprocessing, model architecture building, training, inference, and evaluation, offering a comprehensive guide for understanding and implementing the Veo 3 system. ### Key Features: - Data preprocessing and preparation - Building the Veo 3 model architecture - Model training - Inference - Model evaluation - Safety and ethical considerations ### Tech Stack: The notebook uses Python 3.9+ and JAX for numerical computations. It also includes the following libraries and resources: - PyTorch - Video processing libraries - Model card from Google ### Why it Matters: This notebook is designed for students, researchers, and anyone interested in understanding and implementing the Veo 3 system from scratch. It provides a modular, step-by-step approach to building and training the Veo 3 model, making it a valuable resource for anyone looking to dive into the Veo 3 architecture. ### Repository: The repository is named `FareedKhan-dev/google-veo3-from-scratch` and has been starred by 84 people. It is hosted on GitHub and is under the Apache 2.0 license. ### Primary Language: The notebook is written in Jupyter Notebook, which is a web application that allows you to create and share documents that contain live code, equations, visualizations, and narrative text. It is a great tool for data scientists, engineers

Overview: This repository contains the official code for the paper "Generative Enhancement for 3D Medical Images" presented at IJCV'2026. The project focuses on generating high-quality 3D medical images using generative models, specifically leveraging latent diffusion models. Key features: - Preprocessed datasets and pretrained models available for BraTS and abdomen datasets. - Training scripts for three different models: KL-f8 AE, position conditioned slice generation model, and a modified version of Make-A-Volume. - Inference scripts for various datasets and post-processing settings. Tech stack: Python Why it matters: This project offers a comprehensive solution for generating high-quality 3D medical images, which can significantly improve the accuracy and visualization of medical imaging data, particularly in fields such as neurology and radiology. GitHub stars: 84 Repository: HKU-MedAI/GEM-3D

### Overview: OmniAgent is the first native omni-modal agent for active video perception, achieving a 73% reduction in frames compared to Qwen2.5-VL-72B on LVBench. It treats perception as reasoning, iteratively observing, thinking, and acting through on-demand `get_frames`, `get_audio`, and `get_clip` actions. ### Key features: - **Native Active Perception**: Treats perception as reasoning, iteratively observing, thinking, and acting. - **Efficiency**: Reduces frames by 73% compared to Qwen2.5-VL-72B on LVBench. - **Model Size**: 7B parameters, enabling advanced reasoning capabilities. ### Tech stack: Python ### Why it matters: OmniAgent is a significant advancement in the field of active perception, making it a valuable tool for researchers and developers looking to enhance their models' ability to understand and interact with video data.

Overview: Grok Imagine Image 2.0 API is a high-fidelity image generation service provided by APIMart, offering a one-stop solution for image creation and modification at competitive prices. Key features: - Supports OpenAI-compatible endpoint - Pricing per image by resolution and quality - Supports multiple image resolutions and sizes - Flexible prompt and input options - Real-time billing and refund tracking - Asynchronous task processing Tech stack: Python Why it matters: Grok Imagine Image 2.0 API simplifies the process of generating high-quality images, making it accessible and cost-effective for developers and businesses alike. GitHub stars: 84 Repository: apimartkjp84/grok-imagine-image-2.0-api-ja

### Overview: 虚拟AI基础设施团队是一个开源项目，旨在将本地AI推理优化自动化的流程转变为一个受策略约束、可验证和可回滚的闭环系统。该系统独立地验证模型性能，确保质量标准，并在同一个OpenAI兼容API地址上完成升级或恢复。 ### Key features: - **自动化优化流程**：从候选发现到在线验证，整个过程自动化，无需手动干预。 - **独立验证**：模型只能提出计划，确定性代码负责判定，确保结果的准确性。 - **可回滚升级**：只有质量、速度、内存和错误门禁全部通过的候选优化才会被升级，失败则保留或恢复上一版本。 - **稳定接口**：客户端始终连接到`127.0.0.1:8000/v1`，确保API的稳定性。 - **独立质量验证**：模型性能通过独立验证，确保质量标准。 - **自动回滚**：失败的候选优化不会成为当前Recipe，并触发保留或恢复。 ### Tech stack: Python 3.11–3.13 ### Why it matters: - **自动化优化流程**：将本地AI推理优化工作流自动化，提高效率和准确性。 - **独立验证**：确保模型性能的准确性和稳定性。 - **可回滚升级**：在质量、速度、内存和错误门禁通过后，自动升级或恢复，避免手动干预。 - **稳定接口**：客户端始终连接到同一个OpenAI兼容API地址，

### Overview: Natter is a fast, local-first dictation app for macOS that transcribes speech into clean, formatted text. ### Key features: - **Raw**: Types the final Parakeet transcript after stop with no cleanup or technical formatting. - **Agent**: Starts as a Fast mode for low-latency technical dictation. - **Clean**: Starts as a Refine mode for guarded punctuation and flow improvements. - **Email**: Starts as a Rewrite mode for email content. - **Article**: Starts as a Rewrite mode for article content. ### Tech stack: Swift ### Why it matters: Natter provides a reliable, local-first dictation experience for developers who frequently use macOS, ensuring that their speech input is accurately transcribed and formatted without leaving their workflow.

### Overview: A speech-to-text framework and bot for Discord, designed to be used by hearing-impaired and deaf individuals to control their Discord server through voice commands. The bot supports Google Speech-to-Text and WitAI APIs for transcribing audio into text, making it accessible for those with hearing impairments. ### Key features: - **Voice Command Control**: Users can control their Discord server using voice commands. - **Transcription**: Transcribes audio from text channels into readable text. - **Integration with Discord**: Integrates seamlessly with Discord servers, allowing users to interact with the bot within voice and text channels. - **Accessibility**: Ideal for hearing-impaired and deaf users, providing a way to control their server without relying on visual input. ### Tech stack: - Node.js - Discord.js - Google Speech-to-Text API - WitAI API ### Why it matters: This bot offers a practical solution for those who may face challenges in using traditional voice commands due to hearing impairments, providing a more accessible way to interact with their Discord servers.

Overview: HarmonyDagger is a tool designed to protect audio against generative AI models, ensuring that the human listening experience remains unaltered while effectively masking AI-generated noise. Key features: - Psychoacoustic Masking - Adaptive Scaling - Phase Perturbation - Temporal Forward Masking - Vocal-Specific Mode - Dry/Wet Control - Multi-channel Support - Multiple Audio Format Support - Robustness Testing - Protection Verification - Benchmark Reporting - Visualization Tools - Parallel Batch Processing - Streamlit Web Demo - Docker Support - PyPI Package Tech stack: Python Why it matters: HarmonyDagger provides a robust and effective solution for audio protection against generative AI, ensuring that AI-generated noise does not compromise human listening experience while maintaining the quality of the original audio. Repository: jaschadub/harmonydagger GitHub stars: 84

Overview: This is an unofficial desktop app for Google Bard, a Generative AI chatbot, designed for use on Windows, Linux, and macOS. Key features: - Supports all desktop platforms including Windows, Linux, and macOS - Provides a stable and reliable experience for users Tech stack: JavaScript Why it matters: This app offers a convenient way to interact with Google Bard on a desktop, making it accessible to a wider audience and providing a more familiar interface for those accustomed to using traditional desktop applications. Repository: mantreshkhurana/Google-Bard-electron GitHub stars: 84

Overview: This repository contains the official PyTorch implementation for MaPO, a model designed to align diffusion models without reference images. MaPO aims to improve the quality of generated images by incorporating user preferences into the optimization process. Key features: - MaPO is a model for aligning diffusion models without using reference images. - It uses a preference optimization approach to enhance the quality of generated images. - The repository includes a training script (`train.py`) and a LoRA variant (`train_with_lora.py`) for experimenting with the model on consumer GPUs. Tech stack: Python Why it matters: MaPO offers a novel method for improving image quality by incorporating user preferences, making it valuable for applications that require high-quality image generation without the need for reference images.

### Overview: This comprehensive training course is designed to guide developers from the foundational concepts of AI agents to building complex, production-ready systems using the Google Agent Development Kit (ADK). It covers the complete ADK 2.0 curriculum across 45 modules, providing hands-on practice and real-world application through interactive knowledge checks, student progress tracking, and a student completion certificate. ### Key Features: - Interactive Knowledge Checks (`KnowledgeCheck`) with Socratic feedback, misconception breakdowns, guided hints, and reflection cooldowns. - Student Progress Tracker & Gated Completion with real-time completion tracking, interactive checklist drawer, and multi-tab sync. - Official Student Completion Certificate with custom learner name, date, verification ID, and credential checklist. - Interactive Solution Vault (`SolutionVault`) with fast completion prompt, dedicated workshop facilitator support, and WCAG AAA high-contrast light mode styling. - Instant Offline Local Search powered by `@easyops-cn/docusaurus-search-local`. - Automated Student Grader (`adk-student-evaluator`) with zero-cost empirical verification framework. - Zero-Setup Cloud Onboarding with pre-configured Python 3.11, `uv`, and GCP tooling. - Full ADK 2.0 API Alignment with comprehensive coverage of Human-in-the-Loop (HITL), Context Caching & Compaction, Session Rewind, Callbacks, and Custom Plugins. ### Tech Stack: Python ### Why It Matters: This course is designed to empower developers with the skills needed to build the next generation of AI agents, transforming them

Overview: Seedream 5.0 Pro Prompts offers a comprehensive set of 8 tested recipes for generating high-quality images, including 1K and 2K layers at various resolutions. The API is compatible with OpenAI and supports real-time pricing and async task management. Key features: - 8 tested recipes for generating images - 1K and 2K layer pricing options - Real-time pricing with `cost` field in response - Supports 1K, 1K-layer, and 2K resolutions - Batches and retries for async tasks Tech stack: Python Why it matters: Seedream 5.0 Pro Prompts simplifies the process of generating high-quality images, making it accessible and cost-effective for developers and businesses alike. The API's real-time pricing and async task management features ensure smooth integration and efficient use of resources. Repository: nj4naiem992lmr/seedream5.0-seedream-5.0-prompts Primary language: Python GitHub stars: 84 GitHub description: seedream-5.0 API (seedream5.0 / seedream 5.0): 1K-layer $0.0146; 1K $0.0293; 2K-layer $0.0293. Model id, prompts settings, curl and Python examples over an OpenAI-compatible gateway

Overview: Parsee AI is a Python-based framework designed to simplify the extraction and structuring of data from common unstructured sources like PDFs and HTML files. It supports both LLMs and custom models, making it versatile for various financial and non-financial use cases. Key features: - Extraction of structured data from PDFs and HTML files - Support for tabular data extraction and multimodal queries - Integration with LangChain for advanced natural language processing - Extensive support for financial data extraction Tech stack: Python Why it matters: Parsee AI simplifies the process of extracting and structuring data, making it accessible to developers and data analysts without requiring deep knowledge of natural language processing or complex data extraction techniques. Its ability to handle both LLMs and custom models makes it highly versatile and efficient, especially in the financial domain where it has been successfully used by SimFin.

Overview: vCache is a reliable and efficient semantic prompt caching system designed to reduce the latency and cost of large language models (LLMs) by providing cached responses for semantically similar prompts. Key features: - Online-learned, embedding-specific decision boundaries - Guarantees user-defined error rate bounds - Compatible with OpenAI API and other inference setups Tech stack: Python Why it matters: vCache optimizes LLM API costs and reduces inference latency, making it a valuable tool for developers looking to improve the performance and cost-effectiveness of their applications that rely on LLMs. Repository: vcache-project/vCache GitHub stars: 84 GitHub description: Reliable and Efficient Semantic Prompt Caching with vCache

Overview: A web interface for Jukebox, a music generation model by OpenAI, backed by Google Colab. Key features: - Jukebox Web UI - Google Colab integration - Python-based tech stack Tech stack: Python Why it matters: Provides a user-friendly interface for generating music using Jukebox, backed by Google Colab, making it accessible and easy to use for developers and enthusiasts alike. Repository: vzakharov/jukebox-webui GitHub stars: 84

### Overview: Weaviate is a vector search engine and database that uses machine learning to vectorize and store data. It allows for the deployment of custom ML models to production scale, making it an ideal tool for semantic search, complex document queries, and more. ### Key Features: - **Vector Search Engine**: Stores and retrieves data using vector embeddings. - **Custom ML Models**: Integrates with custom ML models for advanced search capabilities. - **Semantic Search**: Enables semantic search queries to provide more relevant results. - **Complex Document Queries**: Supports finding answers in complex documents using Weaviate's semantic search modules. - **Vector Database**: Stores and retrieves vector data for efficient querying. ### Tech Stack: Python - Weaviate is built using Python, making it easy to integrate with existing Python projects. ### Why It Matters: Weaviate offers a powerful and flexible solution for semantic search, complex document queries, and efficient vector data storage and retrieval. Its integration with custom ML models and advanced semantic search capabilities make it a valuable tool for developers and data scientists alike.

### Overview: HuBERT-ECG is a self-supervised foundation ECG model pre-trained on 9.1 million 12-lead ECGs from four countries and diverse patient populations. It has been evaluated through fine-tuning on 21 independent datasets spanning more than 1.6k diagnostic and prognostic targets across adults and pediatric cohorts, including single-lead settings. The model achieves AUROC ranging from 84% to 99% on ECG-primary diagnostic tasks, 76% to 97% on supportive diagnostic tasks, 74% to 91% on prognostic prediction tasks, and 88% to 92% on single-lead ECG benchmarks. Available in three model sizes to characterize scaling behavior and support diverse computational constraints, HuBERT-ECG is a valuable tool for broad and scalable cardiac applications. ### Key features: - Self-supervised foundation ECG model pre-trained on diverse ECG datasets. - Fine-tuned on 21 independent datasets across multiple tasks. - Achieves high AUROC scores across various diagnostic and prognostic tasks. - Available in three model sizes for different computational constraints. ### Tech stack: Shell ### Why it matters: HuBERT-ECG is a valuable tool for broad and scalable cardiac applications, providing high-quality diagnostic and prognostic information for various conditions, including single-lead ECGs. Its self-supervised pre-training and fine-tuning capabilities make it a robust and versatile model for clinical and research applications.

Overview: The Speech Recognition Benchmark is a program designed to evaluate and compare the performance of various automated speech recognition (ASR) APIs across different corpora. Key features: - Supports multiple ASR APIs including Amazon Lex, Google Cloud, Houndify, IBM Watson, Microsoft, and Speechmatics. - Runs on Mac OS X, Microsoft Windows, and Ubuntu. - Currently benchmarks five corpora: Common Voice, Fotolia, Image Edit Requests, LibriSpeech clean, and LibriSpeech other. - Provides benchmark results in word error rates (WER) expressed as percentages. Tech stack: Python Why it matters: The benchmark helps developers and researchers assess and compare the performance of different ASR APIs across various use cases, ensuring that the chosen technology meets the specific needs of their projects.

### Overview: GAOKAO-MM 是一个以中国高考题目为数据集的多模态评测基准，用于评估多模态模型的感知、理解、知识、推理等综合能力。 ### Key features: - **多模态数据集**：涵盖语文、数学、物理、化学、生物、政治、历史、地理8个科目以及示意图、折线图、地图、照片、几何图形等12种图像类型。 - **数据集示例**：提供646道选择题和897张图片，具体题目示例如下。 - **实验结果**：在10个多模态大模型上进行了实验，所有模型的正确率都低于50%。 ### Tech stack: Python ### Why it matters: GAOKAO-MM 为多模态模型的综合能力评测提供了重要的基准数据集，有助于推动多模态模型的发展和应用。

### Overview: SnapGenAI/Free-Nano-Banana-Pro-API-Ultimate-AI-Image-Generator is a free AI image generator API that offers a full-featured web app and a dedicated platform for seamless image editing and stunning AI visuals. The app utilizes Google's latest, fastest, and most efficient natively multimodal model, Nano Banana Pro, to create ultra-realistic images, edit photos, and generate stunning AI visuals in seconds. ### Key Features: - **Full-Featured Web App**: Seamless UI for creating and editing images with powerful AI tools. - **Text-to-Image Generation**: Generate high-quality images from simple text prompts with support for multiple styles. - **Style Transfer**: Instantly transform your images into different artistic styles (e.g., Anime, Ghibli, Realistic, Cinematic). - **Flexible Aspect Ratios**: Choose from presets like 1:1, 16:9, 3:4, or 9:16 for social media, presentations, or custom designs. - **Image Reference Support**: Upload an image as a reference to guide generation for more accurate and consistent results. - **Developer-Friendly API**: Integrate image generation into your apps easily with simple endpoints. ### Tech Stack: The app is built using Google's latest, fastest, and most efficient natively multimodal model, Nano Banana Pro. ### Why it Matters: This AI image generator platform is perfect for creators and developers who need to create stunning visuals quickly and efficiently. With its cutting-edge model and powerful features, it offers