Projects

Overview: AISIX is an open-source, Rust-based AI Gateway and LLM proxy designed for unified access to various OpenAI-compatible providers. Key features: - Multi-Provider Support: Supports OpenAI, Anthropic, Gemini, DeepSeek, and any OpenAI-compatible providers. - Traffic Management: Rate limiting, per-model and per-key access control, request validation with JSON Schema. - Security & Auth: API key authentication, per-key model allowlist, admin API key protection. - Observability: OpenTelemetry distributed tracing, Prometheus metrics export, structured logging via logforth. - Admin & Management: RESTful Admin API, React-based Admin Dashboard, Model CRUD, API Key CRUD, Chat Playground, etcd-backed dynamic configuration. Tech stack: Rust, Tokio, OpenAI-Compatible, etcd. Why it matters: AISIX simplifies the process of integrating multiple AI models into a single system, ensuring high performance and low resource usage. Its modular architecture and observability features make it an ideal choice for enterprise-level AI gateway solutions. Repository: api7/aisix-archived GitHub stars: 59 GitHub description: An open source, Native AI Gateway and LLM proxy built in Rust.

### Overview: Nano Banana (official) API (`gemini-2.5-flash-image-preview-official`) provides access to multiple image, video, and language models through the OpenAI-compatible APIMart endpoint. This API supports token billing and is designed for teams that need a single key to access a full catalog of models. ### Key Features: - **Model Id and Pricing**: The API supports multiple models, including Nano Banana (Gemini 2.5 Flash Image), with pricing based on token usage. - **Pricing Model**: The API charges per million tokens, with `text_input` and `image_input` priced at $1.60 per million tokens. - **Billing Model**: The billing model is token-based, and the charge is reflected in the response for completed tasks. - **OpenAI-Compatible Endpoint**: The API endpoint is compatible with the OpenAI API, allowing for seamless integration with other OpenAI models. ### Tech Stack: - **Python**: The primary programming language used to develop and interact with the API. ### Why It Matters: - **Token-Based Pricing**: The API charges based on token usage, making it cost-effective and easy to manage. - **OpenAI Compatibility**: The API is compatible with the OpenAI API, allowing seamless integration with other models. - **Token-Billed Route**: The API charges per million tokens, making it ideal for teams that need a single key to access a full catalog of models. ### GitHub Stars: - **Stars**: 58 ### Primary Language: - **Language**: Python ### GitHub Description: -

Overview: Treeffuser is an easy-to-use package designed for probabilistic prediction and regression on tabular data using tree-based diffusion models. It allows for the estimation of distributions of the form \( p(y|x) \) where \( x \) is a feature vector and \( y \) is a target vector, with the ability to model complex conditional distributions. Key features include: - Easy-to-use API adhering closely to scikit-learn's conventions. - Minimal user tuning required. - Supports complex conditional distributions like multimodal, heteroscedastic, non-Gaussian, and heavy-tailed models. Tech stack: This project is implemented using Jupyter Notebook. Why it matters: Treeffuser simplifies the process of probabilistic modeling on tabular data, making it accessible and efficient for practitioners without extensive statistical knowledge. Repository: The project is hosted on GitHub at https://github.com/blei-lab/treeffuser. GitHub stars: The repository has received 59 GitHub stars, indicating its popularity and usefulness among developers. Primary language: The project is written in Python, specifically using Jupyter Notebook, making it suitable for both research and production use. GitHub description: Treeffuser is a package for probabilistic prediction and regression on tabular data with tree-based diffusion models, designed to be easy to use and adaptable to complex conditional distributions.

### Overview: This project is a Minecraft bot built using the Agents for Amazon Bedrock framework, written in TypeScript. It demonstrates how to create an agent that can interact with the Minecraft world using the Return Control Agents. ### Key Features: - **Minecraft Integration**: The bot is designed to work seamlessly with the Minecraft game engine. - **TypeScript Implementation**: The project is written in TypeScript, making it easier to manage and maintain. - **Agents for Amazon Bedrock**: Leverages the Agents for Amazon Bedrock framework to interact with the Minecraft world. ### Tech Stack: - TypeScript - Amazon Bedrock ### Why It Matters: This project showcases the potential of using AI and automation in Minecraft, providing a practical application of the Agents for Amazon Bedrock framework. It is designed to be a starting point for developers interested in exploring the integration of AI and automation in the Minecraft ecosystem.

### Overview: OpenTihui is a Swift-based iOS app that runs on-device language models using the llama.cpp library. It supports both text and image input and can connect to OpenAI-compatible cloud APIs, making it a versatile tool for language processing. ### Key Features: - **Three Tabs**: Chat, Shortcuts, Settings. - **Local or Cloud Models**: Run a GGUF on-device or add an OpenAI-compatible API endpoint for chat. - **Multiple Conversations**: Saved to disk, switchable from the chat list, auto-titled from the first message, swipe to delete. Each chat keeps its own model, system prompt, config, and history. - **Shortcuts**: Reusable presets (Polite, Shorter, Fix Grammar, Reply, Translator, Summarizer, etc.). Tapping one starts a pre-configured chat; they also drive the keyboard's chips. - **Per-Chat & Default Settings**: Model, name, icon, context length, sampling, reasoning effort, system prompt, load-projector toggle, discard-context mode, auto-context (auto-fill from the clipboard on open), and image-compression size. - **Prompt Variables**: Reference variables in a system prompt as `$name` (highlighted in blue in the editor; tap one from the bar above the keyboard to insert it). Each variable's list of options is defined separately in Shortcut settings or Chat settings. ### Tech Stack: Swift ### Why it Matters: OpenTihui is a powerful tool for language processing on iOS, offering a native solution for on-device

### Overview: OBVI is a customizable web component designed for speech recognition in web-based projects, utilizing the Speech Recognition API and a fallback to Google Cloud Speech API for unsupported browsers. ### Key features: - Customizable UI for speech recognition - Supports both web and client-side speech recognition - Easy integration with Polymer 3+ projects - Fallback to Google Cloud Speech API for unsupported browsers ### Tech stack: Polymer 3+ ### Why it matters: OBVI simplifies integrating speech recognition into web projects, making it accessible and customizable for developers.

Overview: vLLM Router simplifies the management and usage of vLLM models by providing a single entry point for accessing multiple vLLM services. Key features: - Single entry point for accessing multiple vLLM services - Dynamic routing based on model name - Seamless integration with Kubernetes using Helm - Built using FastAPI and Uvicorn for high performance and scalability Tech stack: Python Why it matters: vLLM Router simplifies the deployment and management of vLLM models in applications by providing a unified interface, eliminating the need for separate API endpoints for each model. This makes it easier for developers to integrate multiple vLLM services into their applications without worrying about the underlying infrastructure.

### Overview: This repository contains the code and tutorial for inference-time alignment in protein diffusion models, focusing on optimization of multiple reward functions within a protein inverse folding model. The objective is to improve the stability and accuracy of protein design by leveraging inference-time techniques. ### Key Features: - **Inference-Time Techniques**: Utilizes various optimization techniques such as SMC, SVDD, NestedIS, and Classifier guidance to enhance the performance of protein design. - **Reward Functions**: Optimizes stability, prediction confidence (LDDT), and structural robustness (scRMSD) using specific metrics. - **Dataset and Validation**: The repository includes validation protein datasets and generates protein sequences using the diffusion model. ### Tech Stack: - Jupyter Notebook ### Why It Matters: Inference-time alignment significantly improves the stability and accuracy of protein design by optimizing multiple reward functions, making it a crucial technique in the field of protein structure prediction and design.

Overview: VectorLake is a S3 vector database designed for low maintenance, cost-effective storage, and efficient querying of high-dimensional vector data. It supports various features such as native Big Data support, vector data handling, efficient search, data persistence, customizable partitioning, and integration with LLM Agents. Key features: - Native Big Data Support: Ideal for handling large datasets. - Vector Data Handling: Stores and queries high-dimensional vectors. - Efficient Search: Supports nearest neighbor search for high-dimensional spaces. - Data Persistence: Stores data on disk, network volume, and S3 for long-term storage. - Customizable Partitioning: Minimizes database maintenance and cost with trade-offs. - Integration with LLM Agents: Enables seamless integration for language models. Tech stack: Python Why it matters: VectorLake simplifies the storage and querying of high-dimensional vector data, making it ideal for applications requiring efficient and cost-effective storage of large datasets, especially in the context of large language models and retrieval-based applications. Repository: msoedov/vector_lake GitHub stars: 59 GitHub description: S3 vector database for LLM Agents and RAG.

### Overview: This script is designed to generate a question-answer dataset from a given text, specifically from a PDF document. It uses the OLLAMA API to create questions and answers based on the text content. ### Key features: - Extracts text from a PDF document using PyPDF2. - Generates questions and answers using the OLLAMA API. - Saves the generated question-answer pairs in a JSON file. ### Tech stack: Python ### Why it matters: This script is a powerful tool for generating question-answer pairs based on a given text, making it useful for training and testing Language Learning Models (LLMs) in a Question-Answering (QA) context.

### Overview: The OMEGA Labs Bittensor Subnet is a pioneering project that aims to create the world's largest decentralized multimodal dataset, focusing on accelerating AGI research and development. This initiative leverages the Bittensor network and a community of miners and validators to build a dataset that is unparalleled in scale and diversity, encompassing 1 million hours of footage and 30 million video clips across 50+ scenarios and 15,000+ action phrases. ### Key Features: - **Unparalleled Scale and Diversity**: The dataset includes 1 million+ hours of footage and 30 million+ video clips, covering 50+ scenarios and 15,000+ action phrases. - **Latent Representations**: Utilizes state-of-the-art models to translate video components into a unified latent space, enabling efficient processing. - **Incentivized Data Collection**: Rewards miners for contributing high-quality, diverse, and novel videos through a decentralized network. - **Empowering Digital Agents**: Facilitates the development of intelligent agents that can navigate complex workflows. ### Tech Stack: Python ### Why it Matters: This initiative is crucial for accelerating AGI research and development by providing a comprehensive and diverse dataset that can be used to train powerful AGI models. The decentralized nature of the project ensures that the dataset remains accessible and transparent, democratizing access to cutting-edge research.

Overview: Samsar is a generative video and media automation platform that turns prompts, product images, structured content, and creative ideas into finished videos and reusable studio assets. Key features: - One-shot text-to-video up to 3 minutes - One-shot image-list-to-video - One-click scene re-rolls - Full Studio workspace for post-processing - Sandboxed local logging - Multi-provider model routing - Built-in search and recommendations Tech stack: JavaScript Why it matters: Samsar simplifies the process of turning creative ideas into finished videos and reusable studio assets, making it easier for developers to create engaging and polished content without the need for complex setup or extensive technical knowledge.

Overview: This advanced FLUX LoRA manipulation toolkit offers various techniques for manipulating and merging LoRAs, including difference operations, layer-based merging, and targeted layer targeting. It provides a user-friendly interface to help create compatible LoRAs and experiment with different combination approaches. Key features: - LoRA Difference: Removes conflicts between LoRAs. - Traditional Merging: Combines LoRAs with custom weights. - Layer-Based Merging: Surgically combines specific layers from different LoRAs. - Layer Targeting: Selectively mute facial or style layers. - LoRA MetaEditor: Direct metadata editing for any LoRA file. - LoRA MetaViewer: Examine LoRA characteristics, metadata, and compatibility. - LoRA MetaEditor: Direct metadata editing for fixing and customizing LoRAs. - Universal Compatibility: Works with LoRAs from any training tool (AI-Toolkit, FluxGym, sd-scripts). - Automatic Fixes: Auto-detects and resolves common compatibility issues. Tech stack: Python Why it matters: This toolkit is essential for developers and artists looking to create and experiment with LoRAs, enhancing their ability to manipulate and combine different styles and characters effectively. Repository: shootthesound/lora-the-explorer Primary language: Python GitHub stars: 59 GitHub description: Advanced FLUX LoRA manipulation toolkit with GUI interface README excerpt: This tool provides various FLUX LoRA manipulation techniques including difference operations, merging, targeted layer merging between LoRAs, layer targeting for zeroing and analysis. It's designed to help you

Overview: Novel Assistant는 로컬 또는 클라우드 LLM을 사용해 문서 기반으로 검색 및 답변하는 RAG 어시스턴트입니다. Key features: - 文本检索 (RAG 기능) - 文本生成 (AI 답변 생성) - 文本编辑 (Root Folder 기반) Tech stack: Python, OpenAI API, PyTorch, Flask, SQLite, HTML/CSS Why it matters: Novel Assistant는 문서 기반으로 질문에 대한 답변을 제공하는 데 있어 훌륭한 도우미입니다. 이 어시스턴트는 소설, 설정집, 캐릭터 시트, 세계관 노트 등 다양한 텍스트 파일을 활용하여 질문에 대한 정확하고 관련된 답변을 생성합니다. 이는 소설 작성, 설정 정리, 캐릭터 관리 등 다양한 작업을 돕는 데 큰 도움이 됩니다.

Overview: We propose Weak-to-Strong Diffusion (W2SD), a novel framework that utilizes the estimated difference between existing weak and strong models to approximate the gap between an ideal model and a strong model. By employing a reflective operation that alternates between denoising and inversion with weak-to-strong difference, W2SD steers latent variables along sampling trajectories toward regions of the real data distribution. Key features: - Utilizes the estimated difference between weak and strong models to approximate the gap between an ideal model and a strong model. - Employing a reflective operation that alternates between denoising and inversion with weak-to-strong difference. - Highly flexible and broadly applicable, enabling diverse improvements through the strategic selection of weak-to-strong model pairs. Tech stack: Jupyter Notebook Why it matters: W2SD offers a novel approach to refining variable latent representations, bringing them closer to the ground truth, and is highly flexible and broadly applicable, enabling diverse improvements through the strategic selection of weak-to-strong model pairs. Repository: xie-lab-ml/Weak-to-Strong-Diffusion-with-Reflection Primary language: Jupyter Notebook GitHub stars: 59 GitHub description: [ICLR2026] The official code of "Weak-to-Strong Diffusion with Reflection".

Overview: Diffusion Models as Data Mining Tools is a project that introduces a method to extract important patches from large datasets using diffusion models, aiming to enhance data mining capabilities. Key features: - Fine-tuning Stable-Diffusion v1.5 with custom prompts for specific labels. - Computing typicality scores for each image in a dataset. - Extracting top-1000 patches based on typicality scores. - Clustering patches using DIFT-161 features. Tech stack: Python Why it matters: This approach leverages diffusion models to efficiently mine important patches from large datasets, offering a powerful tool for data analysis and visualization.

### Overview: This repository contains the code for data processing used in the dataset and paper "Multimodal Data for the Detection of Freezing of Gait in Parkinson’s Disease" (Zhang et al., 2022). The dataset consists of multimodal data collected from 12 Parkinson's disease patients, including EEG, EMG, ECG, SC, and ACC signals. The data was collected using a high-quality hardware system and a standard experimental procedure to induce FOG in hospital surroundings. The FOG episodes were labeled by two qualified physicians, and the dataset can be used to efficiently discriminate FOG from normal walking. ### Key features: - Multimodal dataset of 12 Parkinson's disease patients. - Rich sensor data including EEG, EMG, ECG, SC, and ACC. - Data length of 3 hours and 42 minutes. - Labeled FOG episodes by two qualified physicians. - Code for data processing and analysis. ### Tech stack: Python ### Why it matters: This dataset is crucial for researchers and clinicians looking to improve the detection and prediction of Freezing of Gait (FOG) in Parkinson's disease. The multimodal approach enhances the accuracy and reliability of FOG detection, making it valuable for patient assessment and rehabilitation.

### Overview: DiET-GS is an innovative method for motion deblurring that leverages diffusion prior and event stream-assisted 3D Gaussian Splatting (DiET-GS). This project introduces a new approach to achieve high-quality novel-view synthesis from blurry multi-view images, preserving precise color and well-defined details. ### Key Features: - **DiET-GS**: A novel method for motion deblurring that uses diffusion prior and event stream-assisted 3D Gaussian Splatting. - **DiET-GS++**: An enhanced version of DiET-GS that includes per-scene optimization for better performance. - **3D Gaussian Splatting (3DGS)**: A technique that reconstructs 3D scenes from 2D images, providing realistic depth and texture. ### Tech Stack: - **Programming Language**: C++ - **Dependencies**: PyTorch, Lightning, Hugging Face Hub, Pillow, NumPy ### Why It Matters: DiET-GS offers a significant improvement in motion deblurring by combining diffusion prior and event stream-assisted 3D Gaussian Splatting, resulting in high-quality novel-view synthesis with precise color and details. The project's approach is particularly useful in scenarios where motion blur is severe, such as in real-world video processing and computer vision applications. ### Repository: - **Repository**: DiET-GS/DiET-GS - **GitHub Stars**: 58 - **GitHub Description**: Official code of "DiET-GS: Diffusion Prior and Event Stream-Assisted Motion Debl

### Overview: JODO is a research project that focuses on generating complete molecules using joint 2D and 3D diffusion models. The project aims to improve the quality and diversity of molecules generated through generative diffusion processes. ### Key Features: - **Joint 2D & 3D Diffusion Models**: Utilizes diffusion models to generate molecules by combining 2D and 3D representations. - **Complete Molecule Generation**: The project addresses the generation of complete molecules, including 3D structures and 2D representations. - **Visualization**: Provides visualizations of molecules generated on different datasets, including GEOM-Drugs and QM9. ### Tech Stack: - Python - PyTorch - PyG 2.1 ### Why It Matters: JODO is significant because it addresses the limitations of existing models in generating complete molecules, particularly in terms of 3D geometry and stability metrics. The project's approach aims to improve the quality and diversity of generated molecules, making it valuable for both research and practical applications in drug discovery and materials science. ### Repository: - Repository: GRAPH-0/JODO - GitHub Stars: 57 - GitHub Description: Learning Joint 2D & 3D Diffusion Models for Complete Molecule Generation ### Dependencies: - [pyTorch >= 1.11] - [PyG 2.1] - See requirements.txt for others. ### Dataset: - **GEOM-Drugs**: Provides a processed dataset for generating molecules. - **QM9**: Provides a dataset with explicit

### Overview: Big-Little Net is an efficient multi-scale feature representation model designed for visual and speech recognition. This repository contains the codes and models used in the paper "Big-Little Net: An Efficient Multi-Scale Feature Representation for Visual and Speech Recognition" by Chun-Fu (Richard) Chen, Quanfu Fan, Neil Mallinar, Tom Sercu, and Rogerio Feris. ### Key features: - **Models and Codes**: The repository includes the codes and models used in the paper. - **Paper Citation**: Please cite the paper if you use the codes and models from this repo. - **Training Script**: The training script is based on the imagenet example of PyTorch, with modifications. - **Usage Instructions**: - **Training**: Use the script to train a bL-ResNeXt-101 model with two GPUs. - **Testing**: Evaluate the model by downloading the pretrained models and running the evaluation script. ### Tech stack: Python ### Why it matters: This repository provides a practical implementation of the Big-Little Net model, demonstrating its effectiveness in visual and speech recognition tasks. The models and scripts are useful for researchers and developers looking to implement similar architectures.