Projects

### Overview: OpenConstructionEstimate is an open multilingual construction cost database designed for AI agents. It includes 55,000+ work items, 27,000+ resources, and 30+ regions, utilizing a semantic search engine via Qdrant vector database. ### Key Features: - **Semantic Search via Qdrant Vector DB:** Allows for efficient and accurate retrieval of relevant information. - **55,000+ Work Items:** Covers a wide range of construction work items, ensuring comprehensive coverage. - **27,000+ Resources:** Includes various resources such as materials, labor, and equipment. - **30+ Regions:** Supports construction projects across different regions, including North America, Europe, Asia, and more. ### Tech Stack: - **HTML** as the primary language. - **Qdrant** for semantic search. - **n8n** for creating workflows that calculate estimates based on descriptions, photos, and CAD (BIM) data. ### Why It Matters: - **Comprehensive Coverage:** Provides a wide range of work items and resources, making it useful for various construction projects. - **Efficient Search:** The semantic search engine helps in quickly finding relevant information, improving efficiency. - **Versatile Integration:** Supports multiple regions and project types, making it suitable for AI agents working across different locations and industries. ### GitHub Stars: - **GitHub Stars:** 258 - **Repository:** datadrivenconstruction/OpenConstructionEstimate-DDC-CWICR ### Languages:

### Overview: three.ws is an open-source platform for creating 3D AI agents using text or photos. It allows users to turn prompts into textured 3D models, register on-chain, and embed them anywhere with minimal setup. ### Key features: - **Text-to-3D Model**: Generate a 3D model from text prompts. - **Image-to-3D Model**: Convert existing GLB models into 3D. - **LLM Integration**: Register an LLM brain and memory for the 3D avatar. - **Embed Anywhere**: Embed the 3D avatar on any web component without plugins or server uploads. ### Tech stack: JavaScript, Three.js, x402 payments, and a remote MCP server. ### Why it matters: three.ws simplifies the process of creating 3D AI agents, making it accessible to developers without requiring plugins, server uploads, or installation. It provides a straightforward way to integrate AI into web applications, enhancing user experience with interactive 3D models.

Overview: Victor is a Rust-based web-optimized vector database designed for efficient storage and retrieval of high-dimensional data. Key features: - Rust API for native filesystem or transient in-memory filesystem - Web API for accessing the database through the Private Origin File System - Efficient vector storage format with minimal representation size - PCA for vector compression when storage space is limited Tech stack: Rust Why it matters: Victor provides a fast and efficient way to store and retrieve high-dimensional data, making it suitable for applications requiring real-time performance and minimal storage footprint. Repository: not-pizza/victor Primary language: Rust GitHub stars: 262 GitHub description: Web-optimized vector database (written in Rust).

Overview: BigCiDian is a pronunciation lexicon project aimed at creating a unified phoneset for both English and Chinese languages for Automatic Speech Recognition (ASR) applications. This project aims to simplify and precisely map English and Chinese phonemes into a common phoneset. Key features: - A unified phoneset covering both English and Chinese languages. - Mapping from ARPA phoneset to target phoneset for English entries. - PinYin to IPA mapping for Chinese entries. - Tone retention in Chinese phonemes. - A unified phoneset with potential for 6 tonal variations. Tech stack: Python Why it matters: This project is crucial for enhancing the accuracy and efficiency of ASR systems by providing a standardized pronunciation reference for both languages, which is essential for improving the overall quality of speech recognition applications.

Overview: This repository provides a Terraform configuration for setting up Google Cloud resources and deploying Dify in a highly available and scalable manner. Key features: - Serverless hosting - Auto-scaling - Data persistence Tech stack: HCL Why it matters: This Terraform configuration ensures that Dify is deployed on Google Cloud with high availability, scalability, and production-level readiness, making it suitable for production environments.

### Overview: DiffusedTexture is a free, open-source Blender add-on that generates textures directly on your 3D mesh using Stable Diffusion 1.5 or SDXL. It ensures the generated textures align with the geometry and supports various refinements like LoRA and IP-Adapter. ### Key Features: - **Stable Diffusion 1.5/SDXL Integration**: Directly generates textures on your mesh using high-resolution images. - **ControlNets**: Aligns the generated textures with the geometry, ensuring high detail and realism. - **LoRA and IP-Adapter Support**: Enhances the quality of generated textures with advanced techniques. - **Mesh Complexity Support**: Works with meshes of up to 50,000 triangles. - **Texture Resolution**: Supports resolutions up to 2048x2048. - **Render Resolution**: Renders at 4096x4096. - **Custom SD Resolution**: Supports Custom SD Resolution settings for better texture quality. - **4 Camera Views**: Generates multiple views for better texture coverage. - **Prompt Customization**: Allows customization of prompts to match specific models and textures. ### Tech Stack: - Python ### Why It Matters: DiffusedTexture simplifies the process of generating high-quality textures directly in Blender, making it easier for artists and designers to create detailed and realistic textures without the need for specialized software or cloud services. The use of Stable Diffusion and ControlNets ensures that the generated textures are highly aligned with the geometry, resulting in a seamless integration into Blender projects.

Overview: docGPT is a free document chatbot that allows users to interact with their documents in various formats (PDF, Word, CSV, TXT) without needing API keys or fees. Key features: - Supports docx, pdf, csv, and txt files - Integrates with LangChain for enhanced functionality - Directly inputs document URLs for parsing - Enables LangChain to answer current questions and provide Google-like search results Tech stack: Python Why it matters: This project offers a user-friendly interface for querying documents and integrating with LangChain for advanced AI capabilities, making it a valuable tool for developers and content creators. Repository: Lin-jun-xiang/docGPT-langchain Primary language: Python GitHub stars: 261 GitHub description: 🔐Free GPT-3.5 chat with your docs (PDF, Word, CSV, TXT)

### Overview: ConsistI2V is a diffusion-based method designed to enhance visual consistency for image-to-video generation. The project is part of the TMLR-2024 paper titled "[ConsistI2V: Enhancing Visual Consistency for Image-to-Video Generation](https://arxiv.org/abs/2402.04324)." The codebase aims to improve the visual consistency of videos generated from images, addressing the need for more consistent and realistic video outputs. ### Key Features: - **Spatiotemporal Attention**: Maintains both spatial and motion consistency by introducing spatiotemporal attention over the first frame. - **Noise Initialization**: Uses noise from the low-frequency band of the first frame to enhance layout consistency, improving overall visual consistency. ### Tech Stack: - Python ### Why It Matters: ConsistI2V is a significant advancement in the field of image-to-video generation, offering a method to produce videos that are more consistent and visually appealing. The project's contributions are particularly valuable for applications requiring high-quality video outputs, such as simulations, animations, and realistic visual effects. ### Repository: - **Repository**: TIGER-AI-Lab/ConsistI2V - **GitHub Stars**: 261 - **GitHub Description**: This repo is under construction. Please stay tuned. - **Primary Language**: Python ### Environment Setup: - **Prepare Codebase and Conda Environment**: ```bash git clone https://github.com/TIGER-AI-Lab/

### Overview: Swap-Anything is an innovative model designed for personalized visual editing, enabling users to swap objects in any size and shape within an image. This project addresses the limitations of existing object swapping techniques by providing a flexible and powerful solution that can handle arbitrary object swapping, general object swapping, and object insertion with high accuracy. ### Key Features: - **Personalized Object Swapping**: Swap-Anything allows users to swap objects in their personal photos, preserving the background perfectly. - **General Object Swapping**: Supports swapping objects of any size and shape in any image size. - **Object Insertion**: Enables the insertion of new objects into an image without altering the background. ### Tech Stack: - **Python**: The primary programming language used for developing the model and its functionalities. ### Why It Matters: Swap-Anything offers a significant improvement in the field of visual editing by providing a versatile solution that can handle various types of object swaps and insertions, making it highly valuable for both personal and professional use cases.

Overview: This project focuses on enhancing the conversational capabilities of the Mixtral-8x7B sparse Mixture-of-Experts model through instruction tuning. The team, including Rongsheng Wang, has successfully created the Aurora model by integrating three Chinese instruction-following datasets. Key features: - Aurora is a Mixtral-8x7B sparse Mixture-of-Experts model. - It is designed to improve the Chinese conversational abilities of the model. - The project involves instruction fine-tuning on carefully processed datasets. Tech stack: Python Why it matters: This work is pioneering in the execution of instruction fine-tuning on a sparse expert-mixed model, marking a significant advancement in enhancing the capabilities of this model architecture. Repository: WangRongsheng/Aurora Primary language: Python GitHub stars: 261 GitHub description: The official codes for "Aurora: Activating chinese chat capability for Mixtral-8x7B sparse Mixture-of-Experts through Instruction-Tuning" README excerpt: 

Overview: LangFair is a Python library designed for conducting use-case level LLM bias and fairness assessments. It provides comprehensive support for choosing bias and fairness metrics, offering illustrative examples, and includes a documentation site, tutorials, and research papers. Key features: - Comprehensive framework for choosing bias and fairness metrics - Documentation site with API reference - Illustrative example notebooks - Tutorial on how to assess LLM use cases for bias and fairness - Software and research paper on evaluating LLMs Tech stack: Python Why it matters: LangFair addresses the gap in static benchmark assessments by adopting a Bring Your Own Prompts (BYOP) approach, ensuring that metrics reflect real-world performance and risk assessment in LLM use cases. Repository: cvs-health/langfair GitHub stars: 261 GitHub description: LangFair is a Python library for conducting bias and fairness assessments of large language model (LLM) use cases.

Overview: The AI SDK for Go provides a unified interface for interacting with multiple AI providers, including OpenAI, Anthropic, and more. It aims to simplify the development of AI applications and agents by offering a developer-friendly experience with idiomatic Go patterns and error handling. Key features: - Multi-Provider Support: Supports multiple AI providers including OpenAI and Anthropic. - Multi-Modal Inputs: Handles text, images, and files in conversations. - Tool Calling: Allows function calling with parallel execution. - Language Models: Supports text generation with streaming and non-streaming capabilities. - Embedding Models: Generates text embeddings for semantic search. - Image Models: Generates images from text prompts. - Structured Outputs: Supports JSON generation with schema validation. Tech stack: The AI SDK for Go is built using Go, a programming language known for its simplicity and efficiency. Why it matters: The AI SDK for Go simplifies the process of integrating AI into Go-based applications, reducing the need for developers to rewrite code across different providers, and improving the developer experience by providing a Go-first design and comprehensive error handling. Repository: The repository is maintained by Jetify, the creators of TestPilot, and is currently in Public Alpha, with plans for General Availability (v1) in the future. GitHub stars: The repository has received 261 GitHub stars, indicating a growing interest in the project. Primary language: The AI SDK for Go is written in Go, the programming language of choice for building scalable and efficient systems. GitHub description: The AI SDK for Go is a unified interface for interacting

Overview: WordLink is an AI-powered vocabulary learning tool that enhances vocabulary retention by showing words in their context, including synonyms, near-synonyms, antonyms, and related words. It uses a fine-tuned model to provide a comprehensive vocabulary workspace. Key features: - Word graph: Displays the selected word with relation types, word families, and progress indicators. - Around the graph: Includes quizzes, graded reading, ambient mode, and a learning passport. - Quick start: Requires Node.js 20+ and Docker, with migrations and dictionaries loaded. Tech stack: TypeScript, Docker, PostgreSQL, Prisma ORM, Next.js, React, ModelScope weights. Why it matters: By providing context and related words, WordLink helps improve vocabulary retention and understanding, making it a valuable tool for language learners. GitHub stars: 227, Live demo: [wordlink.lifeplayertribe.com](https://wordlink.lifeplayertribe.com) Repository: orionsheep/wordlink

### Overview: This repository contains a collection of datasets designed for the instruction tuning of large language models. The datasets are categorized into two main types: gold standard datasets and silver standard datasets generated using large language models. The gold standard datasets include diverse NLP tasks and expert-written instructions, while the silver standard datasets are generated using pre-trained models. These datasets are crucial for evaluating and improving the performance of large language models in various NLP tasks. ### Key Features: - **Gold Standard Datasets**: - P3: Promoted English datasets covering a diverse set of NLP tasks. - xP3: Mixture of 13 training tasks in 46 languages with prompts in 20 languages. - Natural Instructions v2: A benchmark of 1,616 diverse NLP tasks and their expert-written instructions. - The Flan Collection: Superset of some of the datasets. - Open Assistant: Human-annotated assistant-style conversation corpus. - LIMA: 1K high-quality instructions. - databricks-dolly-15k: A large-scale dataset for Android device control. - PRESTO: 550K contextual multilingual conversations. - BB3x: A large-scale dataset for Android device control. - InstructCTG: A framework for controlled generation. - CrossFit: A large-scale dataset for contextual multilingual conversations. - tasksource: A benchmark of 1,616 diverse NLP tasks. - ExMix: A benchmark of

### Overview: Awesome-VLA-Post-Training is a comprehensive collection of cutting-edge resources focused on the post-training of Vision-Language-Action (VLA) systems. This repository serves as a centralized hub for research updates, practical codes, and implementation insights, aiming to enhance the ability of VLA agents to perceive, reason, and act within physical environments. ### Key Features: - **Enhancing Environmental Perception**: Techniques to improve the model's ability to understand and interpret visual and textual data in complex environments. - **Improving Embodiment Awareness**: Methods to enhance the model's understanding of its physical surroundings and interactions with the environment. - **Deepening Task Comprehension and Generalization**: Approaches to improve the model's ability to generalize tasks and perform well across different scenarios. - **Integrating and Tuning Multiple Components**: Tools and resources to integrate and optimize the performance of various components within VLA models. ### Tech Stack: Python, TensorFlow, PyTorch, and other relevant libraries. ### Why it Matters: This repository is essential for researchers and practitioners interested in advancing the field of VLA systems. It provides a structured and high-quality resource for the community, facilitating collaboration and knowledge sharing.

### Overview: This project presents FMBoost, a method that leverages the strengths of Diffusion Models (DMs), Flow Matching models (FMs), and Variational AutoEncoders (VAEs) to generate high-resolution images at a significantly reduced computational cost. The approach combines the diversity of stochastic DMs, the speed of FMs in training and inference, and the efficiency of a convolutional decoder to map latents into pixel space, resulting in a small diffusion model that excels in generating diverse samples at a low resolution. Flow Matching then boosts LCM-SD1.5 with a general Coupling Flow Matching (CFM) model, achieving exceptional speed and quality in image synthesis. ### Key features: - **FMBoost**: A method that combines the strengths of DMs, FMs, and VAEs to generate high-resolution images. - **Latent Consistency Models (LCM)**: Distilled from SD1.5 and SDXL, LCM-SD1.5 is used as the low-resolution representation. - **Coupling Flow Matching (CFM)**: Boosts LCM-SD1.5 with CFM to achieve higher resolution. - **Training Pipeline**: Feeds both low- and high-resolution images through the pre-trained encoder to obtain low- and high-resolution latent codes. - **Inference Pipeline**: Takes any diffusion model, generates the low-res latent, and uses CFM to synthesize the higher-dimensional latent code, then projects it back to pixel space. - **Quantitative Comparison**: Shows zero-shot quantitative comparison against other state

Overview: TSDiff is an unconditional diffusion model designed for probabilistic time series forecasting, presented in the NeurIPS 2023 paper "Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting". It enables conditional inference during training and inference, improving prediction quality and enabling the synthesis of realistic time series. Key features: - Conditional inference during training and inference. - Implicit density refinement of base forecasters. - Generation of realistic synthetic time series for downstream tasks. Tech stack: Python Why it matters: TSDiff offers a novel approach to time series forecasting by leveraging diffusion models, enabling conditional inference and improving prediction accuracy without the need for auxiliary networks or altering the training procedure.

Overview: This repository contains a collection of research papers on the application of diffusion models in weather forecasting, focusing on high-resolution solar forecasts, precipitation nowcasting, and ensemble forecast emulation. Key features: - Papers on diffusion models for solar forecasts - Papers on diffusion models for precipitation nowcasting with accurate uncertainty quantification - Papers on conditional emulation of Earth system models with diffusion models - Papers on ensemble forecast emulation using diffusion models - Papers on a dynamics-informed diffusion model for spatiotemporal forecasting - Papers on integrating SwinRNN with diffusion models for high-resolution weather forecasting - Papers on residual diffusion modeling for atmospheric downscaling - Papers on improving extreme rainfall and wind forecasts with diffusion models - Papers on probabilistic precipitation downscaling with optical flow-guided diffusion Tech stack: The repository uses Markdown for documentation and includes links to papers and code repositories. Why it matters: This collection is valuable for researchers and developers interested in advancing weather forecasting models using diffusion techniques, which can lead to more accurate and reliable weather predictions.

### Overview: BLIVA is a simple multimodal language model designed to handle text-rich visual questions, achieving top performance on benchmark tasks. ### Key features: - **Text-rich Visual Question Answering (VQA)**: BLIVA excels in answering questions that involve text-rich visual information. - **Training Code**: Available for researchers to access and experiment with. - **Benchmark Performance**: Achieved No.3 in Perception tasks and No.2 in Cognition tasks on the MME benchmark. - **Model Versions**: Available for commercial use with weights for BLIVA Vicuna and FLAN T5. - **Datasets**: YTTB-VQA dataset is publicly available. ### Tech stack: Python ### Why it matters: BLIVA demonstrates the effectiveness of simple multimodal language models in handling complex visual questions, showcasing advancements in multimodal AI.

### Overview: Ragpi is an open-source AI assistant designed to answer questions using your documentation, GitHub issues, and READMEs. It leverages LLMs and intelligent search to provide relevant, documentation-backed answers through an API. ### Key features: - 📚 Builds knowledge bases from docs, GitHub issues, and READMEs - 🤖 Agentic RAG system for dynamic document retrieval - 🔌 Supports OpenAI, Ollama, and Deepseek models - 💬 Integrations with Discord and Slack for community support - 🚀 API-first design with Docker deployment ### Tech stack: Python ### Why it matters: Ragpi simplifies the process of integrating AI into projects by providing a straightforward API and documentation, making it easy to integrate and use in various environments.