Projects

Overview: This repository contains code and pre-trained models for the Princeton University Relation Extraction system (PURE), described in the paper "A Frustratingly Easy Approach for Entity and Relation Extraction" (arXiv:2010.12812). Key features: - Entity model - Relation model - Approximation relation model Tech stack: PyTorch Why it matters: PURE offers a simple yet effective solution for entity and relation extraction, making it a valuable tool for researchers and developers working on NLP tasks. GitHub stars: 815 Repository: princeton-nlp/PURE

### Overview: Verbalized Sampling is a training-free prompting strategy that enhances LLM diversity by 2-3x. It works by requesting responses with probabilities, ensuring a broader range of outputs. Versatile across various tasks, it is model-agnostic and orthogonal to temperature. ### Key Features: - **Training-Free**: Works with any LLM via prompting. - **Model-Agnostic**: Compatible with GPT, Claude, Gemini, Llama, etc. - **Orthogonal to Temperature**: Effective without changing the model's temperature. - **Improves Diversity**: Achieves 2-3x diversity improvement while maintaining quality. ### Tech Stack: Python ### Why It Matters: Verbalized Sampling enhances LLM diversity, making it more versatile and adaptable across different tasks. It is particularly useful in creative writing, social simulation, synthetic data generation, and open-ended QA, offering a more comprehensive approach to AI language models. ### Repository: CHATS-lab/verbalized-sampling - **Primary Language**: Python - **GitHub Stars**: 814 - **Homepage**: https://www.verbalized-sampling.com/ - **Paper**: https://arxiv.org/abs/2510.01171 - **Blog**: https://simonucl.notion.site/verbalized-sampling - **HuggingFace**: https://huggingface.co/collections/CHATS-Lab/verbalized-sampling - **Examples**: https://tinyurl.com/vs-gallery - **Podcast

Overview: LongSplat is an innovative framework designed for robust 3D Gaussian Splatting in casual long videos, achieving high-quality novel view synthesis while maintaining efficiency and scalability. Key features: - Incremental joint optimization - Pose estimation with MASt3R - Adaptive octree anchoring Tech stack: Python Why it matters: LongSplat addresses the challenge of robust 3D Gaussian Splatting in long videos, providing a solution that can handle free viewpoints and remains efficient and scalable, making it suitable for various applications in computer vision and video processing. Repository: NVlabs/LongSplat Primary language: Python GitHub stars: 813 GitHub description: [ICCV 2025] LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos README excerpt: [ICCV 2025] LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos

### Overview: BERT-keras is a Keras implementation of Google BERT and OpenAI's Transformer Language Model, featuring pre-trained weights. It includes a finetuning API, making it suitable for various NLP tasks such as sentiment analysis, is_next, and POS tagging. ### Key Features: - **BERT Pre-trained Model**: Uses pre-trained weights from Google BERT and OpenAI's Transformer. - **Finetuning API**: Allows for fine-tuning of the model for specific tasks. - **Sentence and Token-Level Tasks**: Supports both sentence-level tasks (like sentiment analysis) and token-level tasks (like POS tagging). - **Data Generation and Task Meta Data**: Provides a data generator and task meta data for most NLP tasks, compatible with other frameworks. ### Tech Stack: - **Python**: The primary language used for development. - **Keras**: Framework used for building and training the model. ### Why It Matters: - **Compatibility with Official Models**: Supports both Google BERT and OpenAI's Transformer models, making it a versatile tool for NLP tasks. - **Framework Independence**: Abstraction over token-level and sentence-level tasks, allowing for better integration with different frameworks. - **Ease of Use**: Provides a simple API for fine-tuning and training, making it accessible to developers. ### GitHub Stars: - **GitHub Stars**: 813 ### Repository: - **Repository**: Separius/BERT-keras - **Primary Language**: Python - **GitHub Description**: Keras implementation of BERT with pre-trained weights - **GitHub

### Overview: This repository contains a collection of deep learning models for text classification implemented using the Keras framework. The models include FastText, TextCNN, TextRNN, TextBiRNN, TextAttBiRNN, HAN, RCNN, and RCNNVariant. These models are designed to be efficient and effective for various text classification tasks. ### Key features: - **FastText**: Uses a look-up table to convert bags of n-grams into word representations, which are then averaged into a text representation fed to a linear classifier. - **TextCNN**: Represents sentences with static and non-static channels, convolves with multiple filter widths and feature maps, and uses max-over-time pooling followed by a fully connected layer with dropout and softmax output. - **TextRNN**: Represents sentences with a static and non-static channel, convolves with multiple filter widths and feature maps, and uses max-over-time pooling followed by a fully connected layer with dropout and softmax output. - **TextBiRNN**: An improved model based on TextRNN, which uses a bidirectional RNN layer to consider both forward and backward encoding information. - **TextAttBiRNN**: An improved model that introduces attention mechanism to focus on the most relevant information for decision-making, based on the bidirectional RNN encoder. ### Tech stack: Keras ### Why it matters: These models are designed to be efficient and effective for various text classification tasks, making them suitable for a wide range of applications in natural language processing and text analysis.

Overview: Robby-chatbot is an AI chatbot designed to interact with CSV, PDF, TXT files, and YouTube videos using Langchain, OpenAI, and Streamlit. It aims to make AI more accessible and intuitive for geospatial applications. Key features: - Robby-Chat: Conversational memory for discussing documents. - Robby-Sheet: Analyzes tabular data with natural language. - Robby-Youtube: Summarizes YouTube videos using AI. Tech stack: LangChain, OpenAI, PandasAI, Streamlit, FAISS. Why it matters: This project aims to democratize AI by making it easier to use for geospatial applications, building TerraLab with the author's dad and best friend.

### Overview: This repository is the official code repository for the O'Reilly Publication, "Practical Deep Learning for Cloud, Mobile, and Edge," authored by Anirudh Koul, Siddha Ganju, and Meher Kasam. It provides step-by-step guidance on building practical deep learning applications for cloud, mobile, and edge devices using a hands-on approach. ### Key Features: - **Hands-On Projects:** Implementing real-world applications using Keras, TensorFlow, Core ML, and TensorFlow Lite. - **Industry Case Studies:** Exploring projects from Silicon Valley to 40+ industry case studies. - **Mobile Development:** Developing AI for devices like Raspberry Pi, Jetson Nano, and Google Coral. - **Transfer Learning:** Training models in minutes using transfer learning. - **Autonomous Driving:** Simulating an autonomous car in a video game environment and building a miniature version with reinforcement learning. - **Practical Tips:** 50+ tips for maximizing model accuracy, debugging, and scaling to millions of users. ### Tech Stack: Jupyter Notebook ### Why it Matters: Whether you're a software engineer, data scientist, or hobbyist, this book provides a comprehensive guide to building practical deep learning applications for cloud, mobile, and edge devices. It includes step-by-step projects and real-world examples, making it a valuable resource for anyone looking to enter the field of deep learning.

Overview: TF-ICON is a novel Training-Free Image Composition framework that leverages text-driven diffusion models for seamless cross-domain image-guided composition. Key features: - Cross-domain image composition - Training-Free approach - Exceptional prompt for accurate inversion of real images into latent representations Tech stack: Python Why it matters: TF-ICON offers a practical solution for cross-domain image composition by avoiding the need for additional training, finetuning, or optimization, making it accessible and efficient for various applications.

### Overview: TensorRT C++ API is a modern, no-throw C++ library designed for high-performance GPU inference of CNN models with NVIDIA TensorRT. It simplifies the process by providing a small, leak-free API that supports various precision modes and dynamic shapes. The library is built to TensorRT version 11 and targets CUDA 12, C++20, and Linux. ### Key Features: - **Engine Cache Safety:** The library ensures that the engine cache is safe and up-to-date by using ONNX content hash, build options, and GPU UUIDs. - **Dynamic Shapes and Optimization Profiles:** It supports per-input min/opt/max optimization profiles and one optimization profile per execution context for concurrent dynamic-shape inference. - **Concurrency and Thread Safety:** The engine is thread-compatible and the pool is thread-safe, making it suitable for multi-threaded environments. - **No Leaky Abstractions:** The library does not rely on any external libraries such as OpenCV or TensorRT types, ensuring a clean and straightforward API. - **Quantization Support:** It supports various precision modes like FP16, Int8Qdq, and FP8, with version-aware precision settings. - **Optional Preprocessing:** It includes an optional fused preprocessing kernel that can be used to simplify the preprocessing steps. - **Zero-Copy Python Bindings:** It provides optional zero-copy Python bindings for easier integration with other libraries like CuPy, PyTorch, and Numba. ### Performance: - **Single-Stream Inference Latency:** - **YOLOv8n

Overview: This is a real-time 3D face tracking and reconstruction project that uses 2D video to track and reconstruct a 3D face model. Key features: - Real-time 3D face tracking - Reconstruction from 2D video - Integration with EOS library for 3D face model - Support for live webcams and videos Tech stack: C++ Why it matters: This project is a research demo from 2015/2016, but it demonstrates real-time 3D face tracking and reconstruction from 2D video, which is still relevant for applications like augmented reality and virtual reality.

Overview: This project provides PyTorch implementations for the CVPR 2019 paper "APDrawingGAN: Generating Artistic Portrait Drawings from Face Photos with Hierarchical GANs". It generates artistic portrait drawings from face photos using a GAN-based model. Key features: - PyTorch implementations - Preprocessing steps - Training/testing tips Tech stack: Python Why it matters: This project offers a practical solution for generating artistic portrait drawings from face photos using advanced GAN techniques, making it useful for various applications in the field of computer vision and image synthesis. Repository: yiranran/APDrawingGAN GitHub stars: 811 GitHub description: Code for APDrawingGAN: Generating Artistic Portrait Drawings from Face Photos with Hierarchical GANs (CVPR 2019 Oral) README excerpt: # APDrawingGAN We provide PyTorch implementations for our CVPR 2019 paper "APDrawingGAN: Generating Artistic Portrait Drawings from Face Photos with Hierarchical GANs". This project generates artistic portrait drawings from face photos using a GAN-based model. You may find useful information in [preprocessing steps](preprocess/readme.md) and [training/testing tips](docs/tips.md). [[Paper]](http://openaccess.thecvf.com/content_CVPR_2019/html/Yi_APDrawingGAN_Generating_Artistic_Portrait_Drawings_From_Face_Photos_With_Hierarchical_CVPR_2019_paper.html) [[Demo]](https

Overview: Paint3D is a novel generative framework designed to produce high-resolution, lighting-less, and diverse 2K UV texture maps for untextured 3D meshes. The method addresses the challenge of generating textures without embedded illumination information, enabling re-lighting and re-editing within modern graphics pipelines. Key features: - Coarse-to-fine generative framework - Produces high-quality 2K UV textures - Lighting-less and diverse textures - High-resolution output - Supports text or image inputs Tech stack: Python Why it matters: Paint3D significantly advances the state-of-the-art in texturing 3D objects by enabling the creation of lighting-less and diverse textures, which are essential for applications requiring flexibility and re-editing capabilities. Repository: OpenTexture/Paint3D GitHub stars: 810 GitHub description: [CVPR 2024] Paint3D: Paint Anything 3D with Lighting-Less Texture Diffusion Models, a no lighting baked texture generative model

Overview: Learn vector search with Rust and DataFusion; the C++/BusTub track is deprecated. Key features: - Build a small in-memory vector database in Rust - Compare approximate results with exact search - Use indexes from SQL through DataFusion - Create indexes and compare query plans - Benchmark recall, build time, and search latency on SIFT1M Tech stack: Rust Why it matters: The course focuses on the boundary where vector-search algorithms become database features, covering the algorithms, evaluation contracts, query planning, and execution path instead of hiding them behind an HTTP API or ANN library. Repository: skyzh/vector-db-from-scratch Primary language: Rust GitHub stars: 808 GitHub description: Learn vector search with Rust and DataFusion; the C++/BusTub track is deprecated.

### Overview: The iNaturalist Competition datasets provide a rich resource for developers interested in environmental data analysis and machine learning. ### Key features: - Diverse datasets from the iNaturalist project, including photos, observations, and species information. - Comprehensive metadata and annotations for each dataset. - Support for various programming languages, including Python, to facilitate integration and analysis. ### Tech stack: Python ### Why it matters: These datasets offer a valuable resource for developers looking to enhance their environmental data analysis capabilities, particularly in areas such as image recognition and species identification.

Overview: Debug your AI agents in real-time with vLLora, a lightweight, real-time debugging tool that seamlessly integrates with popular AI frameworks like LangChain, Google ADK, and OpenAI. Key features: - Real-time tracing of AI agent interactions - Support for Model Context Protocol (MCP) servers for seamless integration with external tools - Seamless integration with OpenAI-compatible chat completions API Tech stack: Rust Why it matters: vLLora provides developers with a powerful and easy-to-use debugging tool that allows them to monitor and optimize AI agent interactions in real-time, ensuring that their agents are performing as expected and identifying potential issues quickly. Repository: vllora/vllora GitHub stars: 810 GitHub description: Debug your AI agents README excerpt: <div align="center"> <img src="assets/images/logos/logo_dark.svg" width="200px" alt="vLLora Logo"> #### Lightweight, Real-time Debugging for AI Agents Debug your Agents in Real Time. Trace, analyze, and optimize instantly. Seamless with LangChain, Google ADK, OpenAI, and all major frameworks. **[Documentation](https://vllora.dev/docs)** | **[Issues](https://github.com/vllora/vllora/issues)** </div> ## Quick Start First, install [Homebrew](https://brew.sh) if you haven't already, then: ```bash brew tap vllora/vllora brew install vllora ``` ### Start the vLLora: ```bash

### Overview: This repository offers a collection of AI visual prompt styles for various applications, including Nano Banana Pro, GPT Image, and Midjourney. By copying a single `style.json`, users can achieve consistent and reusable visual styles across different AI image generation workflows. ### Key Features: - **Consistent Visuals:** Each style is designed to be easily integrated into existing workflows, ensuring a consistent look across different platforms. - **Structured JSON:** The styles are provided as `style.json` files, making them easy to copy and paste into any LLM-based image generation tool. - **New Styles Weekly:** The project aims to provide a new style every week, keeping the library fresh and relevant. ### Tech Stack: HTML ### Why It Matters: This repository simplifies the process of creating visually consistent styles for AI image generation, making it easier to iterate and compare different styles. By providing a structured JSON format, users can quickly integrate styles into their workflows, enhancing the overall quality and coherence of their visual content.

Overview: MiNLP is a comprehensive toolkit for natural language processing developed by Xiaomi, featuring modules for tokenization, part-of-speech tagging, named entity recognition, and dependency parsing. Key features: - MiNLP-Tokenizer: A Chinese tokenizer that has been optimized and refined through continuous testing and practical use. - Word Sense Disambiguation (WSD): A tool for resolving word meanings in context. - Dependency Parsing: An advanced dependency parsing tool that can analyze the structure of sentences and convert them into structured objects. Tech stack: Scala Why it matters: MiNLP is essential for any developer looking to enhance their NLP capabilities, making it easier to process and analyze text data effectively. Repository: XiaoMi/MiNLP Primary language: Scala GitHub stars: 809 GitHub description: XiaoMi Natural Language Processing Toolkits

### Overview: labelCloud is a lightweight tool designed for labeling 3D bounding boxes in point clouds. It simplifies the process of annotating point clouds, making it easier for developers to integrate and utilize. ### Key Features: - **Labeling Modes:** Supports both picking and spanning modes for easy and precise bounding box labeling. - **Mouse and Keyboard Options:** Provides intuitive mouse and keyboard controls for quick corrections. - **Configuration:** Configurable through `config.ini` or settings to meet specific needs. - **Semantic Segmentation:** Enables the creation of segmentation labels based on bounding boxes, useful for semantic segmentation tasks. ### Tech Stack: Python ### Why It Matters: labelCloud is an essential tool for developers working with point cloud data, providing a straightforward and efficient way to label and segment 3D bounding boxes. Its lightweight nature and flexible configuration make it a valuable asset in various applications, including computer vision and robotics. Repository: ch-sa/labelCloud GitHub Stars: 809

### Overview: Lobe TTS is a high-quality and reliable Text-to-Speech (TTS) and Speech-to-Text (STT) library designed for both server and browser environments. It simplifies the process of integrating voice features into applications, offering smooth long-form playback, AI-powered development, and seamless integration with various voice providers. ### Key Features: - **Server or Browser**: Generate audio files in Node.js or use the browser's built-in speech capabilities. - **AI-Powered Development**: Use Microsoft Edge, Microsoft Speech, OpenAI, or the browser's built-in speech, all through the same simple API. - **React Ready**: Drop-in hooks and an audio player with progress, pause, and download features, making it easy to integrate without building playback yourself. - **Smooth Long-Form Playback**: Speech starts playing quickly, unlike waiting for the entire audio. - **AI-Native**: Supports skills and LLMs for your coding agent to use the real API instead of guessing. ### Tech Stack: TypeScript ### Why It Matters: Lobe TTS simplifies the process of adding natural-sounding speech output and voice input to your app, making it easier to integrate voice features into applications. It supports various voice providers and is designed for both server and browser environments, ensuring a smooth and reliable user experience.

Overview: ACE0 is a learning-based structure-from-motion approach that estimates camera parameters of sets of images by learning a multi-view consistent, implicit scene representation. This method was presented at ECCV 2024 and is part of the ACE0 paper. Key features: - Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer - RGB-D Reconstruction ("SCR Priors" paper, ICCV 2025) - Reconstruction Priors ("SCR Priors" paper, ICCV 2025) - Standard Relocalization ("ACE" paper, CVPR 2023) - Self-Supervised Relocalization - Train NeRF models or Gaussian splats Tech stack: Python Why it matters: ACE0 is a significant contribution to the field of structure-from-motion, offering a novel approach to estimating camera parameters from image sequences. Its ability to reconstruct scene coordinates and utilize scene priors makes it particularly useful for applications requiring accurate 3D reconstruction from images.