Projects

Overview: Styleformer is a Neural Language Style Transfer framework designed to smoothly transfer natural language text between various fine-grained styles such as formal/casual, active/passive, and more. Created by Prithiviraj Damodaran, it aims to enhance the quality and readability of text across different contexts. Key features: - Transfer text from casual to formal - Transfer text from formal to casual - Transfer text from active to passive - Transfer text from passive to active Tech stack: Python Why it matters: Styleformer provides a powerful tool for text style transfer, making it easier to adjust the tone and style of text in various applications, such as improving the readability of machine-generated content or enhancing the style of paraphrased text. Repository: PrithivirajDamodaran/Styleformer Primary language: Python GitHub stars: 496 GitHub description: A Neural Language Style Transfer framework to transfer natural language text smoothly between fine-grained language styles like formal/casual, active/passive, and many more. Created by Prithiviraj Damodaran. Open to pull requests and other forms of collaboration.

### Overview: Diffusion Classifier leverages pretrained diffusion models to perform zero-shot classification without additional training, achieving strong results on a variety of benchmarks and surpassing alternative methods in terms of multimodal compositional reasoning. ### Key features: - Zero-shot classification - Leverages pretrained diffusion models - Strong performance on benchmarks - Significantly better multimodal compositional reasoning compared to discriminative approaches ### Tech stack: Python ### Why it matters: Our approach shows the potential of using generative models for downstream tasks beyond image generation, providing strong results on zero-shot classification and outperforming alternative methods in terms of multimodal compositional reasoning.

Overview: This repository contains implementations of Pervasive Attention, a model that uses two-dimensional convolutions to encode source and target sequences, and Efficient Wait-k Models for Simultaneous Machine Translation, which are transformer wait-k models trained with joint training of multiple wait-k paths. The repository includes pre-trained models and detailed documentation for both models. Key features: - Pervasive Attention: 2D Convolutional Neural Networks for Sequence-to-Sequence Prediction - Efficient Wait-k Models: Transformer wait-k models with unidirectional encoders and joint training of multiple wait-k paths Tech stack: Python Why it matters: These models offer significant improvements in sequence-to-sequence prediction tasks, particularly in natural language processing applications, by leveraging advanced neural network architectures and efficient training techniques.

Overview: Onlytrains is a project that watches a piece of train track, detects passing trains, and stitches together images of them. It uses a simple computer vision approach and is designed to work with any video4linux USB cam or Raspberry Pi camera v3 module. Key features: - Detects trains - Stitches together images of trains - Works with a pre-cropped region - Stable camera and image without movement - No camera calibration, undistortion, or perspective mapping - CPU-based processing Tech stack: Go Why it matters: The project is a simple hobby project focused on train detection and stitching, suitable for use with any video4linux USB cam or Raspberry Pi camera v3 module. It is designed to be lightweight and efficient, making it suitable for real-time applications. Repository: jo-m/trainbot GitHub stars: 496 GitHub description: Watches a piece of train track, detects passing trains, and stitches together images of them.

### Overview: libfaceid is a Python library designed for prototyping face recognition solutions, seamlessly integrating multiple detection, recognition, and liveness models with speech synthesis and speech recognition capabilities. It simplifies the process of facial recognition by providing a comprehensive list of models and enabling users to mix and match models based on their specific use cases and system requirements. ### Key Features: - **Face Detection and Recognition Models**: Seamless integration of multiple detection and recognition models. - **Model Compatibility**: Models are compatible with each other, abstracting users from differences in speed, accuracy, memory requirements, and third-party library dependencies. - **Face Liveness Detection**: Provides models for anti-face spoofing attacks (photo-based, video-based, 3D-mask-based). - **Voice-Enabled and Voice-Activated Features**: Models for predicting age, gender, emotion, and facial landmarks, along with TTS and STT for voice-enabled and voice-activated capabilities. - **Web App Support**: Flask-based web app for remote video capture viewing on another computer in the same network. ### Tech Stack: Python ### Why It Matters: libfaceid simplifies the process of facial recognition by providing a user-friendly interface, modular design, and robust functionality. It enables beginners to learn various models and experiment with different solutions appropriate for their specific use cases and system requirements. The library's design allows for easy integration of new models in the future, making it a valuable tool for researchers and developers in the field of deep learning and facial recognition.

### Overview: This project is a PyTorch implementation of a 3D bounding box estimation method using deep learning and geometry. It aims to provide accurate and efficient 3D bounding box detection for various applications, such as autonomous driving and robotics. ### Key Features: - PyTorch-based implementation - Uses deep learning techniques for orientation and dimension prediction - Integrates YOLOv3 for 2D bounding box estimation - Supports real-time inference with a latency of approximately 0.4 seconds per frame - Available on GitHub with 496 stars - Includes a Slack workspace for discussion and collaboration ### Tech Stack: - Python - PyTorch - CUDA - OpenCV ### Why It Matters: This project is crucial for applications requiring precise 3D bounding box detection, such as autonomous vehicles, robotics, and augmented reality systems. It offers a practical solution for real-time object detection tasks. ### Repository: Repository: skhadem/3D-BoundingBox Primary Language: Python GitHub Stars: 496 GitHub Description: PyTorch implementation for 3D Bounding Box Estimation Using Deep Learning and Geometry

Overview: Pytorch-SVGRender is a library designed for differentiable rendering methods, enabling the generation of vector graphics using neural networks. It supports text-to-SVG, image-to-SVG, and SVG editing functionalities. Key features: - Supports text-to-SVG, image-to-SVG, and SVG editing. - Based on PyTorch and uses differentiable rendering techniques. - Open-sourced version 1.0 in 2023. Tech stack: Python, PyTorch. Why it matters: Pytorch-SVGRender simplifies the process of generating vector graphics, making it easier to create and manipulate SVGs using neural networks, which is particularly useful in fields such as design, engineering, and education. Repository: ximinng/PyTorch-SVGRender GitHub stars: 496 GitHub description: SVG Differentiable Rendering: Generating vector graphics using neural networks. Support: text-to-SVG, Image-to-SVG, SVG Editing.

Overview: This project involves generating multiple-choice questions from text using Machine Learning techniques, specifically utilizing T5 Transformers to create question-answer pairs with distractors. Key features: - Identifying keywords from text as potential answers. - Transforming sentences to resemble question-like sentences. - Generating distractors that are similar to the correct answer. Tech stack: Python, Jupyter Notebook, spaCy, T5 Transformers. Why it matters: The approach for identifying keywords as answers has been accepted in the RANLP2021 conference, demonstrating the project's relevance in the field of natural language processing and question generation.

### Overview Many formulas or equations are floating around in papers, blogs, etc., about how to calculate training or inference latency and memory for Large Language Models (LLMs) or Transformers. Rather than doing math on papers or typing in Excel sheets, `let's automate the boring stuff with llm-analysis` :gear:! Given the specified model, GPU, data type, and parallelism configurations, llm-analysis estimates the latency and memory usage of LLMs for training or inference. With llm-analysis, one can easily try out different training/inference setups theoretically, and better understand the system performance for different scenarios. ### Key Features - **Model Configuration**: llm-analysis supports various model configurations, including LLaMA, Megatron-LM, and FasterTransformer models. - **Latency and Memory Estimation**: It calculates the latency and memory usage based on the specified model, GPU, data type, and parallelism configurations. - **Training/Inference Setup**: llm-analysis helps determine the optimal batch size, data type, parallelism scheme, and other parameters for training or inference. - **Benchmarking**: It provides a benchmark for different setups, helping to understand the performance impact of various factors. ### Tech Stack - Python ### Why It Matters Understanding the performance of LLMs and Transformers is crucial for optimizing training and inference setups. llm-analysis simplifies the process by automating the calculations, making it easier to experiment with different configurations and quickly assess the impact on latency and memory usage. ### Examples Check the example use cases. With llm

Overview: `Deep Clothes Detector` is a clothes detection framework based on Fast R-CNN, designed to identify potential upper-body, lower-body, and full-body clothes in fashion images. Key features: - Uses Fast R-CNN for detecting clothes - Supports upper-body, lower-body, and full-body detection - Provides robust performance for fashion image analysis Tech stack: MATLAB Why it matters: This project is crucial for enhancing the accuracy and reliability of clothing recognition and retrieval, particularly in the context of fashion images, making it valuable for both research and practical applications in the field of computer vision.

### Overview: MirrorMate is a self-hosted personalized AI that lives in a mirror, providing quick and efficient access to AI without the need for a screen. It runs entirely locally with Ollama and VOICEVOX, offering a seamless and intuitive experience. ### Key Features: - **Voice-First Interaction**: Activated by a wake word, MirrorMate responds to voice commands. - **Personalized Memory**: Remembers and retrieves information from your interactions, enhancing the user experience. - **RAG-Based Memory**: Utilizes Rich Auditory Graphs to store and recall information effectively. ### Tech Stack: TypeScript ### Why It Matters: MirrorMate simplifies access to AI by integrating seamlessly with everyday objects, making it accessible and practical for users. Its local implementation ensures privacy and reduces the need for cloud services, making it an ideal solution for those seeking a more secure and personalized AI experience.

### Overview: A linguistic linter for Traditional Chinese (zh-TW) that enforces Taiwan Ministry of Education (MoE) standards on vocabulary, punctuation, and character shapes. It plugs into AI coding assistants through the Model Context Protocol (MCP) and catches Mainland Chinese (zh-CN) regional drift before it reaches the user. ### Key Features: - **Cross-Strait Vocabulary Enforcement**: Compiles over 1100 vocabulary rules and 15 casing rules into the binary. - **Character Shape Compliance**: Enforces MoE standard character shapes through datasets like OpenCC's TWPhrases/TWVariants. - **Ambiguous Term Resolution**: Uses AI assistance to decide on ambiguous terms. - **Profile-Based Enforcement**: Two profiles—`base` and `strict`—enforce different levels of MoE standards. ### Tech Stack: Rust ### Why It Matters: ### Modern Chinese is an inadequately standardized language. The PRC simplification effort reduced not just stroke counts but vocabulary precision. Terms that should vary by domain got flattened into single catch-all translations. Many PRC translations were coined hastily: if a term worked in one context, it spread uncritically to others. AI models amplify the problem by favoring zh-CN terminology when asked to write zh-TW, leading to inconsistencies and inaccuracies. This goes beyond character conversion, as the same word often means different things across the strait. This project ensures that zh-TW text produced by AI is consistent with Taiwan's official standards, preventing cross-strait terminology leaks and maintaining linguistic

### Overview: OCRAutoScore 是一个 OCR 自动阅卷项目，旨在通过自动化方式处理和评分各种类型的文本输入，包括填空题、选择题和作文。该项目使用了 Python 作为主要编程语言，并且通过 GitHub 上的仓库地址 [https://github.com/vkgo/OCRAutoScore](https://github.com/vkgo/OCRAutoScore) 可以访问到源代码和相关文档。 ### Key features: - **OCR 模型**: 使用了多种 OCR 模型，如 SpinalVGG 和 WaveMix，以提高文本识别的准确性。 - **模型集成**: 将 OCR 模型、AES 模型和 CAN 模型整合在一起，形成一个完整的阅卷系统。 - **前端和后端**: 分别构建了前端网页和后端服务，通过 RESTful API 进行数据交互。 - **数据处理**: 支持多种输入格式，包括图片和文本，通过 OCR 技术进行自动评分和批改。 ### Tech stack: - **Python**: 主要编程语言。 - **PaddleOCR**: OCR 模型之一，用于文本识别。 - **PIL**: 图像处理库。 - **PyTorch**: 用于处理图像数据。 - **CLIP**: 用于文本和图像的相似度判断。 ### Why it matters: 通过自动化阅卷系统，可以显著提高阅卷效率，减少人为错误，同时提高评分的准确性和一致性。这对于教育领域中的大规模阅卷任务

Overview: VibeVoiceFusion is a web application for multi-speaker voice generation and speech recognition, built on Microsoft's VibeVoice models. It features text-to-speech, file transcription, and real-time microphone transcription, with support for LoRA fine-tuning and OpenAI-compatible APIs. Key features: - Multi-speaker dialogue with voice cloning - Narration mode for single-speaker content - Quick Generate and Multi-Generation - Dialog editor - File transcription with VibeVoice-ASR - Live transcription from browser microphone - LoRA fine-tuning with adjustable weight - Runs on consumer GPUs with 10GB+ of VRAM - OpenAI-compatible API integration - Full English/Chinese UI - Docker image on Docker Hub - CLI tools for TTS, ASR, and model conversion Tech stack: Python Why it matters: VibeVoiceFusion combines text-to-speech, speech recognition, and multi-speaker voice generation into a single, user-friendly web application, making it a valuable tool for creators and enthusiasts in the fields of voice cloning, speech synthesis, and AI-driven content creation.

### Overview: This project provides an implementation of "[End-to-End Object Detection with Fully Convolutional Network](https://arxiv.org/abs/2012.03544)" on PyTorch. The implementation is based on the internal framework of the Megvii-BaseDetection organization and has been reimplemented on the cvpods framework for better reproducibility and performance. ### Key features: - Fully Convolutional Network (FCN) for object detection. - End-to-End training pipeline. - Supports multiple backbones and architectures. - Includes detailed experiments and results on the COCO 2017 validation set. ### Tech stack: Python ### Why it matters: This project offers a practical and reproducible implementation of advanced object detection techniques, making it valuable for researchers and developers looking to improve their own projects or applications in the field of computer vision.

### Overview: TeleMem is a high-performance drop-in replacement for Mem0, offering semantic deduplication, long-term dialogue memory, and multimodal video reasoning. ### Key Features: - **Semantic Deduplication**: Ensures unique representations for similar content. - **Long-Term Dialogue Memory**: Stores and retrieves multi-turn dialogues efficiently. - **Multimodal Video Reasoning**: Handles video frames, captions, and vector databases for comprehensive reasoning. ### Tech Stack: Python ### Why It Matters: TeleMem provides a robust and efficient solution for complex conversational AI applications, enhancing accuracy, performance, and character memory capabilities.

### Overview: HOLD is a method that reconstructs articulated hands and objects from monocular videos without assuming a pre-scanned object template or 3D hand-object training data. It has been awarded as a CVPR highlight and is part of the 2024 CVPR conference. ### Key features: - Joint reconstruction of hands and objects from monocular videos - Template-free bimanual hand-object reconstruction - Textureless object interaction with hands - Volumetric rendering framework for dynamic hand-object interaction - Generalized codebase for single and two-hand interaction with objects - Viewer for interactive prediction - Code for evaluating and comparing with HOLD in HO3D ### Tech stack: Python ### Why it matters: HOLD is a significant contribution to the field of 3D reconstruction of articulated hands and objects from monocular videos, offering a novel approach to category-agnostic reconstruction without the need for pre-scanned templates or 3D hand-object training data. Its ability to handle novel objects and hands makes it highly versatile and useful in various applications, including robotics, augmented reality, and interactive simulations.

Overview: DD3D is an official PyTorch implementation of the DD3D project, which addresses the need for monocular 3D object detection. This project aims to evaluate the necessity of pseudo-lidar in such scenarios. Key features: - DD3D implementation of monocular 3D object detection - Supports PyTorch framework - Includes datasets and experiments for evaluation - Compatible with Docker for reproducibility Tech stack: Python Why it matters: DD3D provides a practical solution for evaluating the performance of monocular 3D object detection methods, making it valuable for researchers and developers looking to improve their models without the need for expensive hardware. Repository: TRI-ML/dd3d GitHub stars: 493 GitHub description: Official PyTorch implementation of DD3D: Is Pseudo-Lidar needed for Monocular 3D Object detection? (ICCV 2021), Dennis Park*, Rares Ambrus*, Vitor Guizilini, Jie Li, and Adrien Gaidon.

### Overview: This repository is a comprehensive guide and resource hub for developers interested in building and designing agentic AI systems. It covers foundational concepts, frameworks, system design patterns, real-world use cases, and practical examples. The content is designed to help developers create production-ready AI agents and multi-agent systems, making it a valuable resource for anyone looking to advance their knowledge in AI and agentic systems. ### Key Features: - **[Agentic AI Design Patterns](./03_system_design/design-patterns/agentic-ai-design-patterns.md):** A detailed overview of reusable design patterns for agentic AI systems. - **[2026 Agentic AI System Design Update](./03_system_design/2026-agentic-ai-system-design.md):** Current framing for modern AI engineering, long-running agents, AgentOps, context engineering, MCP/A2A, security, and cost-aware design. - **[RAGs](./03_system_design/RAGs/README.md):** Comprehensive overview of RAG architectures, including basic RAGs, agentic RAGs, multi-modal RAGs, and advanced RAGs. - **[Cookbooks](./03_system_design/cookbooks/README.md):** A comprehensive overview of the most useful cookbooks for designing and building agentic AI systems. - **[OpenAI Agents](./02_frameworks/1_openai/README.md):** Comprehensive guides, tutorials, and code examples for building AI agents using the OpenAI API and Agents SDK. - **[

Overview: Anarchy Labs' LLM-VM is a highly optimized backend for running LLMs, providing features like tool usage, persistent memory, and web playgrounds. Key features: - Tool usage - Persistent stateful memory - Live data augmentation - Data and task fine-tuning - Output templating - Web playground - API endpoints - Student-teacher distillation - Data synthesis - Load-balancing and orchestration - Large context-window mimicry Tech stack: Python Why it matters: This project aims to democratize AGI development by providing an open-source backend for LLMs, optimizing performance and flexibility, and fostering a vibrant community of developers. Repository: anarchy-ai/LLM-VM GitHub stars: 493 GitHub description: irresponsible innovation. Try now at https://chat.dev/