Projects

Overview: uSpeech is a speech recognition toolkit designed for Arduino, providing phoneme-based recognition with a focus on accuracy and simplicity. Key features: - Letter-based recognition - Small memory footprint - Arduino-compatible - No training required - Fixed point arithmetic - 30% to 40% accuracy for phoneme-based recognition, up to 80% for words Tech stack: C++ Why it matters: uSpeech offers a straightforward and efficient solution for speech recognition on Arduino, suitable for projects requiring real-time speech processing without the need for complex training. Repository: arjo129/uSpeech GitHub stars: 476 GitHub description: Speech recognition toolkit for the arduino

Overview: SpanMarker is a framework for training powerful Named Entity Recognition models using familiar encoders such as BERT, RoBERTa, and ELECTRA. It leverages the Transformers library for performance and ease of use. Key features: - Built on top of the Transformers library - Works with common encoders like `bert-base-cased`, `roberta-large`, and `bert-base-multilingual-cased` - Automatically supports datasets with `IOB`, `IOB2`, `BIOES`, `BILOU`, or no label annotation scheme - Integrates with Hugging Face Hub and Inference API for free inference and production deployment Tech stack: Python, Transformers Why it matters: SpanMarker simplifies the process of training and deploying NER models, making it accessible to developers and researchers alike. Its ease of use and compatibility with various encoders make it a valuable tool in the NER toolkit.

### Overview: SimCTG is a novel contrastive framework designed for neural text generation, showcasing significant improvements over traditional autoregressive methods. ### Key Features: - **SimCTG Training**: A contrastive training approach that leverages the contrastive search method to enhance the performance of neural text generation models. - **Off-the-Shelf Models**: Demonstrates superior performance on various language models, including 12 out of 16 evaluated languages. - **Versatile Applications**: The framework can be applied to both PyTorch and TensorFlow platforms, making it highly versatile. ### Tech Stack: Python ### Why It Matters: SimCTG provides a new perspective on neural text generation, offering significant advantages over traditional autoregressive methods. Its ability to improve performance on off-the-shelf models makes it a valuable tool for researchers and practitioners alike.

Overview: This project focuses on Large Language Models (LLMs), providing a comprehensive guide to creating enriched prompts for GPT, Wikipedia, ArXiv, Google Scholar, Stack Exchange, and Google search. The project includes a detailed textbook and solutions for various projects related to NLP and LLMs. Key features: - Enriched prompts for GPT, Wikipedia, ArXiv, Google Scholar, Stack Exchange, and Google search - Category structure creation using Wolfram directory and associated URLs - Good word embeddings generation for LLMs - Focus on Probability & Statistics subdirectory Tech stack: TeX Why it matters: This project offers valuable insights into the creation of enriched prompts for large language models, making it essential for developers looking to improve their models' performance and applications. Repository: VincentGranville/Large-Language-Models GitHub stars: 475 GitHub description: xLLM 1.0, smart crawling, knowledge graph discovery.

### Overview: Audio Development Tools (ADT) is a comprehensive project designed to advance various aspects of sound, speech, and music technologies. It includes components for machine learning, audio generation, signal processing, sound synthesis, game audio, digital audio workstations, spatial audio, music information retrieval, music generation, speech recognition, speech synthesis, singing voice synthesis, and audio evaluation. ### Key Features: - **Machine Learning (ML):** Librosa, Essentia, DDSP, MIDI-DDSP, DDSP-VST, torchsynth, aubio, audioFlux, Polymath. - **Audio Generation (AG):** Librosa, Essentia, DDSP, MIDI-DDSP, DDSP-VST, torchsynth, aubio, audioFlux, Polymath. - **Audio Signal Processing (ASP):** Librosa, Essentia, DDSP, MIDI-DDSP, DDSP-VST, torchsynth, aubio, audioFlux, Polymath. - **Sound Synthesis (SS):** Librosa, Essentia, DDSP, MIDI-DDSP, DDSP-VST, torchsynth, aubio, audioFlux, Polymath. - **Game Audio (GA):** Librosa, Essentia, DDSP, MIDI-DDSP, DDSP-VST, torchsynth, aubio, audioFlux, Polymath. - **Digital Audio Workstation (DAW):** Librosa, Essent

### Overview: Multi-Modal Mamba (MultiModalMamba) is a novel implementation of fusing Vision Transformer (ViT) with Mamba, creating a high-performance multi-modal model. This model is built on Zeta, a simple yet powerful AI framework designed to streamline and enhance machine learning model management. MultiModalMamba is particularly useful for tasks that require processing and interpreting both text and image data, addressing the need for efficient handling of multiple data types. ### Key Features: - **ViT and Mamba Fusion**: Integrates Vision Transformer and Mamba to create a multi-modal model. - **Zeta Framework**: Built on Zeta, a minimalist yet powerful AI framework. - **Versatile Application**: Suitable for a wide range of AI tasks, including text and image data processing. - **Customizable Architecture**: Parameters such as depth, dropout, heads, d_state, image_size, patch_size, encoder_dim, encoder_depth, encoder_heads, and fusion_method can be adjusted for optimal performance. - **Return Embeddings**: Option to return embeddings instead of the final output, useful for transfer learning or feature extraction tasks. ### Tech Stack: - Python ### Why It Matters: - **Versatility**: Handles both text and image data, making it suitable for a broad spectrum of AI tasks. - **Customizability**: Allows for easy adjustment of parameters to fit specific task requirements. - **Efficiency**: Streamlines model management and enhances performance through efficient data handling. ### GitHub Stars: 475 ### Primary Language: Python ### GitHub Description: A novel implementation

Overview: This repository is a simplified implementation of all RAG techniques, focusing on text-based approaches without the need for multimodal integration. The project aims to make RAG accessible and understandable by providing clear, practical code examples. Key features: - Simple RAG: Basic RAG implementation - Semantic Chunking: Segments text based on semantic similarity - Context Enriched Retrieval: Retrieves adjacent blocks to enhance context - Contextual Chunk Headers: Adds descriptive titles before embedding - Document Augmentation RAG: Generates questions to enhance retrieval - Query Transformation: Reorganizes queries to improve search effectiveness - Reranker: Reorders results using LLMs for better relevance - RSE: Extracts and reconstructs continuous text segments - Contextual Compression: Filters and compresses retrieved blocks for better information - Feedback Loop: Learns and improves RAG system over time - Adaptive RAG: Selects optimal retrieval strategies based on query types - Self RAG: Decides when and how to retrieve, evaluates relevance, and assesses support and utility - Fusion RAG: Combines vector search and keyword-based retrieval for improved results - Graph RAG: Organizes knowledge into a graph for concept traversal - Hierarchy RAG: Builds a hierarchical index for efficient retrieval - HyDE RAG: Enhances semantic matching using hypothesis documents - CRAG: Corrects retrieval quality and uses network search as a backup - Proposition Chunking: Breaks down documents into atomic fact statements for precise retrieval - Chunk Size Selector: Investigates

Overview: PyNLPl is a Python library designed for Natural Language Processing (NLP) tasks. It includes modules for common and less common NLP tasks, such as extracting n-grams, building language models, and working with FoLiA XML format. Key features: - Various modules for NLP tasks - Extensive library for FoLiA XML format - Clients for interfacing with NLP-specific servers - Support for FoLiA, Giza, Moses, ARPA, Timbl, and CQL file formats Tech stack: Python Why it matters: PyNLPl provides a comprehensive set of tools for NLP tasks, making it a valuable resource for developers working with natural language processing.

### Overview: This repository contains official skills for the GLM family of models, designed for agent architectures including Claude Code, OpenCode, OpenClaw, AutoClaw, and other AI coding agents. Skills are consolidated from individual model repos into a single, unified collection. ### Key Features: - **GLM-V (Multimodal):** - **glmv-caption:** Generate captions and descriptions for images, videos, and documents. - **glmv-doc-based-writing:** Write content based on PDF/DOCX documents. - **glmv-grounding:** Object localization with bounding-box visualization in images and videos. - **glmv-pdf-to-ppt:** Convert PDF documents into multi-slide HTML presentations. - **glmv-pdf-to-web:** Convert research papers into polished academic project websites. - **glmv-prd-to-app:** Build full-stack web applications from PRD documents and prototypes. - **glmv-prompt-gen:** Generate AI art prompts from visual references (Midjourney, SD, DALL-E, etc.). - **glmv-resume-screen:** Screen and evaluate resumes against user-defined criteria. - **glmv-stock-analyst:** Multi-source stock analysis and report generation for HK/A-share/US markets. - **glmv-web-replication:** Create frontend visual replicas of existing websites. - **glmocr:** General text extraction from images and PDFs. - **glmocr-formula:** Extract mathematical formulas into LaTeX. - **glmocr-handwriting:** Recognize handwritten text.

Overview: DB2Rest is a modern low-code REST DATA API platform that automates the creation of secure REST API endpoints for relational databases, enabling developers to build intelligent applications 30x faster without the need for ORM or code generation. Key features: - Instant REST API creation - Secure database gateway - Integration of existing/new databases - Context-aware, reasoning applications - No code generation - No vendor lock-in Tech stack: Java Why it matters: DB2Rest simplifies the process of integrating data from various databases, enabling developers to build intelligent applications 30x faster and without the need for code, making it an essential tool for accelerating application delivery and innovation with Gen AI. GitHub stars: 474 Repository: 9tigerio/db2rest

Overview: OmniShow is a unified model designed for Human-Object Interaction Video Generation (HOIVG), incorporating text, reference image, audio, and pose conditioning. It is the first all-in-one model for this task, providing flexibility and versatility in various applications. Key features: - Unified Channel-wise Conditioning - Gated Local-Context Attention - Decoupled-Then-Joint Training Tech stack: Python Why it matters: OmniShow offers a comprehensive solution for generating human-object interaction videos with diverse multimodal inputs, making it a significant advancement in the field of video generation and human-object interaction research.

Overview: ResuLLMe is a project that utilizes Large Language Models to enhance résumés, aiming to help candidates avoid common mistakes in job applications. It provides a smart career advisor to improve résumés and convert them into a JSON format for better readability. Key features: - Enhances résumés following published guidelines - Converts résumés to JSON format - Generates a new PDF of the enhanced resume using LaTeX Tech stack: Jinja Why it matters: ResuLLMe offers a valuable tool for job seekers to improve their résumés, making them more appealing and professional. The use of LLMs ensures that the recommendations are accurate and tailored to the candidate's specific needs. Repository: IvanIsCoding/ResuLLMe GitHub stars: 474 GitHub description: Enhance your résumé with Large Language Models README excerpt: # ResuLLMe [](https://resullme.streamlit.app/) [](https://github.com/IvanIsCoding/ResuLLMe/blob/main/LICENSE)  ---  👉 Check some sample résumés generated by ResuLLMe ([1](./.github/samples/Simple.pdf), [2](./.github/samples/

Overview: This repository provides a PyTorch implementation of the "High-Fidelity Generative Image Compression" (HIFIC) paper, focusing on image compression techniques that maintain perceptual quality. It includes a model for learnable image compression and general utilities for neural image compression. Key features: - High-Fidelity Generative Image Compression model - Lossless compression up to two orders of magnitude in size - Maintains perceptually similar reconstructions - Includes a partial port of the TensorFlow Compression library Tech stack: PyTorch Why it matters: This repository offers a practical solution for image compression that preserves quality, making it suitable for applications requiring high fidelity and lossless compression, such as medical imaging or important documents.

Overview: This project presents a novel high-fidelity GAN inversion framework designed for image attribute editing, specifically focusing on preserving background, appearance, and illumination details. Key features: - Inference code and pretrained model release - Training code coming soon - Colab notebook for interactive play Tech stack: Jupyter Notebook Why it matters: This framework offers high-fidelity image attribute editing, ensuring that the edited images retain detailed background, appearance, and lighting conditions, making it suitable for applications requiring precise and realistic modifications. Repository: Tengfei-Wang/HFGI Primary language: Jupyter Notebook GitHub stars: 474 GitHub description: CVPR 2022 HFGI: High-Fidelity GAN Inversion for Image Attribute Editing README excerpt: The project includes a detailed introduction, a list of models and their descriptions, instructions for setting up the environment, and a quick start guide for both inference and training.

### Overview: This repository demonstrates the process of converting a Basel Face Model (BFM) into a FLAME head model, enabling the creation of a texture-based model from the BFM's vertex color space. ### Key Features: - Convert BFM vertex color space to FLAME texture model. - Convert BFM mesh to a FLAME mesh. - Include detailed instructions for setting up the environment and running the conversion process. ### Tech Stack: Python ### Why it Matters: This conversion is crucial for researchers and developers who need to leverage the FLAME model for applications such as photometric optimization and 3D face reconstruction, while maintaining the BFM's original features and vertex colors.

### Overview: D-Fire is an extensive image dataset designed for fire and smoke detection, containing over 21,000 images compiled from various sources including surveillance cameras, the web, and public datasets. It features annotations following the YOLO format, making it suitable for machine learning and object detection algorithms. ### Key features: - **Image Categories**: D-Fire includes images categorized into only fire, only smoke, fire and smoke, and no instances. - **Bounding Boxes**: The dataset contains annotations for fire and smoke instances, with a total of 14,692 bounding boxes for fire and 11,865 for smoke. - **Normalization**: Annotations are provided in the YOLO format, with normalized coordinates between 0 and 1, facilitating pixel coordinate conversion. ### Tech stack: Python ### Why it matters: D-Fire serves as a valuable resource for researchers and developers looking to improve fire and smoke detection algorithms, offering comprehensive and diverse data for training and validation purposes.

Overview: CrystalAI is a fast, scalable, and extensible decision-making AI utility for C# and Unity, built on utility theory concepts. Key features: - Fast implementation of the core simulation loop - Increased code coverage to 95% - Semantic versioning for future releases - Basic example project showcasing the AI Tech stack: C# Why it matters: CrystalAI is a powerful tool for developers looking to incorporate decision-making capabilities into their C# and Unity projects, making it easier to create intelligent AI systems without the complexity of traditional machine learning methods.

### Overview GP-GAN is an official Chainer implementation of the high-resolution image blending algorithm presented in "GP-GAN: Towards Realistic High-Resolution Image Blending" by Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. This project aims to generate realistic and high-resolution blended images using deep generative models. ### Key Features - **Realistic Blending**: The algorithm can seamlessly blend source and destination images, creating realistic high-resolution results. - **High Resolution**: The generated images are of high resolution, suitable for various applications such as augmented reality and image editing. - **Chainer Implementation**: The project is built using the Chainer framework, ensuring efficient and scalable performance. ### Tech Stack - **Language**: Python - **Framework**: Chainer ### Why It Matters GP-GAN addresses the challenge of blending images with high realism and high resolution, which is crucial in fields such as augmented reality, image editing, and content-aware image synthesis. The project's implementation and the associated research provide valuable insights into the state-of-the-art in image blending and generative models. ### Repository - **Repository**: wuhuikai/GP-GAN - **GitHub Stars**: 474 - **Primary Language**: Python - **GitHub Description**: Official Chainer implementation of GP-GAN: Towards Realistic High-Resolution Image Blending (ACMMM 2019, oral) - **Colab**: [Link](https://colab.research.google.com/drive/11ESQZ

### Overview: Surgeon is a library designed to inspect and extract intermediate layers of PyTorch models without modifying the original code. It provides tools to visualize, analyze, and manipulate the layers of models, making it useful for various applications such as debugging, visualization, and training separate components. ### Key features: - **Inspect Layers**: Display all layers of a PyTorch model. - **Extract Nodes**: Display all intermediate nodes of a PyTorch model's graph. - **Named Outputs**: Provide named outputs for specific layers. ### Tech stack: Python Surgeon is written in Python, leveraging the power of PyTorch to inspect and extract model layers. ### Why it matters: - **Inspection without modification**: Useful for understanding and debugging models without altering the original code. - **Visualization and analysis**: Helps in visualizing layer embeddings and analyzing model behavior. - **Model separation**: Enables the extraction of subparts of the model for independent training or analysis. ### Why it matters: It's often the case that we want to inspect or extract intermediate layers of PyTorch models without modifying the code. This can be useful for various applications such as debugging, visualization, and training separate components.

### Overview: A community-driven PHP client for DeepSeek AI, designed for clean API access, fluent developer experience, and framework-friendly integration. ### Key Features: - Seamless API integration - Fluent builder pattern for intuitive request building - Enterprise-ready PSR-18 compliant HTTP client integration - Latest DeepSeek V4 models, including flagship `Models::V4_PRO` and `Models::V4_FLASH` - Real-time response handling - Easy to use `Guzzle` and `symfony/http client` - Framework-friendly Laravel & Symfony packages ### Tech Stack: PHP ### Why It Matters: This client simplifies the integration of DeepSeek AI capabilities into PHP applications, providing a clean, expressive, and production-ready solution.