Projects

Overview: PolyLangVITS is a comprehensive multilingual speech synthesis system utilizing VITS for Korean, Japanese, English, and Chinese. This project is designed for developers and researchers interested in speech synthesis and natural language processing. Key features: - Multi-speaker support for Korean, Japanese, English, and Chinese - Uses VITS model for high-quality speech synthesis - Supports Windows/Linux systems with GPU requirements - Requires Python 3.8 and Anaconda installation - Pre-trained models available for inference and training Tech stack: Python Why it matters: PolyLangVITS provides a powerful tool for developers working on multilingual speech synthesis, offering a robust and efficient solution for creating natural-sounding text-to-speech applications.

### Overview: FyroDB is a blazing fast Redis-compatible in-memory KV store and PostgreSQL query caching proxy written in Rust. It offers sub-microsecond query cache, vector search, and seamless integration with Redis and PostgreSQL. ### Key Features: - **KV Store**: Redis-compatible, supports any Redis client. - **PostgreSQL Cache**: Caches PostgreSQL queries, serving results from memory in microseconds. - **Vector Search**: Supports vector search for full-text and geospatial queries. - **Drop-in Redis Replacement**: Perfectly integrates with Redis, providing a high-performance alternative. ### Tech Stack: Rust ### Why It Matters: FyroDB is essential for developers looking to optimize query performance and reduce database load by caching frequently accessed data. Its ability to handle vector search and seamless integration with Redis and PostgreSQL makes it a valuable tool in modern database management systems. ### Repository: Repository: Rana718/FyroDB ### GitHub Stars: GitHub stars: 75 ### GitHub Description: A Redis-compatible in-memory KV store and PostgreSQL query caching proxy — both in one Rust binary. The KV engine speaks RESP2/RESP3; any Redis client works. The proxy speaks full PostgreSQL wire protocol; any Postgres driver works. v0.3.0 — zero-allocation hot paths across the KV engine, cluster slot accelerations, mio proxy event loop, +32% extended-protocol throughput, deep proxy test coverage. → [Changelog](https://fyrodb.vercel.app/docs/changelog-0-3-0) ### Install: ```bash

### Overview: LocalS3 is an Amazon S3 mock service implementation based on Netty, designed for testing and local development. It supports S3 object versioning, vector storage and similarity search, and in-memory and persistence modes. ### Key features: - **Support for S3 Object Versioning**: Ensures data integrity and version tracking. - **Support for S3 Vectors**: Enables vector storage and similarity search. - **In Memory and Persistence Modes**: Supports both in-memory and persistence storage. ### Tech stack: Java ### Why it matters: LocalS3 simplifies testing and development by providing a reliable Amazon S3 mock service without the need for external dependencies. It is particularly useful for developers working on applications that require S3 functionality but do not have access to a real AWS environment.

Overview: TrashClaw is a versatile local agent designed to automate tasks on various hardware, including vintage Macs, using Python and a local LLM server. It supports a wide range of commands and tools, making it a powerful tool for developers and system administrators. Key features: - Supports 14 built-in tools and unlimited plugin system - Provides 21 commands for various tasks - Runs on any local LLM server or endpoint - No dependencies beyond Python 3.7 and local LLM server Tech stack: Python Why it matters: TrashClaw simplifies task automation by leveraging local LLM servers, making it ideal for developers and system administrators who need to streamline their workflows without the need for cloud services or external dependencies. Repository: Scottcjn/trashclaw GitHub stars: 75 GitHub description: Local LLM agent for vintage Mac hardware. Built on a 2013 Mac Pro trashcan with dual AMD FirePro D500s. Elyan Labs.

Overview: GenDB is an LLM-Powered Generative Query Engine designed to optimize query execution on specific data, workloads, and hardware. It accelerates performance by up to 2.8x compared to DuckDB on TPC-H and 5.0x on real-world financial data. Key features: - Optimized query execution code generation tailored to specific data, workloads, and hardware. - Five specialized LLM agents collaborating through structured JSON to generate optimized storage structures, indexes, and code. - Supports database schema, SQL queries, data files, and hardware specs for personalized query execution. - Real-world performance benchmarks show up to 5.0x faster than DuckDB on real-world financial data and 2.8x faster on TPC-H. Tech stack: C++ Why it matters: GenDB is designed to significantly improve query performance and efficiency, making it an essential tool for database developers and analysts looking to optimize their query execution processes.

Overview: This project aims to provide a high-performance KV cache management framework for large language model inference, focusing on improving memory utilization and inference speed in high-concurrency and long-context generation scenarios. Key features: - KV-cache compression - KV-cache offloading - KV-cache retrieval Tech stack: Python Why it matters: This project aims to address the KV-cache memory bottleneck in large language model inference, leading to improved inference efficiency and reduced memory usage. Repository: TheToughCrane/nano-kvllm Primary language: Python GitHub stars: 75 GitHub description: This project aims to provide a high effective KV cache manage framework for llm inference and improve memory utilization and inference speed. README excerpt: # `nano-kvllm v0.2.0` <p align="center"> <img src="assets/logo.png" alt="nano-kvllm logo" width="350"/> </p> **Here comes version 2.0!** **KvLLM** is an **AI/LLM inference framework** built on top of `nano-vllm`, focused on efficient KV-cache memory management for large language models. The current release lands **KV-cache compression** in a lightweight research-and-development framework, aiming to improve inference efficiency in **high-concurrency** and **long-context generation** scenarios by alleviating the KV-cache memory bottleneck. In the coming weeks, **KvLLM** will continue integrating frontier KV-cache management techniques, including: - **KV-cache compression** - **KV-cache

Overview: This repository showcases architectural patterns for Modular RAG using Spring AI and Ollama, including pre-requisites and running the application. Key features: - Modular RAG samples - Spring AI integration - Ollama inference server - Tavily Search API integration - Arconia Dev Services setup - PostgreSQL database - Grafana observability platform - Accessible at http://localhost:8080 Tech stack: Java Why it matters: This project provides a practical demonstration of how to implement Modular RAG using modern technologies, including Spring AI, Ollama, and Arconia Dev Services, making it valuable for developers looking to enhance their knowledge in these areas.

Overview: Seedance 1.5 Pro Prompts offers tested image prompt recipes with example renders, aspect ratios, the model ID (`seedance-1.5-pro`), and the cost per image. Key features: - 4 tested recipes - Example renders, durations, and costs - Compatible with OpenAI endpoint Tech stack: Python Why it matters: Provides a practical guide for developers to create effective prompts for Seedance 1.5 Pro, ensuring efficient use of resources and accurate pricing.

### Overview: Drax is an open-source project that provides an implementation of **Drax: Speech Recognition with Discrete Flow Matching**. This project aims to improve speech recognition by leveraging discrete flow matching techniques. ### Key Features: - **Speech Recognition with Discrete Flow Matching**: Drax is designed for speech recognition tasks, utilizing discrete flow matching to enhance the accuracy and efficiency of the recognition process. - **Model Repository**: The project includes an official implementation of the Drax model, accessible through the Hugging Face repository. - **GitHub Stars**: The repository has received 75 GitHub stars, indicating its popularity and community interest. ### Tech Stack: - **Python**: The primary programming language used for this project. - **Torch**: The deep learning framework used for implementing the model. ### Why It Matters: - **Improved Speech Recognition**: Drax aims to enhance the accuracy of speech recognition systems by incorporating discrete flow matching, offering a more effective approach to speech-to-text conversion. ### Repository: - **Repository Name**: aiola-lab/drax - **Primary Language**: Python - **GitHub Stars**: 75 - **GitHub Description**: Drax: Speech Recognition with Discrete Flow Matching ### Installation: - **Recommended Installation**: ```bash conda create -n drax python=3.10 -y conda activate drax pip install -U pip pip install torch==2.7.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl

Overview: ThermalGen is an official repository for a style-disentangled flow-based generative model designed for RGB-to-Thermal image translation. This project aims to enhance the accuracy and style of thermal image generation by disentangling the features. Key features: - Style-disentangled flow-based generative models for RGB-to-Thermal image translation - Pretrained weights available for various models (ThermalGen-B/2, ThermalGen-L/2, ThermalGen-XL/2, ThermalGen-L/2-concat) - Simple inference demo script for quick start - Preprocessing scripts for satellite-aerial datasets and other raw datasets Tech stack: Python, Hugging Face datasets and models, Conda environment setup Why it matters: This project offers a high-quality solution for generating accurate and style-disentangled thermal images, which is crucial for applications in thermal imaging and remote sensing.

Overview: This repository is a curated library of AI/ML, Deep Learning, and Generative AI resources, including notebooks, demos, and hands-on projects. Key features: - AI/ML notebooks - Non-convex optimization benchmark functions - LangChain, LlamaIndex, RAG systems, and agents - Model Context Protocol (MCP) - Observability tools - Interview preparation and question banks - Useful skills for AI agents - External resources - Utility functions - Repository rules and contributing guidelines Tech stack: Jupyter Notebook Why it matters: This repository serves as a comprehensive resource for developers looking to learn and apply AI/ML and Generative AI concepts, providing practical examples and tools for hands-on learning and development. GitHub stars: 75 Repository: dotaiz/AI-ML

Overview: This is an unofficial Python API for Perchance, a platform for creating and managing digital experiences. It provides tools for text and image generation, making it easy to create engaging content. Key features: - Text generation - Image generation Tech stack: Python Why it matters: This API streamlines the process of creating digital experiences, making it accessible to developers without requiring deep knowledge of the Perchance platform. Repository: eeemoon/perchance Primary language: Python GitHub stars: 75

### Overview: SincNet is a novel neural architecture designed for processing raw audio samples. It is implemented using Keras, specifically the Functional Framework v2+, and supports only the Tensorflow backend. The model is based on parametrized sinc functions, which are used to implement band-pass filters. ### Key Features: - **Model Conversion**: Models are converted from original torch networks. - **Supports Tensorflow Backend Only**: The model can only be trained using Tensorflow. - **Config File**: The configuration file is the same as the original code but does not support all parameters. ### Tech Stack: Keras (Tensorflow backend) ### Why it Matters: SincNet is a significant advancement in the field of audio processing, offering a novel approach to convolutional neural networks that enhances the discovery of meaningful filters. This architecture is particularly useful for applications requiring high-quality audio processing, such as speaker recognition and speech enhancement.

### Overview ARES is an open-source platform designed to simplify the evaluation and understanding of robot data, particularly for robotics research. It provides a scalable solution for researchers and developers to quickly and accurately analyze their robot datasets without the need for complex infrastructure. ### Key Features - **Ingestion**: Automatically transforms raw robot data into structured formats. - **Annotation**: Adds pseudo-labels to the data for improved downstream tasks. - **Curation and Modeling**: Analyzes data distributions and selects appropriate datasets for training or evaluation. ### Tech Stack - **Databases**: MongoDB, SQLAlchemy, FAISS - **Model Inference**: litellm, configured with various model providers - **Compute Orchestration**: Modal - **Frontend**: Streamlit, Plotly - **Development Tools**: Docker, Cursor, VSCode, Pydantic ### Why It Matters ARES aims to address the challenges faced by researchers and developers by providing a generalized platform for long-term robot research, enabling rapid development and easy scalability. It simplifies the process of data ingestion, annotation, curation, and model training, making it accessible and user-friendly. ### GitHub Stars 75 ### Repository [jacobphillips99/ares](https://github.com/jacobphillips99/ares) ### Primary Language Python

### Overview: AEROBLADE is a novel method for detecting latent diffusion images using autoencoder reconstruction error. This project presents a training-free detection approach that leverages the reconstruction error of autoencoders trained on the latent diffusion images. The method is evaluated at CVPR 2024 and achieved state-of-the-art performance on the CIFAR-100 dataset. ### Key features: - **Latent Diffusion Image Detection**: AEROBLADE identifies latent diffusion images by measuring the reconstruction error of autoencoders trained on these images. - **Training-Free Detection**: The method eliminates the need for training separate models for detection, significantly reducing computational overhead. - **Autoencoder Reconstruction Error**: Utilizes the second LPIPS layer to compute the reconstruction error, which is shown to perform best in experiments. ### Tech stack: - Python ### Why it matters: AEROBLADE offers a practical solution for detecting latent diffusion images, particularly useful in applications where real-time detection is crucial. The method's training-free nature makes it highly efficient and scalable, making it an attractive choice for various domains such as image synthesis, content-aware image editing, and more.

### Overview: This repository contains code for improved sampling techniques using learned diffusions and an optimal control perspective on diffusion-based generative modeling. The primary focus is on the `sde_sampler` method, which optimizes a neural network to control stochastic differential equations (SDEs) for improved performance. ### Key Features: - **Improved Sampling via Learned Diffusions (ICLR 2024):** This method enhances sampling by learning to control SDEs. - **Optimal Control Perspective on Diffusion-Based Generative Modeling (TMLR 2024):** This approach views diffusion-based generative modeling through an optimal control lens. - **Methods:** Various solvers for sampling from unnormalized densities, including DIS, Bridge, DDS, and PIS. ### Tech Stack: Python ### Why it Matters: This repository provides advanced techniques for sampling from unnormalized densities, which is crucial in various applications such as generative modeling, machine learning, and statistical physics. The methods aim to improve performance and efficiency in sampling tasks.

Overview: This project implements the KOSMOS-2.5 model from the Microsoft research paper "KOSMOS-2.5: A Multimodal Literate Model." Key features: - Implementation of KOSMOS-2.5 model - Uses various multimodal datasets including scanned documents, research papers, presentation slides, PDFs, web screenshots, README files, WORD documents, and webpages. Tech stack: Python Why it matters: This implementation provides a practical tool for researchers and developers interested in multimodal learning and literate models. Repository: kyegomez/Kosmos2.5 GitHub stars: 75 GitHub description: My implementation of Kosmos2.5 from the paper: "KOSMOS-2.5: A Multimodal Literate Model"

Overview: This project demonstrates the creation of a Stable Diffusion application using React and FastAPI, allowing users to generate images with Stable Diffusion through a text-based prompt. Key features: - Stable Diffusion image generation - React frontend for user interaction - FastAPI backend for processing and generating images Tech stack: JavaScript, React, FastAPI Why it matters: This project serves as an educational tool for developers to understand and implement image generation using AI techniques, specifically Stable Diffusion, and to familiarize themselves with modern web development frameworks like FastAPI.

### Overview: Qwen-Image 3.0 API is a cost-effective image generation service provided by Alibaba Cloud, offering a 1K/2K output at a fixed price of $0.02 per image. This API uses an OpenAI-compatible gateway and supports various resolutions and prompts, making it ideal for developers looking to integrate image generation into their applications without the need for a monthly subscription. ### Key features: - **Fixed Price**: $0.02 per image for 1K/2K output. - **OpenAI-Compatible Gateway**: Supports various resolutions and prompts. - **Batching and Retries**: Batches and retries tasks for seamless integration. - **No Monthly Plan Required**: Pay as you go, no subscription needed. ### Tech stack: - **Python**: The primary language used for the API. ### Why it matters: - **Cost-Effective**: Ideal for developers who need a cost-effective solution for image generation. - **OpenAI-Compatible**: Compatible with OpenAI's API, ensuring seamless integration with existing projects. - **No Subscription Required**: Pay as you go, no need for a monthly plan. - **Batching and Retries**: Ensures reliable task execution even in case of failures. ### GitHub Repository: Repository: lfzgvaqci0occqg/qwenimage3.0-qwen-image-3.0-gateway GitHub Stars: 75 ### Primary Language: Python ### GitHub Description: Qwen-Image 3.0 API (qwenimage3.0 / qwen image

Overview: This guide provides instructions on how to create a custom model compatible with Vosk using Kaldi. The process involves downloading and compiling Kaldi, preparing a dataset, and training the model on a GPU or CPU. Key features: - Custom model creation - Kaldi project compilation - Dataset preparation - Model training on GPU or CPU - Troubleshooting section Tech stack: Shell Why it matters: This guide helps developers create a custom Vosk model, which can improve speech recognition accuracy, especially for specific languages or dialects. Repository: matteo-convertino/vosk-build-model Primary language: Shell GitHub stars: 75 GitHub description: How to create your own model for vosk README excerpt: This guide tries to explain how to create your own compatible model with Vosk, using Kaldi. It includes tips on using a GPU for faster training, downloading datasets, and preparing the necessary files for training. The guide also provides a troubleshooting section for common issues.