Projects

### Overview: Nano Banana 2 API の日本語ガイド：モデル ID、1枚あたり $0.015 の料金、パラメータ一覧、curl・Python の実行例と OpenAI 互換エンドポイントの解説。 ### Key features: - **Model ID**: 1枚あたり $0.015 の料金 - **1枚あたりの価格**: $0.015 - **パラメータ一覧**: モデル、リソース、サイズ、呼び出し回数 - **実行例**: curl・Python の実行例 - **OpenAI 互換エンドポイント**: 解説 ### Tech stack: - Python ### Why it matters: - **1 つのキーで全カタログ**：同じベース URL と認証ヘッダで Nano Banana 2 を含む 300 以上の画像・動画・言語モデルに到達でき、切り替えは `model` フィールドのみ。 - **最低 1 ドル、従量課金**：サブスクも前払いプランも不要。無料枠を使い切る必要もなく、表の単価がそのまま実単価です。 - **課金額がレスポンスで返る**：各呼び出しが `cost` / `credits_cost` を返すため、月末に推測する必要がありません。 - **

Overview: The Aussie AI Base C++ Library is a comprehensive source code repository for the book "Generative AI in C++," focusing on low-level Transformer and LLM components. Key features: - Implementations of Transformer and LLM components in C++ - Educational material for the book "Generative AI in C++" - General-purpose programming methods including bitwise tricks, floating-point tricks, and advanced C++ coding techniques Tech stack: C++ Why it matters: It serves as a practical resource for understanding and implementing AI and ML concepts in C++, complementing the book's theoretical explanations. Repository: aussie-ai-labs/aussie-ai-base GitHub stars: 26

Overview: Research4j is a comprehensive Java library designed for automating research workflows, providing intelligent query analysis, dynamic citation fetching, adaptive reasoning strategies, and seamless integration with various LLMs and APIs. Key features: - Dynamic Query Analysis - Adaptive Reasoning Selection - Multi-Source Citation Fetching - User Profile Personalization Tech stack: Java Why it matters: Research4j offers a robust and flexible solution for applications requiring automated research capabilities, making it a valuable tool for developers looking to enhance their applications with intelligent research automation. Repository: bhavuklabs/research4j Primary language: Java GitHub stars: 26 GitHub description: Build your own perplexity for your applications using research4j and integrate them for any domain specific usecases

Overview: This is a Lua library that provides bindings for DeepSpeech, an open-source speech recognition library. It is designed to be used with LÖVR and LÖVE, and should work with any Lua program that has audio samples in a table or lightuserdata. Key features: - Lua bindings for DeepSpeech - Intended for use with LÖVR and LÖVE - Can work with any Lua program that has audio samples in a table or lightuserdata - Includes a scorer package for custom word recognition Tech stack: The library is written in Lua and uses C for the backend. Why it matters: This library is useful for developers who want to integrate speech recognition functionality into their game or application, making it easier to implement speech-to-text capabilities in their projects.

### Overview: The CMU Sphinx Models repository contains acoustic and language models for minorised languages, specifically Catalan. It includes the necessary files for training and testing the models, as well as the training data for the ParlamentParla corpus. ### Key features: - **Models**: The repository includes language models, phonetic dictionaries, and acoustic models for Catalan. - **Training Data**: The training data consists of 240 hours of TV3Parla corpus and 320 hours of Parliament plenary sessions. - **Acoustic Models**: The models are trained using the *sphinxtrain 5prealpha* tool. - **Resources**: The project provides resources for speech technologies and links to download the data. ### Tech stack: Python ### Why it matters: The models and training data are essential for speech recognition and natural language processing tasks in Catalan, making them valuable tools for developers working in the field of ASR and NLP.

### Overview: contextdb is an embedded database system designed for agentic memory systems. It provides relational, graph, and vector storage under unified MVCC transactions, ensuring enforceable policy constraints and seamless data synchronization. ### Key Features: - **Unified MVCC Transactions**: Ensures data consistency and enforceability. - **Relational Storage**: Supports traditional SQL queries. - **Graph Traversal**: Enables efficient graph-based operations. - **Vector Similarity Search**: Facilitates fast vector similarity queries. ### Tech Stack: Rust library and CLI. ### Why It Matters: contextdb simplifies the process of building agent memory systems by replacing multiple databases with a single, unified system. It ensures enforceable policy constraints and seamless data synchronization, making it ideal for systems where state, graph traversal, and vector similarity search are required. ### Repository: - **Repository**: context-graph-ai/contextdb - **GitHub Stars**: 24 - **GitHub Description**: Embedded database for agentic memory systems. ### Use It As a Library: - **Rust API**: Use the Rust API for interacting with contextdb. - **Python and TypeScript Bindings**: Contributions welcome. ### Website: - **Docs**: [contextdb.tech/docs](https://contextdb.tech/docs/) - **How It Works**: [contextdb.tech/how-it-works](https://contextdb.tech/how-it-works) ### Getting Started: - **2 Minutes**: Try it out in 2 minutes. - **Documentation**: [Why contextdb?](docs/why-contextdb.md) for the full problem statement.

Overview: Coqui STT Model Manager is a tool designed to install and manage Coqui STT models from the Coqui Model Zoo. Key features: - Install and manage Coqui STT models - Support for Coqui Model Zoo models - Python 3.7 compatibility for Raspberry Pi Tech stack: Python, JavaScript, Node.js, Yarn, Python 3.7 Why it matters: Coqui STT Model Manager simplifies the process of installing and trying out Coqui STT models, making it easier for developers to experiment with different models.

Overview: This hands-on lab guides you through the step-by-step process of efficiently serving and fine-tuning large-scale Korean models on AWS infrastructure. Key features: - Fine-tuning of Generative AI foundation models using QLoRA - Local training and debugging for large-scale model tuning - Model serving using SageMaker DJL, TGI, and FasterTransformer Tech stack: AWS, SageMaker, DeepSpeed, DeepJava Library Why it matters: This lab provides practical guidance on how to quickly evaluate and deploy large-scale Korean models using AWS infrastructure, ensuring efficient and effective model tuning and serving.

Overview: This project provides a unified API for accessing multiple speech-to-text and natural language processing (NLP) providers, enabling developers to easily switch between different services without vendor lock-in. Key features: - Access to a wide catalog of speech-to-text and NLP providers - No vendor lock-in, allowing seamless switching between providers - Standardized transcript formats and open adapters for easy integration and future updates Tech stack: Node.js, TypeScript, AWS Why it matters: This API simplifies the process of building Conversational AI applications, providing developers with the flexibility to choose the best technology for each task and stay updated with the latest advancements from different providers.

### Overview: 这是一款强大的大模型微调数据集生成和管理工具，支持一键爬取指定域名的所有链接，将链接转换成大模型友好的Markdown文件，以及通过多种大模型将Markdown文件转换为训练大模型所需的数据集。 ### Key features: - 支持深度爬取指定域名的所有链接 - 支持将链接转换成大模型友好的Markdown文件 - 支持上传.md、.txt、.pdf、.docx、.doc等文件，自动转换成.md文件 - 支持智能算法对Markdown文件分段 - 支持通过DeepSeek、ChatGPT、Gemma等大模型将Markdown转换成训练大模型可用的数据集 - 支持数据集数据的自定义增加、编辑、修改 - 支持导出 JSONL 和 JSON 两种输出格式，Alpaca、ShareGPT 和自定义格式 - 支持预览转换结果 ### Tech stack: Python ### Why it matters: 这款工具简化了大模型微调数据集的生成和管理过程，提高了效率并减少了手动操作的需要，使得开发者能够专注于模型的训练和优化。

### Overview: Taris is an online speech recognition system that uses a Transformer architecture to dynamically segment spoken sentences and improve accuracy. ### Key features: - **Dynamic Segmentation**: Taris segments a spoken sentence into discrete segments based on the number of words it contains. - **Conditional Decoding**: The decoder uses a dynamic window of segments to condition on the entire utterance, improving accuracy. - **Audio-Visual Fusion**: Implements the AV Align strategy with a Transformer stack, replacing recurrent networks in the original architecture. ### Tech stack: Python ### Why it matters: Taris leverages the Transformer architecture to enhance online speech recognition, improving accuracy and efficiency.

### Overview: This project provides an easy-to-use interface for fine-tuning a pre-trained Stable Diffusion XL model using LoRA (Low-Rank Adaptation) with a custom dataset of fashion products collected from KREAM, a Korean online resale market. The dataset includes detailed captions for each product, making it ideal for generating realistic and high-quality fashion images. ### Key Features: - Custom dataset of fashion products from KREAM - Fine-tuning of Stable Diffusion XL using LoRA - Detailed captions for each product in the dataset - Easy-to-follow instructions for both dataset collection and fine-tuning ### Tech Stack: - Python 3.11 - XFormers - PyTorch 2.0.1 - Hugging Face Transformers, Diffusers, and Datasets ### Why It Matters: - Provides a straightforward way to fine-tune a pre-trained model for specific applications - Utilizes a custom dataset for better results - Facilitates the generation of realistic, high-quality fashion images Repository: hahminlew/fashion-product-generator GitHub Stars: 26 GitHub Description: Easy demo for finetuning a pre-trained Stable Diffusion XL with LoRA using the collected fashion dataset from scratch.

### Overview: GPT Image Studio 是一个本地优先的 AI 图片创作工作台。通过聊天式界面调用 OpenAI 兼容 Images API，生成和编辑图片。所有数据保存在本地，无需后端即可使用；需要保护 API 凭据或多人共用时，可接入本地 Companion 或部署为服务器模式。 ### Key features: - **图片创作**：支持文生图、图片编辑和遮罩编辑功能。 - **提示词控制**：提供安全、创意和开放模式，可在请求前注入对应模式说明和词库灵感。 - **本地存储**：所有数据本地存储，无需后端，保护 API 凭据或多人共用时可接入 Companion 或部署为服务器模式。 ### Tech stack: TypeScript, Vue 3, Tailwind CSS ### Why it matters: GPT Image Studio 是一个本地优先的 AI 图片创作工作台，通过聊天式界面调用 OpenAI 兼容 Images API，生成和编辑图片。所有数据保存在本地，无需后端即可使用；需要保护 API 凭据或多人共用时，可接入本地 Companion 或部署为服务器模式。

Overview: Katip is a macOS application designed for transcribing voice recordings using the SFSpeechRecognizer framework. It supports various audio formats and requires local transcription capabilities. Transcription speeds vary depending on the performance of the local macOS computer. Key features: - Transcribes voice recordings - Supports mp3, m4a, wav, and mpa formats - Requires local transcription capabilities - Estimated transcription time based on first run and subsequent runs Tech stack: Swift Why it matters: Katip simplifies the process of transcribing voice recordings, making it accessible for macOS users without the need for internet connectivity. Repository: i-celeste-aurora/katip Primary language: Swift GitHub stars: 26 GitHub description: A SFSpeechRecognizer-based voice recordings transcriber for macOS

### Overview: StreamUni is a framework designed to efficiently enable unified Large Speech-Language Models for streaming speech translation, achieving state-of-the-art performance on various tasks. ### Key Features: - Unified Large Speech-Language Models - Efficient Streaming Speech Translation - State-of-the-Art Performance on Streaming En-De and Simultaneous En-Zh Tasks ### Tech Stack: Python ### Why it Matters: StreamUni offers a comprehensive solution for real-time speech translation, making it highly valuable for applications requiring immediate translation capabilities.

Overview: This project provides Pepper Robot with basic conversational abilities, enabling it to engage in free, open-domain dialogue. Key features: - Speech Recognition Module - Conversational Engine - Speech Synthesis Tech stack: Python 2.7, Java, NAOqi SDK Why it matters: This system allows Pepper to interact naturally and pleasantly with users, serving as a fallback for general conversation or task-oriented interactions. Repository: igor-lirussi/Dialogue-Pepper-Robot GitHub stars: 26 GitHub description: Provides Pepper Robot with basic conversation abilities, including speech recognition, conversational engine, and speech synthesis.

### Overview: Midi-Tuning is an efficient Multi-round Interactive Dialogue Tuning framework that models agents and users separately using large language models. It leverages round-level memory caching to improve performance in dialogue tasks. ### Key features: - **Efficient Tuning**: Models agents and users individually with two adapters. - **Round-by-Round Tuning**: Adapters are tuned using respective utterances in alternating order. - **Memory Caching**: Rounds are cached to improve performance. - **Large Language Models**: Uses LLaMA-7B, Mistral-7B-v0.1, Vicuna-7B-v1.3, and Llama 2 Chat-7B. ### Tech stack: Python ### Why it matters: Midi-Tuning is designed to be efficient and effective in dialogue tasks, making it suitable for various applications where consistent and interactive dialogue is required.

### Overview: Mini-SGLang is a lightweight educational implementation of SGLang, focusing on the core mechanisms of modern LLM inference engines. It dissects key components like continuous batching, PagedAttention, RadixCache, and tensor parallelism into clean, Python code for developers to understand and learn from. ### Key Features: - **Continuous Batching**: Prefill / Decode scheduling for maximum throughput. - **PagedAttention**: Paged KV cache management to avoid GPU memory fragmentation. - **RadixCache**: Prefix-aware caching with automatic shared-prefix reuse. - **CUDA / NPU Graph**: Elimination of decode-phase kernel launch overhead. - **Tensor Parallelism**: Column / Row Parallel Linear layers ready; multi-process TP launch not yet implemented. - **Multi-device Support**: NVIDIA CUDA / Huawei Ascend NPU / CPU with automatic priority detection. - **Pluggable Attention Backends**: FlashAttention (auto-falls back to PyTorch SDPA) / PyTorch SDPA. - **OpenAI-compatible API**: `/v1/chat/completions` + `/v1/completions` + SSE streaming. ### Tech Stack: Python ### Why it Matters: Developers can learn the intricate workings of an LLM inference engine by understanding the core mechanisms implemented in Mini-SGLang. This project provides a clear, line-by-line explanation of how modern LLMs work, making it an invaluable resource for developers and researchers alike.

### Overview: X2HDR is a NeurIPS 2026 project that introduces a method for generating high-dynamic-range (HDR) images in a perceptually uniform space. The technique leverages a pretrained model and various LoRAs (Low-Resolution Models) to enhance image quality and realism. ### Key Features: - **Text-to-HDR Image Generation**: Converts textual descriptions into high-quality HDR images. - **RAW-to-HDR Image Reconstruction**: Recovers raw images into HDR format. - **LDR-to-HDR Image Reconstruction**: Converts low-dynamic-range (LDR) images to HDR. - **Pretrained Models**: Utilizes FLUX.1-dev and various LoRAs for improved performance. ### Tech Stack: - **Python**: The primary programming language used for the project. - **Torch**: Used for training and inference processes. ### Why It Matters: X2HDR offers a significant improvement in image quality and realism, making it particularly useful for applications requiring high-dynamic-range visuals, such as virtual reality, augmented reality, and photorealistic rendering in video games and film production. ### Repository: - **GitHub Repository**: kingnobro/X2HDR - **Stars**: 25 - **Description**: A project from NeurIPS 2026 focusing on HDR image generation in a perceptually uniform space. ### Setup: - **Environment**: Requires `conda` and `pip` to be installed. - **Models**: Downloads pretrained models and LoRAs for text-to-HDR, RAW-to-HDR, and L

### Overview: This project focuses on optimizing the scheduling and transfer paths in the SGLang runtime on NVIDIA H100 GPUs, specifically using C++ descriptors, Triton, and pinned memory. The primary goal is to reduce overhead and improve performance by consolidating and optimizing these paths, resulting in significant latency reductions and throughput enhancements. ### Key Features: - **Pinned Transfers:** Reduced preparation latency by 51.1%–54.0% for cache-hit batches and 21.9%–43.1% for longer prefills. - **Expert Scheduling:** Improved throughput by 3.9%–5.1% for ordinary-call serving. - **Verification Transfer:** Enhanced throughput by 2.25× for the largest two-rank case, with a throughput change of −0.9% to +9.4% across nine serving configurations. ### Tech Stack: - Python ### Why It Matters: This project significantly improves the performance of SGLang by optimizing the scheduling and transfer paths, leading to substantial reductions in latency and throughput improvements, particularly in cache-hit scenarios and serving operations.