Projects

Overview: hnswlib-wasm is a WebAssembly (Wasm) version of the hnswlib index library, designed to be used in browsers. It aims to provide a browser-friendly implementation of hierarchical navigable small world graphs for approximate nearest-neighbor search. Key features: - WebAssembly (Wasm) port of hnswlib - Supports indexing and searching using the cosine distance metric - Utilizes IndexedDB (IDBFS) for efficient storage and retrieval of the index - Provides TypeScript definitions for better compatibility with Wasm - Supports both browser and Emscripten environments Tech stack: C++ Why it matters: hnswlib-wasm simplifies the use of hnswlib in the browser, making it accessible for applications that require approximate nearest-neighbor search in the web environment.

### Overview: AI Book Generator is a project that helps developers kickstart their book writing process with the assistance of artificial intelligence. It provides a step-by-step or fully automated approach to generating complete, publication-ready books. ### Key features: - **Multi-LLM Support**: Supports OpenAI GPT (4o, o4-mini, GPT-5), Anthropic Claude (3.5 Sonnet, 3.5 Haiku, Opus, Sonnet 4.5, Opus 4.5), and Google Gemini (1.5 Flash/Pro, 2.0 Flash, 2.5 Pro). - **Agent Mode**: One-click "Generate Complete Book" button that autonomously generates title suggestions, book concept, chapter outline, and all chapter content in sequence. - **Step-by-Step Workflow**: Manual control with individual steps: concept → outline → chapters. - **Rich Init-Prompting**: Structured, chain-of-thought system prompts that produce higher-quality, more consistent output. - **Live Streaming**: Watch chapters appear word-by-word as they're generated (OpenAI Chat API). - **Cover Image Generation**: AI-generated cover art via OpenAI Images API. - **Multiple Export Formats**: TXT, HTML, Markdown, JSON, PDF, and clipboard copy. - **Dark/Light Theme**: Auto-detected or manually toggled. - **Auto-Save & Drafts**: Projects are automatically saved to browser local storage. - **Keyboard Shortcuts**: `Ctrl+S` (save), `Ctrl+G` (generate

Overview: This repository contains a script designed to train a Latent Diffusion Model on the MIMIC-CXR dataset using MONAI Generative Models. The project aims to generate synthetic images of chest X-rays for medical research and training purposes. Key features: - Training a Latent Diffusion Model on the MIMIC-CXR dataset - Preprocessing of images and reports - Training the model using MONAI Generative Models - Inference and evaluation of the model Tech stack: Python Why it matters: This project provides a valuable tool for generating synthetic images of chest X-rays, which can be used for medical research, training, and validation of medical imaging models.

Overview: WaveSpeed nodes for ComfyUI enable the integration of hundreds of AI models into workflows, providing a unified interface for multi-modal generation and performance optimization. Key features: - Support for 20+ AI generation categories including Text to Image, Image to Image, Text to Video, Video to Video, Text to Audio, Image to 3D, and more. - A single predictor node that accesses 600+ AI models through one node. - Dynamic parameters auto-configure, making model-specific adjustments easy. - Fuzzy search for quick model discovery. - Category tabs for easy browsing by type. - Smart caching for fast model loading. - Real-time progress indicators for generation. - Workflow support for saving and restoring complete workflows. - Connection flexibility for any compatible input/output. Tech stack: JavaScript Why it matters: This plugin allows users to seamlessly integrate a wide range of AI models into ComfyUI workflows, enhancing their creativity and productivity by providing a unified interface and optimized performance. Repository: WaveSpeedAI/wavespeed-comfyui GitHub stars: 65

### Overview: Zengram is a multi-agent memory management system designed for shared memory across multiple AI agents running on different machines. It leverages a single Postgres database to store and retrieve events, facts, statuses, and decisions, ensuring that all agents can access and share information seamlessly. ### Key Features: - **Typed Memory**: Events, facts, statuses, and decisions are stored in a structured manner, with each type having its own lifecycle and decay rules. - **Unified Storage**: All memory types are stored in a single Postgres database, ensuring efficient querying and retrieval. - **Multi-Path Search**: The system supports three retrieval paths: vector search, full-text search, and entity graph, providing a robust search experience. - **Cross-Encoder Reranker**: An optional cross-encoder reranker enhances precision by reading query and document together, improving recall and precision. ### Tech Stack: - **Node.js**: The primary programming language used to build the system. - **Postgres**: The database used to store and retrieve memory types. - **pgvector**: A vector search engine for semantic vector search. - **tsvector**: Postgres's full-text search engine for text-based queries. ### Why It Matters: Zengram addresses the common problem of agents in multi-agent systems losing information between sessions. By providing a shared memory system, it ensures that all agents can access and benefit from each other's discoveries, enhancing collaboration and efficiency in complex AI environments.

Overview: This repository contains the latest version of vLLM, a native Windows build, optimized for Windows 10/11, with support for CUDA 13.0 and Python 3.13. It includes features such as FlashAttention/Rust, Multi-TurboQuant, and experimental CPU/NVMe KV offload, making it suitable for high-performance computing tasks. Key features: - Native Windows build of vLLM 0.29.0 - Python 3.14/CUDA 13.2 release (CPU TorchAudio) - Python 3.13/CUDA 13.0 prerelease - PyTorch 2.13, FlashAttention/Rust, OpenAI serving, 10 KV formats, Multi-TurboQuant, experimental CPU/NVMe KV offload Tech stack: Python 3.14/CUDA 13.2 Why it matters: This repository provides a pre-built wheel for vLLM 0.27.1, optimized for Windows 10/11, with support for CUDA 13.0 and Python 3.13. It includes features such as FlashAttention/Rust, Multi-TurboQuant, and experimental CPU/NVMe KV offload, making it suitable for high-performance computing tasks.

### Overview: Frink is a pure-Rust GGUF inference engine that supports inference on CPU, Apple Metal, and CUDA. It includes quantized CPU, Metal, and CUDA kernels, and supports MoE. Frink is designed to be benchmarked against LLaMA.cpp, offering comparable performance and features. ### Key Features: - **Quantized Inference**: Runs on quantized CPU, Metal, and CUDA kernels. - **MoE Support**: Supports Multi-Head Embedding (MoE) for efficient inference. - **OpenAI-Compatible Server**: Compatible with OpenAI's API, including support for concurrent clients and paged KV. - **Quantized Weights**: Uses K-quants, IQ tiers, MXFP4, F16, and BF16 formats, ensuring efficient memory usage. - **Drop-in Compatibility**: Works seamlessly with LLaMA.cpp, maintaining the same flags and toolchain. - **Server Features**: - OpenAI-compatible server with continuous batching, slot save and restore, and speculative decoding. - Supports 102 architectures, measured at 1.28x aggregate throughput and 736 of 757 prompt tokens reused. - Logit comparison and lossless resumable streams. - Silent handling of unsupported fields, ensuring the model cannot produce invalid outputs. ### Tech Stack: - **Rust**: The entire engine is written in Rust, ensuring performance and reliability. ### Why It Matters: Frink is a pure-Rust implementation of a GGUF inference engine that supports inference on multiple hardware

Overview: The Nano Banana Pro (official) API is an OpenAI-compatible endpoint that provides image generation and preview services for Nano Banana Pro, offering flexible pricing and seamless integration. Key features: - Supports both text and image inputs - Offers tiered pricing based on resolution and number of images - Provides detailed documentation and examples for developers Tech stack: Python Why it matters: This API simplifies integration with Nano Banana Pro, making it accessible through a familiar OpenAI-compatible endpoint, and allows for flexible billing based on usage. Repository: apimartbv/gemini-3-pro-image-preview-official-api-fr Primary language: Python GitHub stars: 64 GitHub description: Nano Banana Pro (official) API (`gemini-3-pro-image-preview-official`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint. README excerpt: The Nano Banana Pro (official) API is an OpenAI-compatible endpoint that provides image generation and preview services for Nano Banana Pro, offering flexible pricing and seamless integration.

Overview: This repository offers Generative AI and AI/ML code samples, blueprints, and proof-of-concepts tailored for the Latin American market. Key features: - Multipage Document Analysis - Contract Compliance Analysis - Smart Product Onboarding - RFP Answers with Generative AI Tech stack: Python Why it matters: This repository provides practical solutions and blueprints for developers looking to implement Generative AI and AI/ML technologies in the Latin American market, focusing on efficiency and relevance in regulatory and compliance-heavy industries. Repository: aws-samples/generative-ai-ml-latam-samples Primary language: Python GitHub stars: 64

### Overview: This project introduces a Keras implementation for translating text into images using Generative Adversarial Networks (GANs) and Word2Vec embeddings. The repository includes three versions of the DCGan model: one with noisy input, one without noise, and one with a configurable amount of noise. The models are trained on a dataset of Pokémon images and texts, allowing for the generation of new Pokémon images based on textual descriptions. ### Key Features: - **DCGAN with Noise**: Uses half of the input as pure noise and the other half as a glove embedding of the input text. - **DCGAN without Noise**: Removes the noise from the input, keeping only the glove embedding. - **DCGAN with Noise Configurable**: Allows for a configurable amount of noise in the input. ### Tech Stack: - Keras - GANs - Word2Vec embeddings - Recurrent neural networks ### Why it Matters: This project provides a practical demonstration of how to use Keras to create an image generation model that can translate textual descriptions into visually coherent images. It showcases the integration of GANs with other neural network techniques, specifically Word2Vec, to enhance the quality of generated images. The ability to generate new Pokémon images based on textual descriptions is particularly interesting and could have applications in various fields, such as creative writing or art generation.

### Overview: This repository contains the **text-to-video (Wan2.1-1.3B)** and **image-to-video (Cosmos-Predict2.5-2B)** experiments for **Data-Forcing Distillation (DFD)**. DFD is a simple post-training framework designed to restore diversity and fidelity in few-step video generators, addressing the issue of collapsed sample diversity and over-saturation in DMD/DMD2. It achieves this by adding explicit real-data supervision to the distribution-matching loss, pulling the student towards modes it has missed and away from over-saturated modes absent in real data. ### Key Features: - **Text-to-Video (Wan2.1-1.3B)** - **Image-to-Video (Cosmos-Predict2.5-2B)** ### Tech Stack: Python ### Why It Matters: This codebase is adapted from [NVIDIA FastGen](https://github.com/NVlabs/FastGen). The autoregressive (Self-Forcing) experiments live in a separate repository: [DFD-self-forcing](https://github.com/csy2077/DFD-self-forcing). ### One-Line Change: DFD stochastically replaces the teacher-score input with a matched **real** video. In FastGen, this is exposed as a single flag, `model.gen_real_replace_prob` (the mixing probability `p`); the rest of the DMD2 pipeline is unchanged. ### Repository Structure: ``` fastgen/ ├── fastgen

### Overview: Graph Distillation for Action Detection This repository contains the code for the paper "Graph Distillation for Action Detection with Privileged Modalities" presented at ECCV 2018. The method, termed graph distillation, incorporates rich privileged information from a large-scale multi-modal dataset to improve learning in the target domain where training data and modalities are scarce. ### Key Features: - **Graph Distillation**: A technique that distills information from a large-scale multi-modal dataset to enhance learning in a target domain. - **Action Detection**: The primary application of the method, focusing on action detection tasks. - **Multi-Modal Data**: Utilizes a variety of modalities such as RGB, depth, and semantic segmentation for improved performance. ### Tech Stack: Python ### Why It Matters: This method offers a significant improvement in action detection by leveraging privileged information from a large dataset, making it particularly useful in scenarios with limited training data and diverse modalities. ### Repository: google/graph_distillation ### Primary Language: Python ### GitHub Stars: 64 ### GitHub Description: Graph Distillation for Action Detection with Privileged Modalities ### Pretrained Models: - Available for download via `scripts/download_models.sh` or `gsutil` commands. ### Running Models: - **Classification**: - Scripts: `classification/run.py` - Arguments: `--xfer_to` for the modality being trained, `--modalities` for the modalities to distill from. - **Detection**: - Scripts: `detection/run

Overview: 将微信语音输入整合为一个Windows应用，提供OpenAI兼容的音频转写API，不发送语音给OpenAI。 Key features: - Windows应用 - OpenAI兼容的音频转写API - 本地音频转写 - 三个并发池 - 自定义并发数 - 调度独立性 - 原生客户端与私有网页协议 Tech stack: Python Why it matters: 提供一个Windows上的语音转写API，简化了语音识别的集成，适用于需要本地音频转写需求的开发者。

Overview: This project is a web application that utilizes Whisper ASR and ChatGPT API to provide voice-based assistance for users interested in multiple languages. Key features: - Transcribes spoken input into text - Sends requests to ChatGPT API to obtain answers - Speaks out the generated text using computer synthetic voice Tech stack: Python Why it matters: This project showcases the integration of advanced speech recognition and language generation technologies, offering a practical tool for users seeking accurate language translations and AI-driven responses.

Overview: PromptML is a domain-specific language (DSL) designed specifically for AI systems, built to enable prompt engineers to define AI prompts in a deterministic and explicit manner. Key features: - Standardization instead of fragmentation - Collaboration instead of confusion - Enabling version control-ability - Promoting verbosity for better results Tech stack: Python Why it matters: PromptML simplifies the process of defining AI prompts, making them more predictable and easier to manage, thus improving the reliability and effectiveness of AI systems.

### Overview: Oracle LangChain integration provides official LangChain support for Oracle’s Generative AI and AI Platform services, enabling developers to build intelligent, secure, and scalable applications powered by Oracle’s cloud technologies. ### Key Features: - **OCI Generative AI & Data Science**: Chat models (`ChatOCIGenAI`, `ChatOCIOpenAI`, `ChatOCIModelDeployment`), embeddings (`OCIGenAIEmbeddings`), agents (`create_oci_agent`, `create_deepagents_agent`), guardrails (`OCIGuardrails`), and multimodal (vision, PDF, video, audio) support across Cohere, Google, Meta, OpenAI, and xAI models on OCI. - **Oracle AI Vector Search**: Vector store (`OracleVS`), hybrid & text-search retrievers, document loaders (`OracleDocLoader`), text splitter, in-database embeddings and summaries, semantic cache, and chat message history — all powered by Oracle Database. ### Tech Stack: Python ### Why it Matters: This integration allows developers to leverage Oracle’s cloud technologies to build intelligent applications, ensuring secure and scalable solutions.

### Overview: Hierarchical Continuous Diffusion Language Models (HC-DLM) is an innovative approach to continuous diffusion language models that combines the strengths of discrete and continuous diffusion methods. By keeping both strengths, HC-DLM ensures that tokens are planned jointly in the latent space and that the latent is continuously guided by tokens at every step, enabling more flexible and context-aware language generation. ### Key Features: 1. **Joint Token Planning**: Tokens are planned jointly in the shared latent space, ensuring that the model can handle complex language tasks more effectively. 2. **Continuous Latent Update**: At every step, tokens are read out of the latent space and guide the next update, allowing for more flexible and context-aware language generation. 3. **Interactive Demonstration**: A simulated example of HC-DLM's reverse trajectory on a sentence is provided, demonstrating how tokens guide the latent update at every step. ### Tech Stack: The codebase is written in Python and utilizes libraries such as PyTorch and Hugging Face for training and inference. ### Why It Matters: HC-DLM addresses the limitations of discrete and continuous diffusion methods by providing a more flexible and context-aware approach to language generation. By keeping both strengths, HC-DLM enables more effective and context-sensitive language modeling, making it a significant advancement in the field of continuous diffusion language models.

### Overview: Med-Banana is an end-to-end system designed to learn quality-controlled medical image editing from success-and-failure trajectories. The system consists of an editor, verifier, and refiner, each trained on specific trajectories to ensure the edits are both medically and clinically valid. ### Key Features: - **Med-Banana-80K Dataset**: The official dataset used for training and evaluation, comprising 80,000 image records with 50,635 accepted edits and 37,822 rejected attempts. - **Editor/Verifier/Refiner Pipeline**: A Gemini-based pipeline that constructs the Med-Banana-80K dataset, trains the editor, verifier, and refiner, and provides an edit-verify-refine inference runtime. - **Inference and Evaluation**: The system includes an evaluation suite that provides insights into the quality and effectiveness of the edits. ### Tech Stack: - **Python**: The primary programming language used for implementing the system. ### Why It Matters: Med-Banana offers a novel approach to medical image editing by focusing on the failures as well as the successes. It uses the failures to teach quality control, ensuring that the edits are not only correct but also medically and clinically appropriate. This method provides a more comprehensive and effective learning mechanism compared to existing datasets that only supervise with final accepted edits. ### Repository: - **Repository Name**: richardChenzhihui/Med-Banana-80K - **GitHub Stars**: 65 - **GitHub Description**: Official code release

Overview: Vectra Core is an ultra-fast in-memory vector database and Approximate Nearest Neighbor (ANN) search engine built in Java 21 (Project Loom Virtual Threads). It implements the Hierarchical Navigable Small World (HNSW) graph indexing algorithm alongside 8-bit Scalar Quantization (SQ8) and Metadata Predicate Filtering, delivering sub-millisecond query latencies and over 16,000 QPS on high-dimensional embedding workloads. Key features: - Sub-millisecond query latencies - Over 16,000 QPS - 8-bit Scalar Quantization (SQ8) for 75% RAM reduction - Metadata Predicate Filtering for improved efficiency Tech stack: Java 21 (Project Loom Virtual Threads) Why it matters: Vectra Core is designed for high-dimensional vector workloads and offers exceptional performance, making it suitable for applications requiring fast and accurate nearest neighbor search.

### Overview: Fine-tuning Stable Diffusion using Keras. ### Key features: - Fine-tuning of Stable Diffusion model using Keras. - Adapted from Hugging Face script for text-to-image fine-tuning. - Uses KerasCV's Stable Diffusion model for fine-tuning. - Supports fine-tuning on two resolutions: 256x256 and 512x512. ### Tech stack: Python ### Why it matters: This repository provides a way to fine-tune Stable Diffusion using Keras, which can be useful for researchers and developers interested in applying machine learning to image generation tasks.