Projects

### Overview: This project provides runnable examples for the Dreamina API, a third-party REST API for Dreamina (CapCut) by ByteDance. It allows you to generate videos using Seedance 2.5, 2.0, 2.0 Fast, 2.0 Mini, Seedance 1.0 Pro (CA only), 1.0 Mini (US only), 1.0 Fast (CA only), HappyHorse 1.1 by Alibaba (CA only), and Wan 3.0 by Tongyi Lab (CA only), as well as images with Seedream 5.0 Pro (CA only), 5.0 Lite, 4.7 (CA only), 4.6, 4.5, 4.1, 4.0, and 3.0, Nano Banana Pro, Nano Banana (CA only), GPT Image 2.5 Flare and Sunburst, and GPT Image 2. CA only means a Canada-region Dreamina account, and US only a US one. ### Key Features: - Batch-generate Seedance 2.5 and 2.0 video: text-to-video and first/last-frame image-to-video - Batch-generate Seedream, Nano Banana, and GPT Image pictures with optional reference images - Supports multiple Dreamina accounts with different regions ### Tech Stack: - Node.js v21 or newer ### Why It Matters: This project is useful for developers who want to automate video generation using the Dreamina API,

### Overview: Libravdb is a persistent local memory engine for OpenClaw, designed to replace the default memory system with a full context lifecycle. It includes hybrid vector recall, automatic compaction, and domain-adaptive gating, providing enhanced cognitive memory capabilities. ### Key Features: - **Causal Graph Reasoning**: Enables causal graph-based reasoning and identity tracking. - **Keyword-Based Reply Rules**: Facilitates efficient keyword-based queries and replies. - **Local Inference Support**: Provides local inference capabilities for faster responses. ### Tech Stack: TypeScript ### Why It Matters: Libravdb offers advanced cognitive memory features that enhance the performance and reliability of OpenClaw agents, making them more effective in complex cognitive tasks. ### Repository: - **Repository**: xDarkicex/openclaw-memory-libravdb - **GitHub Stars**: 10 - **GitHub Description**: Persistent local memory engine for OpenClaw — replaces default memory-system with a full context lifecycle: hybrid vector recall, automatic compaction, and domain-adaptive gating over LibraVDB ### Primary Language: TypeScript ### GitHub Description: <div align="center"> <img src="./docs/assets/libravdb-logo.svg" alt="LibraVDB" width="480"> <p><strong>Local-first memory kernel with causal graph reasoning, identity tracking, and keyword-based reply rules.</strong><br/> <em>768-dim vector search • 6-tier cognitive classification • Local inference supported</em></p> </div>

### Overview: This project is an AI-powered quantitative trading platform designed for US stocks. It leverages Alpaca Markets API and DeepSeek LLM to provide intelligent trading decisions, risk management, and strategy execution. ### Key features: - **💰 Account Management**: Real-time account information, portfolio value, buying power, and equity tracking. - **📊 Position Management**: View all positions with real-time P&L, entry prices, and market values. - **⚡ Order Management**: Place buy/sell orders with multiple order types (market, limit, stop, stop-limit). - **🤖 AI Trading Decisions**: DeepSeek LLM-powered analysis providing buy/sell/hold recommendations with confidence scores. - **🎯 Strategy Management**: Create, manage, and execute trading strategies with automated monitoring. - **🔄 Auto Trading**: Automated trading execution based on AI decisions and strategy rules. - **🛡️ Risk Management**: Stop loss and take profit monitoring with automatic alerts. - **📋 Trade Records**: Complete history of all trades and trading signals. - **⚙️ Configuration**: Easy web-based configuration for API keys and trading settings. ### Tech stack: Python ### Why it matters: This platform offers a comprehensive solution for quantitative traders, providing advanced AI-driven trading decisions, risk management, and strategy execution, all within a user-friendly interface.

### Overview: AI-Memory is a biomimetic memory management framework designed for AI agents, implementing a human-like STM/LTM funnel, cognitive filtering, and Ebbinghaus forgetting curve. It uses Go and Vue technologies to manage memory lifecycle, ensuring efficient and effective information retention and recall. ### Key Features: - **Biomimetic Funnel Architecture**: Mimics human memory processes with three-tier filtering: STM, Staging Area, and LTM. - **Intelligent Value Judgment**: Uses LLM to evaluate memory importance, relevance, and uniqueness, and applies recurrence validation, time window verification, and LLM-based scoring. - **Semantic Deduplication**: Prevents duplicate memories from entering the funnel and ensures uniqueness before final storage. - **Automatic Decay & Forgetting**: Simulates natural memory decay over time, with configurable half-life and auto-cleanup for low-value memories. - **Monitoring & Dashboard**: Real-time visibility into the memory system's health and performance, including metric tracking, visual trends, and system status checks. - **Intelligent Alert System**: Automatically monitors health and detects potential issues, with dynamic rule configuration and built-in rules for queue backlog detection, low promotion success rate, cache hit rate anomalies, and memory decay spike detection. ### Tech Stack: Go ### Why it Matters: AI-Memory addresses critical memory challenges faced by traditional AI conversation systems, such as full retention being expensive, inability to distinguish valuable insights, and starting from scratch every conversation. By mimicking human memory processes and implementing a biologically-inspired architecture, it enables AI agents to intelligently filter,

**Overview:** ofxLlamaCpp is an openFrameworks addon that enables running large language models (LLMs) directly in C++ using llama.cpp, providing a seamless integration for creative coding projects. **Key features:** - Seamless integration of llama.cpp for LLM inference - Minimal dependencies, focusing solely on C++ and llama.cpp - Support for GGUF models and remote OpenAI-compatible APIs - Examples demonstrating both local and remote inference modes **Tech stack:** C++ **Why it matters:** This addon simplifies the integration of LLMs into openFrameworks projects, making it easier to create interactive and generative applications without the need for Python or external dependencies.

### Overview: This project provides a Python wrapper class for the OpenVINO Model Server (OVMS), allowing users to submit inference requests with minimal code. It simplifies the process of interacting with OVMS, making it accessible for developers without extensive knowledge of TensorFlow or TensorFlow Serving. ### Key Features: - **Wrapper Class**: A Python class for interacting with OVMS. - **Model Repository**: Automatically generates a model repository for OVMS using a single command. - **Alternative Solution**: Provides a bundled GRPC handler for non-IA clients, eliminating the need for TensorFlow and TensorFlow Serving API. - **Demo Programs**: Demonstrates how to use the wrapper class with a sample client inference code. ### Tech Stack: - Python ### Why It Matters: This project simplifies the integration of OVMS into applications, making it easier for developers to deploy and utilize AI inference services without the complexity of setting up TensorFlow or TensorFlow Serving.

### Overview: AI Forge 是一个专注于 AI 工程实践的仓库，包含文章、笔记本和实验，旨在研究模型微调、LangChain、LangGraph、RAG、MCP、Sec for AI、Agent 工程化、记忆、工具调用与安全控制等方面。 ### Key features: - 模型微调 - LangChain - LangGraph - RAG - MCP - Sec for AI - Skills - Model - Coding ### Tech stack: Jupyter Notebook ### Why it matters: 通过持续的研究和实践，帮助开发者掌握 AI 工程的关键技术和最佳实践，提高开发效率和项目质量。 ### Repository: yauld/ai-forge ### GitHub stars: 9 ### GitHub description: AI Engineering 实战学习仓库

Overview: nano-llama.cpp is a small, readable, and extensible C++/CUDA LLM inference engine designed for Qwen3 and Qwen3.5 models, featuring fast performance, multimodal support, developer-friendly architecture, continuous-batching server, and lightweight dependencies. Key features: - Fast performance on a single GPU, equivalent to the reference ggml runtime. - Multimodal support with an integrated ViT vision encoder. - Developer-friendly codebase with a small, readable 2,730 lines of C++. - Continuous-batching server for OpenAI-compatible chat completions. - Backend support for CPU and CUDA. - No heavy dependencies, using vendored ggml and two small libraries (HTTP + JSON). Tech stack: C++/CUDA. Why it matters: nano-llama.cpp offers a lightweight, developer-friendly solution for LLM inference, making it suitable for applications requiring fast performance and easy integration into existing projects.

Overview: vecdb is a Rust library that provides an embeddable vector database, supporting dense, sparse, hybrid, and filtered search. It is designed to be used directly in applications without the need for a separate server. Key features: - **Dense Search**: HNSW (Instant Distance) for fast nearest neighbor queries. - **Sparse Search**: BM25 inverted index for efficient full-text search. - **Hybrid Search**: Fusion of dense and sparse results for optimal recall. - **Filtered Search**: Predicate pushdown on JSON metadata for selective filtering. - **Quantization**: Opt-in int8 or 1-bit binary indexing for reduced index size. - **Durable Indexing**: Write-ahead log for crash recovery and persistence. - **Embedded**: No dependencies beyond Rust, SQLite for metadata. Tech stack: Rust. Why it matters: vecdb simplifies vector search by providing a fast, in-process database that can be easily integrated into applications, eliminating the need for a separate server or database management system. Repository: zaydmulani09/vecdb GitHub stars: 10 GitHub description: An embeddable vector database in Rust — no server, SQLite-style. HNSW + BM25 + hybrid + filtered search, in-process.

### Overview: SLO Router is a high-performance LLM inference router designed to optimize latency and cost for applications with strict service level objectives (SLOs). It leverages Jev decision-making to route requests to the backend with the lowest cost and highest quality that meets the SLO. ### Key features: - **SLO-aware Routing**: Routes requests to the backend with the lowest cost and highest quality that meets the SLO. - **Jev Decisions**: Utilizes Jev 1.13 for decision-making, providing bounded semantic features. - **Live Queue Metrics**: Tracks and evaluates the live queue depth and context capacity of the backends. - **Counterfactual Evaluation**: Allows for the analysis of different routing scenarios to understand their impact. - **Reproducible Benchmarks**: Offers benchmarks with live queue metrics, counterfactual collection, and trace replay paths for accurate performance measurement. ### Tech stack: - Python ### Why it matters: SLO Router is essential for applications that require high levels of service quality and low latency, ensuring that requests are routed efficiently and cost-effectively while meeting the defined SLOs. The ability to dynamically adjust routing based on the SLO and backend metrics makes it highly valuable for real-time applications and microservices architectures.

### Overview: StyleBlend is an official implementation of a finetuning-based approach for enhancing style-specific text-to-image generation in text-to-image diffusion models. This method ensures robust style consistency and aligns textual semantics, making it particularly useful for creating consistent and contextually relevant images. ### Key Features: - **Robust Style Consistency:** Ensures that the generated images maintain the specified style throughout the process. - **Aligns Textual Semantics:** Ensures that the generated images are semantically consistent with the provided text. - **Flexible Training and Inference:** Supports both training and inference using the provided scripts and configuration files. ### Tech Stack: Python ### Why it Matters: StyleBlend is a significant contribution to the field of text-to-image diffusion models, offering a practical solution for creating style-specific and contextually relevant images. Its robust style consistency and aligned textual semantics make it a valuable tool for various applications, including art generation and creative design. Repository: zichongc/StyleBlend GitHub stars: 10 GitHub description: [Eurographics '25] Official Implementation of StyleBlend: Enhancing Style-Specific Content Creation in Text-to-Image Diffusion Models

Overview: This repository contains code and datasets for the CCL 2023 paper "Revisiting k-NN for Fine-tuning Pre-trained Language Models". The project aims to revisit k-NN classifiers for augmenting PLMs-based classifiers, proposing a method to leverage k-NN predicted probabilities as indications for easy vs. hard examples during training. Key features: - Revisiting k-NN classifiers for fine-tuning pre-trained language models. - Proposed method to use k-NN predicted probabilities as indications for easy vs. hard examples during training. - Extensive experiments on fine-tuning and prompt-tuning across 8 diverse end-tasks. Tech stack: Python Why it matters: This project introduces a novel approach to fine-tuning pre-trained language models by leveraging k-NN classifiers, which can improve the performance of PLMs-based classifiers in various tasks. Repository: zjunlp/Revisit-KNN GitHub stars: 10 GitHub description: [CCL 2023] Revisiting k-NN for Fine-tuning Pre-trained Language Models

### Overview: OpenTQ is an open-source tool designed to simplify and audit the quantization process for low-bit weight formats, making it easier to release quantized models while ensuring reproducibility and transparency. ### Key Features: - **Stock GGUF Deployment:** Supports the stock-compatible `Qwen3.6-27B-OTQ-GGUF` release, allowing users to run quantized models with stock `llama.cpp`. - **Transparent Allocation:** Norms and states remain high precision, while projection-heavy families absorb most compression. - **Practical Quality Checks:** Pairs BF16 vs GGUF mini-subsets with pinned task IDs and public reproducibility data. - **Runtime Gates:** Local Apple Silicon checks with `llama.cpp`/Metal, bounded generation, release evals, and 8K prefill/decode measurements. - **Custom Policies:** Users can define their own dynamic allocation policy with YAML/JSON via `--policy-file`. - **Run Monitor:** Long quantization jobs can be watched through the terminal dashboard and machine-readable status output. - **Research Runtime Tracks:** Separate native OpenTQ payloads and compressed-domain runtime work are tracked from stock GGUF releases. ### Tech Stack: Python ### Why It Matters: OpenTQ simplifies and ensures the reproducibility of quantized models, making it easier to release and audit quantized versions of models while maintaining the quality of the original models.

Overview: Memorable-AI is a memory system that automatically remembers everything with just 2 lines of code, using research-backed techniques and a production-ready architecture. It integrates seamlessly with any SQL database and supports 100+ LLM models, including optional graph capabilities. Key features: - Zero-code integration - Research-validated techniques - Optional graph support - Hybrid retrieval - SQL-first storage - Multi-model support Tech stack: Python Why it matters: Memorable-AI simplifies memory management for developers, ensuring that their applications can learn from conversations and previous interactions without the need for extensive code changes. Repository: 0sparsh2/memorable-ai GitHub stars: 9

### Overview: Unlocking the full potential of the NVIDIA CMP 100-100 mining card from 6 GB to 12 GB HBM2 on Linux. ### Key Features: - **Unlocking 12 GB HBM2:** The card's driver reports 12,288 MiB instead of the default 6144 MiB. - **PCIe and HBM Support:** The project aims to achieve Gen2 and physical x16 functionality, including the fourth HBM site. - **Linux Compatibility:** The solution works on Linux x86_64 with glibc >=2.38, NVIDIA driver, and standard PCI tools. ### Tech Stack: - **Python 3.8+ Controller:** The project uses a Python 3.8+ controller for initialization and management. - **NVIDIA Driver:** The driver is required for the card to report the correct memory size. ### Why It Matters: Unlocking the full potential of the CMP 100-100 mining card can significantly enhance its performance and utility, making it more suitable for high-performance computing tasks.

### Overview: This project implements a deep learning model called StackGAN to generate photo-realistic images based on textual descriptions. The model aims to simplify the concept of attention mechanisms, making it more accessible for generating images with modern tools. ### Key Features: - **StackGAN Implementation**: Directly uses the StackGAN model from the paper "StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks" by Han Zhang et al. - **Textual Descriptions**: Generates images given textual descriptions, focusing on simplifying the attention mechanism. - **Visual Descriptions**: Supports visual descriptions for generating images. - **GitHub Repository**: Contains detailed documentation and dependencies for easy implementation. ### Tech Stack: - Python 3.0 - Pytorch - torchfile - NLTK-('punkt') - Pandas - Scikit-learn - Python-dateutil - EasyDict - Shutil - Matplotlib - Scikit-Image(skimage) ### Why It Matters: This project simplifies the concept of attention mechanisms, making it easier to generate images based on textual descriptions using modern tools. It aims to bridge the gap between complex attention mechanisms and simpler, more accessible techniques, making it valuable for developers and researchers in the field of image generation and text-to-image synthesis.

Overview: TeEFusion is a simple yet powerful distillation method that fuses classifier-free guidance directly into text embeddings, enabling student models to mimic complex teacher sampling strategies with a single forward pass. Key features: - Incorporates guidance magnitude *w* by linearly combining conditional and null prompt embeddings. - Distills from complex teacher sampling strategy (Euler + CFG, Z-Sampling + CFG, W2SD + CFG) into a simple student that uses standard Euler sampling. - No extra network modules beyond the pretrained model’s encoder and decoder. Tech stack: Python Why it matters: TeEFusion achieves up to 6× faster inference without sacrificing image quality, making it highly efficient and practical for real-time applications.

Overview: This application provides comprehensive examples of the llmedge library, showcasing language model inference, retrieval-augmented generation, vision and multimodal processing, generative media, and speech processing. Key features: - Language Model Inference - Retrieval-Augmented Generation - Vision and Multimodal Processing - Generative Media - Speech Processing Tech stack: Kotlin Why it matters: These examples demonstrate the llmedge library's capabilities and best practices for efficient on-device inference and advanced AI applications.

### Overview: Start Journey is a text-to-image web application that leverages Stable Diffusion and FastAPI for backend and React for frontend. It runs locally on GPU, CPU, or Apple MPS, serving a single app from the same origin. ### Key Features: - **Stable Diffusion Integration**: Stable Diffusion is used behind a FastAPI backend. - **React Frontend**: Built with React for the frontend. - **Live Streaming Generation Progress**: Displays the generation progress in real-time. - **Device Support**: Auto-detects hardware and picks the fastest option, including CUDA (NVIDIA GPU), Apple MPS, and CPU. ### Tech Stack: - **Frontend**: TypeScript, React - **Backend**: FastAPI, Vite ### Why It Matters: Start Journey simplifies the process of generating images from text prompts, making it accessible and efficient for developers and artists alike. The single-origin approach ensures consistent performance and reduces CORS issues, making it ideal for local development and production environments. ### Repository: - **Adi-UA/start-journey** - **GitHub Stars**: 9 - **GitHub Description**: A text-to-image web app using Stable Diffusion behind a FastAPI backend and a React frontend. Runs locally on GPU, CPU, or Apple MPS. Supports live streaming generation progress. ### Primary Language: - **TypeScript** ### Device Support: - **CUDA (NVIDIA GPU)**: fp16, ~1-3 minutes - **Apple MPS**: fp32, tens of seconds - **CPU**: fp32, 1

Overview: Build interactive and visually appealing applications with the power of Gemini API and Streamlit. This repository serves as a foundation for creating data-driven web applications that seamlessly integrate data from the Gemini API while harnessing the user-friendly features of Streamlit for a dynamic and engaging user experience. Key features: - Seamlessly integrate data from the Gemini API - Utilize Streamlit's user-friendly features for a dynamic and engaging user experience - Create interactive and visually appealing applications Tech stack: Python Why it matters: This repository provides a comprehensive solution for developers looking to create data-driven web applications with the power of Gemini API and Streamlit, offering a seamless integration and user-friendly interface for a dynamic and engaging user experience.