Projects

Overview: HAG is an advanced knowledge-enhanced generation framework that leverages the power of vector databases and knowledge graphs to provide intelligent Q&A capabilities. Built on LangChain, Neo4j, and Weaviate, it excels in domain-specific knowledge retrieval and reasoning. Key features: - **Intelligent Intent Recognition**: Deep analysis of user query intent with precise knowledge need matching. - **Dual Database Integration Architecture**: Weaviate for semantic similarity search and Neo4j for complex relationship reasoning. - **Document Storage Management**: Support for multiple document formats with real-time processing and comprehensive analytics. Tech stack: Python, LangChain, Neo4j, Weaviate. Why it matters: HAG is a cutting-edge solution that combines the strengths of vector databases and knowledge graphs to deliver intelligent Q&A capabilities, making it a valuable tool for developers looking to enhance their applications with advanced knowledge retrieval and reasoning. Repository: yankmo/HAG Primary language: Python GitHub stars: 9 GitHub description: 🚀 HAG: Next-Gen AI | Neo4j + Weaviate Fusion | Dual-Similarity Retrieval | 100% Local & Private | Graph Intelligence Meets Vector Search

Overview: SentinelRAG is a self-healing academic knowledge assistant powered by Agentic RAG, LangChain, ChromaDB, and Google Gemini. Key features: - PDF upload and automatic storage - Text extraction and metadata preservation - Intelligent chunking with configurable size and overlap - Sentence Transformers embeddings - Semantic retrieval with source tracking - Multi-LLM support including Gemini, Qwen, Llama 3.3 - Source citations and conversation history Tech stack: Python Why it matters: SentinelRAG demonstrates a production-ready implementation of Retrieval-Augmented Generation, addressing limitations of LLMs by retrieving relevant information from user documents before generating responses. Repository: yashrajagawane/SentinelRAG GitHub stars: 9 GitHub description: Self-Healing Academic Knowledge Assistant powered by Agentic RAG, LangChain, ChromaDB, and Google Gemini.

Overview: Wan 2.7 R2V API (`wan2.7-r2v`) is a high-performance AI video model compatible with OpenAI endpoints, offering superior visual quality and professional-grade performance. The API provides a base URL and a single key for access, eliminating the need for subscription or prepayment. Key features: - Supports prompt and reference image/video inputs for style and character control. - Offers a tiered pricing model with flexible resolution options. - Returns the exact cost in the response for billing purposes. - Operates asynchronously by default, allowing for batch processing and retries. Tech stack: Python Why it matters: This API is ideal for developers looking to integrate AI video capabilities into their projects without the need for complex setup or ongoing costs. GitHub stars: 9 Repository: yesjbbz38yev4/wan-2-7-r2v-api-ru

### Overview: Weiyue Text Reader is a versatile text-to-speech application designed for Windows users, offering a warm, human-like voice for reading PDFs, EPUBs, and more. It supports offline reading, customizable speed, and features like bookmarking and pronunciation dictionary for enhanced listening practice. ### Key features: - Reads PDF, EPUB, TXT, MOBI, FB2, and even photos of printed pages. - Supports offline reading and customizable speed. - Features a pronunciation dictionary for improved listening practice. - Offers a sleep timer and cloud library sync for multiple devices. ### Tech stack: Python ### Why it matters: Weiyue Text Reader provides a convenient and engaging way to read documents, enhancing the listening experience with its warm and natural voice. Its offline capabilities and customizable features make it a valuable tool for various reading and learning purposes. ### Repository: - Repository: yshi40198-glitch/weiyue-text-reader - GitHub stars: 9 - GitHub description: 中文文字朗读工具（Windows）：打开文档即朗读，支持倍速、定时关闭、打标记续读、离线可用 ### Primary language: - Python ### GitHub description: 中文文字朗读工具（Windows）：打开文档即朗读，支持倍速、定时关闭、打标记续读、离线可用

Overview: This repository contains a sample .NET project demonstrating the use of Kernel Memory for semantic search and Retrieval-Augmented Generation (RAG) on a small commercial products dataset. It mimics an e-shop environment where users can search for products, and the application retrieves the most relevant results. Key features: - Kernel Memory integration for semantic search and RAG - Serverless setup of Kernel Memory using Docker - Embedded services for easy integration - Support for multiple vector databases: Postgres and Qdrant - OpenAI's gpt-4o-mini language model and text-embedding-ada-002 embedding model - Swagger API documentation for easy access to the application's endpoints Tech stack: C# Why it matters: This sample project showcases the capabilities of Kernel Memory in e-commerce, providing a practical demonstration of how AI can enhance the search functionality of an online store, making it easier for users to find products they are interested in.

Overview: This project provides a lightweight, native .NET implementation of the Byte-Pair Encoding (BPE) tokenizer specifically tailored for Qwen3 models. It supports various Qwen3 variants and integrates seamlessly with HuggingFace, making it ideal for developers working with Qwen3 models in C#/.NET applications. Key features: - Native .NET Implementation: Built for .NET applications, ensuring compatibility with existing C#/.NET projects. - Identical to HuggingFace: Produces the same token IDs and outputs as the official HuggingFace tokenizer, ensuring compatibility with Qwen3 models. - Qwen Model Family Support: Compatible with all Qwen3 variants (LLM, Embedding, Reranker, and Vision-Language models). - HuggingFace Integration: Load tokenizer files directly from HuggingFace model repositories, facilitating easy integration with Qwen3 models. - Configurable: Customize tokenization behavior through options and custom file providers, allowing for tailored configurations. - No External Dependencies: Does not require Python or other runtime dependencies, making it a standalone solution. Tech stack: C# Why it matters: This project is essential for developers working with Qwen3 models, providing a lightweight and efficient solution for tokenization that integrates seamlessly with HuggingFace, ensuring compatibility with various Qwen3 variants and reducing the need for external dependencies.

Overview: This project focuses on the few-shot image generation (FSIG) problem, analyzing current methods and proposing a new approach based on mutual-information contrastive learning (DCL) to enhance the diversity of generated images. Key features: - Analysis of FSIG methods - Proposal of DCL algorithm - Preservation of diversity during adaptation Tech stack: Python, PyTorch, lmdb, tqdm Why it matters: The project provides a detailed analysis of FSIG methods and introduces a novel approach that improves the diversity of generated images, which is crucial for applications requiring high-quality and varied outputs. Repository: yunqing-me/A-Closer-Look-at-FSIG GitHub stars: 9 GitHub description: The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) 2022

Overview: Nano Banana 2 Lite API is a cheaper alternative to Gemini 3.1 Flash Lite Image, providing access to 300+ models including OpenAI-compatible endpoints. Key features: - Supports OpenAI-compatible API endpoint - Pricing per image, with a base price of $0.0125 - Compatible with various models and resolutions - Non-blocking design, polling for task completion Tech stack: Python Why it matters: Nano Banana 2 Lite is an economical solution for developers looking to integrate OpenAI-compatible models without the need for a subscription or upfront payment.

### Overview: Seedance 2.5 Prompts: 8 Tested Recipes With Outputs Seedance 2.5 Prompts offers a comprehensive set of tested recipes for generating high-quality images using the Seedance 2.5 model. These recipes are designed to provide clear, effective prompts that can be easily integrated into any project. ### Key Features: - **Model ID:** `seedance-2.5` - **Image Prompt Recipes:** 8 tested recipes with detailed examples and cost estimates. - **Aspect Ratios:** 8 different aspect ratios for versatility. - **Model Compatibility:** Compatible with OpenAI's API, ensuring seamless integration. ### Tech Stack: - Python-based project using the Seedance 2.5 model. ### Why It Matters: - **Cost Efficiency:** The effective price of $0.0961 per second is highly competitive, making it an ideal choice for budget-conscious projects. - **Versatility:** The model supports multiple aspect ratios, ensuring that the output can be adapted to various project needs. - **Model Compatibility:** Ensures smooth integration with other OpenAI models, maintaining consistency across different projects. ### Repository: - Repository: yyoetry8jo2djt/doubao-seedance-2-5-prompts-de - GitHub Stars: 9 - GitHub Description: Seedance 2.5 Prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-2.5`) and the cost per image. ### Primary Language: - Python ### GitHub Description: -

Overview: Facial Region-Aware Makeup Transfer aims to enable diffusion-based makeup transfer with regional controllability, addressing limitations in current diffusion-based methods that struggle with capturing makeup styles and overlooking facial region-aware makeup features. Key features: - Makeup CLIP fine-tuning - Identity and facial region-aware makeup injection Tech stack: Python Why it matters: Facial Region-Aware Makeup Transfer (FRAM) addresses the limitations of current diffusion-based makeup transfer methods by enabling regional controllability and facial region-aware makeup features, thereby improving the quality and effectiveness of makeup transfer.

Overview: This comprehensive LangChain MasterClass course is designed for students, Python developers, backend engineers, and AI practitioners looking to build real-world AI applications using the LangChain ecosystem. The curriculum includes 18 modules and 100+ labs, covering from basic model invocation to advanced agent development, with all labs executable as standalone scripts. Key features: - 18 progressive modules - 100+ runnable labs - Multi-provider support (OpenAI, Anthropic, Google Gemini, Ollama, HuggingFace, OpenRouter) - Structured outputs using Pydantic, TypedDict, and JSON-schema - Vector databases (InMemoryVectorStore, Chroma, FAISS) - Retrieval-Augmented Generation (RAG) Tech stack: Python Why it matters: This course provides a practical, hands-on approach to mastering LangChain, making it suitable for transitioning into AI engineering, deepening a skill set, or building the foundational knowledge needed for RAG and agent systems.

Overview: DebugMind 是一个基于 AI 技术的 Bug 诊断系统，它通过结构化写入知识库，帮助工程师快速定位和修复重复性问题，从而显著提高工作效率和准确性。 Key features: - 结构化写入知识库，每次诊断结果都能被记录和利用。 - 通过记忆库中的历史案例，快速定位和解决重复性问题。 - 与 Hugging Face Spaces 集成，提供在线体验和搜索功能。 Tech stack: Python Why it matters: DebugMind 通过自动化和智能化的 Bug 诊断，帮助工程师提高工作效率，减少重复性工作，从而提升整体软件质量。

Overview: QwenForge is a lightweight white-box inference engine designed for Qwen2.5-0.5B-Instruct models, specifically optimized for legacy consumer GPUs (Turing sm_75) running sub-1B parameter models. It achieves a 3.34x speedup over the original vLLM benchmark on a single RTX 2080 Ti (Turing sm_75) GPU. Key features: - Custom Triton kernels for efficient inference - Handcrafted Transformer forward pass and attention management - Batch scheduling and fixed per-token overhead for optimal performance on older GPUs Tech stack: Python Why it matters: QwenForge provides a detailed benchmark and educational reference for inference optimization, offering a white-box approach to performance tuning for legacy consumer GPUs.

### Overview: govector-bm25 is a high-performance vector store written in Go that combines the precision of keyword search with the semantic understanding of embeddings. This library uses BM25 for term-based retrieval and a vector index for approximate nearest neighbor search, providing a balanced approach to both keyword matching and semantic search. ### Key Features: - **Hybrid Retrieval**: Combines BM25 term-based scoring with vector-based cosine similarity for a comprehensive search experience. - **Go Implementation**: Built in Go for optimal performance and ease of deployment without requiring Python dependencies. - **Efficient Memory Usage**: Utilizes HNSW (Hierarchical Navigable Small World) for approximate nearest neighbor search, reducing memory footprint. ### Tech Stack: Go ### Why It Matters: govector-bm25 offers a solution for developers looking to integrate semantic search capabilities into their Go applications. By combining the strengths of keyword-based search with the semantic insights of vector embeddings, it provides a powerful tool for applications requiring both exact matches and contextual understanding. ### GitHub Stars: 8 ### Repository: 2aronS/govector-bm25

Overview: This project provides an exact implementation of FlashAttention for AMD RDNA2 GPUs, specifically targeting the gfx1030 / RX 6800 XT. It is designed as a drop-in replacement for the FlashAttention library, offering 19.8 TFLOP/s forward and 99% of HBM bandwidth decode performance. Key features: - Exact implementation of FlashAttention for AMD RDNA2 GPUs - Supports causal and sliding window attention - Compatible with the FlashAttention library API - Measured performance on a RX 6800 XT: 19.8 TFLOP/s forward, 505-528 GB/s decode, 36 TFLOP/s raw v_dot2 issue rate Tech stack: HIP (Heterogeneous Instruction Processing) Why it matters: This project offers a high-performance alternative to existing FlashAttention implementations, providing significant performance improvements on AMD RDNA2 GPUs. The exact implementation ensures that the results are equivalent to a standard attention implementation, making it suitable for applications requiring high accuracy and throughput. Repository: 2kiss/flash-attention-rdna2 GitHub stars: 8 GitHub description: Exact FlashAttention for AMD RDNA2 (gfx1030 / RX 6800 XT), written in HIP. Drop-in replacement for flash-attn: 19.8 TFLOP/s forward, decode at 99% of HBM bandwidth.

Overview: R-Mem is a lightweight Rust implementation of mem0's memory architecture, designed to be a study of their approach using Rust. It features a single binary, SQLite-backed, and includes a tiered archive system, all in 2,826 lines of Rust code. Key features: - SQLite-backed memory - Vector memory - Graph memory - History tier - MCP (Memory Control Protocol) support Tech stack: Rust Why it matters: R-Mem is a study of mem0's memory architecture in Rust, offering a minimal footprint with a single binary and SQLite-backed memory. It's designed for AI agents and aims to be as close to mem0 as possible while focusing on simplicity and efficiency. Repository: Adaimade/R-Mem GitHub stars: 8 GitHub description: Lightweight Rust alternative to mem0. 3.2MB binary, SQLite vector + graph memory, Ollama support.

Overview: This repository is a production-grade multi-modal generation system designed for text-to-image (T2I) and text-to-video (T2V) synthesis. It includes end-to-end implementations of VAE, GAN, VQ-VAE, U-Net, DiT, and Video DiT with DDPM/DDIM/DPM-Solver++ samplers, along with a full FID/IS/CLIP evaluation suite. Key features: - End-to-end implementation of modern generative models - Production system architecture with async job API, Redis reliable queue, GPU workers, continuous batching, S3 artifact store, moderation gates, Prometheus/OpenTelemetry observability, and Kubernetes autoscaling - Includes 186 tests, CI, and operations documentation Tech stack: Python Why it matters: This repository provides a coherent, tested platform architecture for multi-modal generation, including system architecture, data flow, failure modes, and scaling to 100M+ users. Repository: AdilShamim8/Multi-Modal-105 GitHub stars: 8 Primary language: Python GitHub description: Production-grade multi-modal generation system for Text-to-Image (T2I) and Text-to-Video (T2V) synthesis. End-to-end implementation of VAE, GAN, VQ-VAE, U-Net, DiT, and Video DiT with DDPM/DDIM/DPM-Solver++ samplers, distributed training, and full FID/IS/CLIP evaluation suite.

### Overview: This project is a one-command recipe for running GLM-5.3 on four RTX PRO 6000 GPUs, utilizing TensorFold for a full model experience. The recipe includes checkpoints from Mia's AI Lab, patches from Aevonix Research, and optimized single-host PCIe setups. ### Key Features: - **Checkpoint:** Mia-AiLab/GLM-5.3-EXL3-2.75bpw-TensorFold - **Drafter:** RedHatAI/GLM-5.3-speculator.dspark (DSpark) - **API Model ID:** glm-5.3 - **Context:** 1,048,576 tokens per request - **Concurrency:** 5 sessions by default - **One Command:** `./start.sh`; restart with `./start.sh restart`, stop with `./stop.sh` ### Tech Stack: Python ### Why It Matters: This recipe provides a streamlined, efficient way to run GLM-5.3 on powerful hardware, leveraging TensorFold for optimal performance and Mia's AI Lab's expertise. The single-host PCIe optimizations ensure optimal resource utilization and performance.

### Overview: `docs2vecs` is a CLI tool that simplifies the process of vectorizing and exposing your documentation. It helps create RAG-powered applications by simplifying data ingestion, proof of concept development, and seamless integration with vector databases. ### Key features: - Vectorizes your data - Simplifies data ingestion - Enables RAG-powered applications - Supports local and Docker-based deployment ### Tech stack: Python ### Why it matters: `docs2vecs` streamlines the process of vectorizing documents, making it easier to create and deploy RAG-powered applications. It simplifies the integration with vector databases like ChromaDB, making it a valuable tool for developers looking to enhance their documentation with AI capabilities.

Overview: API-LLM-Hub is a lightweight JavaScript library that simplifies the integration of various AI language models directly into static web projects, offering a straightforward interface without the need for build steps or backend. Key features: - Simple Integration - Browser-Focused - CORS-Friendly - Consistent Interface across Multiple Providers - Quick Prototyping Tech stack: HTML Why it matters: API-LLM-Hub simplifies the integration of AI language models into static web projects, making it ideal for projects without a backend server and avoiding common deployment issues like CORS. Repository: AmanPriyanshu/API-LLM-Hub GitHub stars: 8 GitHub description: A static-page vanilla-js interface for various LLM APIs (OpenAI, Claude, Gemini, Together).