Projects

### Overview: Makra is a memory layer designed to bridge the gap between the web and your AI agents, enabling real-time structured data retrieval from the open web. It offers a cost-effective vector search solution, allowing you to access structured data without the need for expensive vector-based search. ### Key Features: - **Real-time Structured Data Retrieval**: Makra fetches structured data from the web and returns it in a JSON format, making it easy to integrate with AI agents. - **Vector Search Integration**: It supports vector search, allowing for more advanced querying capabilities. - **Customizable Schemas**: Users can define schemas to specify the fields they need, ensuring precise data extraction. ### Tech Stack: - **Python**: The primary language used for the SDKs and playground. - **JavaScript**: Used for the JavaScript SDK, providing a TypeScript declaration file for better type safety. ### Why It Matters: Makra simplifies the process of extracting structured data from the web, making it accessible and cost-effective for developers and AI agents alike. Its integration with vector search enhances the querying capabilities, offering a unique solution for data retrieval and analysis. ### Repository: - **Repository**: `makralabs/makra` - **Primary Language**: Python - **GitHub Stars**: 30 - **GitHub Description**: Makra is a memory layer between the web and your AI agents. It serves real-time structured data from the open web at the cost of a vector search. - **Readme Excerpt**: A lockup image, a quote about crawling the web like a database query, and

Overview: This example demonstrates how to integrate Semantic Kernel and Kernel Memory with SQL Server as a vector database, showcasing the efficient performance of vector search using COLUMNSTORE indexes. Key features: - Semantic Kernel integration - Kernel Memory for efficient vector storage and retrieval - SQL Server as vector database - Efficient vector search with COLUMNSTORE indexes Tech stack: C# Why it matters: This example provides a practical demonstration of how to leverage Semantic Kernel and Kernel Memory for working with embeddings in a .NET application, particularly useful for applications requiring fast and accurate vector search capabilities.

### Overview: Fine-tuning an LLM using a Generic Workflow and Best Practices with PyTorch. ### Key features: - Custom PyTorch model creation - Utilizing a generic workflow for LLM fine-tuning - Incorporating best practices for model training and evaluation ### Tech stack: PyTorch ### Why it matters: This workflow provides a comprehensive framework for beginners to understand LLM project structure and PyTorch, showcasing a practical example of fine-tuning an LLM. It includes recent advancements in NLP and open-source software, making it a valuable resource for adapting PyTorch for various deep learning applications.

Overview: WhatsApp voice transcriber is an audio message transcriber app created with APIMatic auto-generated SDKs for WhatsApp and OpenAI Whisper API. Key features: - Transcribes incoming voice messages in real-time - Supports multiple languages - Integrates with Meta's WhatsApp API and OpenAI Whisper API Tech stack: TypeScript Why it matters: This app enables easy communication across languages by transcribing voice messages in real-time, making it a valuable tool for developers and businesses looking to improve their communication capabilities.

### Overview: GigaPisar is a command-line tool that allows you to transcribe speech into text using GigaAM, a state-of-the-art Russian speech-to-text model. It works offline, on your own machine, and is optimized for performance and efficiency. ### Key Features: - **Offline Transcription**: Transcribes speech directly from audio files without internet access. - **Fast Processing**: Utilizes a CPU-only approach for quick transcriptions. - **Model Conversion**: Converts the GigaAM model to ONNX and quantizes it to int8, making it lightweight and suitable for server-side processing. - **Command Line Interface**: Supports various input formats (ogg, m4a, mp3, wav) and splits recordings on pauses between phrases. - **Output Quality**: Outputs clean text with diagnostics directed to stderr, making it easy to integrate with pipelines. - **Model Files**: Provides pre-made model files that don't require manual export or additional dependencies. - **Dependencies**: Minimal dependencies, including `onnxruntime`, `numpy`, `sentencepiece`, `pyyaml`, and `gigaam` are included, totaling approximately 150 MB instead of a gigabyte. - **Server Mode**: Runs as a server that can handle multiple recordings in memory, making it suitable for continuous transcription. - **Customization**: Supports custom model folders and can be run as a systemd service. - **Hotkey Dictation**: Allows dictation on a Linux desktop using a hotkey, with the text pasted into the active window. ### Tech Stack: The tool is

### Overview: LinguFlow is a low-code tool designed for LLM application development, simplifying the building, debugging, and deployment process for developers. ### Key features: - **Technical Characteristics**: Construction based on a DAG of information flows. - **Business Characteristics**: Suitable for those with a clear understanding of how to solve business problems using LLM, enhancing accuracy and logic-specific solutions. ### Tech stack: TypeScript ### Why it matters: LinguFlow simplifies the integration of LLMs into complex business processes, emphasizing accuracy and logic-specific solutions.

### Overview: This project compares the performance of LanceDB and Elasticsearch in full-text search (FTS) and vector search using the Wine Reviews dataset. ### Key features: - **LanceDB**: A distributed database optimized for full-text search. - **Elasticsearch**: A popular full-text search engine. - **Embedding Model**: ModernBERT-Embed-BASE for generating vector embeddings. - **Benchmark Queries**: Keyword-based queries and vector search queries. - **Setup**: Runs on Docker containers for both LanceDB and Elasticsearch. - **Benchmark Protocol**: Runs 1000 queries per search type, averaged over 3 runs. - **Embedding Model**: 256-dimensional embeddings for text fields. - **LanceDB Workflow**: Ingests data, creates indexes, and runs benchmarks. - **Elasticsearch Workflow**: Ingests data, creates indexes, and runs benchmarks. - **Direct Client Benchmark**: Runs against both databases without FastAPI overhead. - **End-to-End API Benchmark**: Runs against both databases via a FastAPI server. ### Tech stack: Python ### Why it matters: This project provides a comprehensive comparison of LanceDB and Elasticsearch in terms of their performance in full-text search and vector search, using the Wine Reviews dataset. It helps developers understand the strengths and weaknesses of each database in different use cases.

Overview: The LLM Sampling Library is a Python package designed to facilitate text generation using various sampling methods with large language models (LLMs). Key features: - Multiple sampling methods including Unconstrained sampling, Top-k sampling, Top-p (nucleus) sampling, Min-p sampling, Typical sampling, Epsilon sampling, Eta sampling, Beam search, Chain-of-Thought (CoT) decoding, Constrained JSON decoding, Speculative sampling, and Medusa decoding. - Chat template support for Instruct models. - Memory efficient implementation using KV-cache. Tech stack: Pure Pytorch. Why it matters: The library provides a simple command-line interface for generating text based on input prompts, making it accessible for developers and researchers alike.

Overview: Coqui STT offline engine API for Node.js developers, providing a simple HTTP ASR server interface. Key features: - Transcription of audio files or buffers - Model loading and freeing - Example usage in Node.js Tech stack: JavaScript Why it matters: Simplifies access to Coqui STT's speech-to-text transcription capabilities for Node.js developers, making it easier to integrate into their projects without needing to directly interact with the native library. Repository: solyarisoftware/CoquiSTTJs GitHub stars: 30 GitHub description: Coqui STT runtime transcript Node.js client, including multithread (server) usage examples.

### Overview: This curated map of 3,000+ open-source AI agent tools is organized into 29 categories, sorted by GitHub stars. It provides a comprehensive overview of the tools used in AI agent systems, from coding agents in the terminal to model serving the tokens. The repository is maintained by Tayyab Imam, with 29 GitHub stars and a curated list of tools that are real, public, not archived, and actively maintained. ### Key Features: - **Coding Agents & Harnesses**: Tools for building AI agents. - **Skills, Plugins & Context Engineering**: Enhancements to AI agents. - **Agent Frameworks & Orchestration**: Frameworks for orchestrating AI agents. - **Memory & Persistent Context**: Tools for managing and persisting context. - **RAG & Retrieval**: Retrieval-based AI agents. - **MCP Servers & Tool Integration**: Servers and integration tools. - **Sandboxes, Browsers & Computer Use**: Tools for sandboxing and browser-based AI. - **Document Processing & Data Ingestion**: Tools for processing and ingesting documents. - **Workflow Automation**: Automation tools for workflows. - **Model Gateways & Routing**: Tools for model gateways and routing. ### Tech Stack: TypeScript ### Why It Matters: This curated list helps developers find the right tools for each layer of an AI agent system, ensuring a smooth and efficient development process. It's a valuable resource for anyone looking to build or integrate AI agents into their projects.

Overview: This project is a Taiwanese Hokkien speech-to-text transcriber built on the MediaTek Breeze-ASR-26 model optimized for NVIDIA RTX 3050 GPUs with 4GB VRAM. It supports various formats like SRT, VTT, TXT, JSON, and provides a Gradio UI, CLI, Docker, and WSL2/Docker integration. Key features: - Taiwanese Hokkien (Taigi) speech-to-text transcriber - Uses MediaTek Breeze-ASR-26 model for high accuracy - Optimized for low VRAM GPUs - Supports multiple formats and outputs - Gradio web interface, command line interface, Docker, and WSL2/Docker integration - Built-in GPU profiler for efficient use of resources Tech stack: Python Why it matters: This project is essential for Taiwanese language enthusiasts and those working in the Taiwanese language community, providing accurate and efficient speech-to-text transcriptions for Taiwanese Hokkien (Taigi) content. Repository: thc1006/breeze-asr-taigi Primary language: Python GitHub stars: 30 GitHub description: Taiwanese Hokkien (Taigi) speech-to-text transcriber - MediaTek Breeze-ASR-26 with faster-whisper, tuned for RTX 3050 4GB low-VRAM GPUs. Gradio UI, CLI, Docker, SRT/VTT/TXT/JSON.

Overview: Cargo Chat is a Rust-based tool designed to help developers understand and navigate their codebases using natural language queries. Key features: - Smart chunking of codebase into semantic chunks - Semantic search and AI analysis - Indexing phase for repository scanning and smart chunking - Query phase for processing user queries and generating intelligent answers Tech stack: Rust, Jina, Qwen3, tree-sitter, ANN index, OpenAI Why it matters: Cargo Chat simplifies the process of understanding and debugging code by leveraging AI and semantic search, making it easier for developers to find relevant information and solutions in their repositories.

Overview: XAI 是一个透明接入的 AI API 路由与治理平台，支持多租户管理、BYOK模式、端到端加密存储和无超级管理员的架构。 Key features: - XAI Router (云端聚合，预充值) - XAI Control (云端 BYOK) - XAI Router Private (私有化部署) Tech stack: Node.js, Docker, Kubernetes, RESTful API Why it matters: XAI 提供了透明接入、安全加密、多租户管理和成本效益的解决方案，适用于快速验证、生产环境和合规隔离场景。 Repository: xaixapi/xai GitHub stars: 30 GitHub description: XAI

### Overview CoRe2 is an official implementation of our paper "Improved and Accelerated Text-to-Image Generation with Collect, Reflect, and Refine." This project provides a comprehensive guide on how to use the CoRe2 pipeline, which includes inference code for various models such as Stable Diffusion 3.5 Large and Stable Diffusion XL. The project is designed to enhance text-to-image generation efficiency and effectiveness, making it a valuable resource for developers and researchers in the field of computer vision and artificial intelligence. ### Key Features - **Inference Code for Multiple Models**: CoRe2 supports inference for Stable Diffusion 3.5 Large and Stable Diffusion XL, providing enhanced text-to-image generation capabilities. - **Noise Model Weights**: The project includes noise model weights for both models, ensuring better image quality and performance. - **Model Weights and Demo**: The script provides model weights and a demo space on Hugging Face, making it easy to start using the CoRe2 pipeline. - **Command-Line Arguments**: The script supports various command-line arguments for customization, including pipeline selection, prompt, inference steps, classifier-free guidance scale, and more. ### Tech Stack - **Python**: The primary programming language used for the project. - **Hugging Face**: The library used for model weights and demo space. - **SD3.5 and SDXL**: Models used for inference. - **Flux**: The inference method for the Collect phase. - **LlamaGen**: The inference method for the Reflect phase. ### Why It Matters CoRe2 is

### Overview: CanvasChat 是一个基于 Gemini 模型的精美聊天界面，专为 Canvas 预览模式设计，提供免费、大量使用 Gemini 3 Flash 模型的机会。它支持多种多模态输入，包括图片、视频、音频、PDF 等，以及高效的对话和多模态交互体验。 ### Key features: - **多模态输入**：支持图片、视频、音频、PDF 等多种输入方式。 - **高效对话**：专为 Canvas 环境优化的聊天界面。 - **多模态交互**：支持图片、视频、音频、PDF 等多种输入。 - **生产级体验**：完整的功能支持，不输付费版本。 ### Tech stack: React, TailwindCSS, KaTeX, Google Gemini API, TTS API. ### Why it matters: 通过本项目，你可以免费、大量地使用 Gemini 3 Flash 模型进行对话，体验到 Gemini 的强大功能，同时享受生产级的聊天界面和多模态交互体验。

Overview: This project, I2T2I: Text-to-Image Synthesis with textual data augmentation, leverages text-to-image synthesis techniques to create images from textual descriptions. The code is compatible with Python 2.7 and TensorFlow 0.11, requiring the setup of Google's setup guide and Inception model for image captioning. It includes scripts for training image captioning models and text-to-image synthesis using these models. Key features: - Text-to-Image Synthesis - Image Captioning Model Training - Data Augmentation for Improved Results Tech stack: Python Why it matters: This project offers a novel approach to text-to-image synthesis, enhancing the ability to generate images from textual descriptions, which has applications in various fields including art, design, and virtual reality. The use of text augmentation techniques improves the quality and diversity of generated images, making it a valuable tool for creators and researchers in the field of image synthesis.

### Overview: 这是一个基于MiniMax CLI的统一Model Context Protocol（MCP）服务器，旨在提供全面的模态支持，包括文本、搜索、图像理解、图像生成、语音合成、视频生成、音乐生成和配额查询。 ### Key features: - **Model Context Protocol (MCP) Server**: 实现统一的MCP协议，支持多种模态。 - **MiniMax CLI Integration**: 无缝集成MiniMax CLI工具，简化配置过程。 - **API Key Management**: 支持环境变量、命令行参数和本地登录认证。 ### Tech stack: JavaScript ### Why it matters: 该服务器为MiniMax CLI用户提供了一个统一的MCP接口，简化了模态处理和API调用，提高了开发效率和用户体验。

Overview: 10xGraph is a production-grade Python framework designed for building multi-agent AI systems. It uses a graph-based approach, supports LLM-agnostic tools, and includes features like live agents, parallel tool execution, and a 3-layer memory system. Key features: - Graph-based orchestration - LLM-agnostic tools (OpenAI, Google GenAI, Anthropic) - Redis cache + Postgres + vector store for 3-layer memory - Live agents - Parallel tool execution - Versioned state writes - Replay-safe tools - Real node and tool timeouts - Human approval inside a tool Tech stack: Python Why it matters: 10xGraph is a robust and scalable solution for building AI agents, ensuring correct operation even in the face of failures, and providing a production-ready server with various features like event publishing and long-term memory.

**Overview:** 本项目介绍了一款在8GB显存的显卡上运行的35B MoE模型Cyber-Tiel-Coder-35B-A3B，通过专家权重的内存按需搬运和缓存机制，实现了专家权重的高效处理。实测结果表明，该模型在RTX 4080 SUPER显卡上，8GB显存档的解码速度为57.7-63.5 tok/s，全常驻档速率为134-136 tok/s。 **Key features:** - 专家权重留在系统内存中，按需搬运到GPU上计算。 - 专家缓存按剩余显存自适应，KV数据可压位宽为int8与fp16数值逐位一致。 - 未命中GPU缓存的专家在CPU线程池中直接计算。 **Tech stack:** C++ **Why it matters:** 本项目解决了35B MoE模型在小显存卡上高效处理专家权重的问题，通过内存按需搬运和缓存机制，实现了专家权重的高效计算，显著提升了模型的性能和能效比。 Repository: 1314521gjy/cyber-on-strata GitHub stars: 29

### Overview: OPSIE is a terminal-native self-centered intelligence prototype designed for technical work. It offers semantic recall, image generation, market data retrieval, and more, all within a single session and infrastructure. ### Key Features: - **Semantic Recall**: Recall previous conversations and inputs. - **Image Generation**: Create images, music, and videos. - **Market Data Retrieval**: Access real-time market data. - **Mail Drafting and Sending**: Draft and send emails. - **Structured DNA Workflows**: Run DNA workflows in a single session. - **Collaborative Rooms**: Facilitate multi-agent collaboration. - **Durable Conversation Memory**: Keeps conversation history and context. ### Tech Stack: Python ### Why it Matters: OPSIE is a groundbreaking prototype that explores the future of AI-human interaction, offering a unified platform for technical work and continuous learning.