Projects

### Overview: This repository is a hands-on curriculum and reference implementation for building production-grade agentic AI systems. It focuses on engineering deterministic, production-grade systems around non-deterministic LLMs, including FSM, durable execution, retries, DAGs, model routing, edge inference, RAG, memory, multi-agent orchestration, security, and observability. The code is intentionally small, with each folder representing a self-contained, minimal, runnable proof-of-concept for one architectural concern in agentic AI systems. The goal is transferable mental models, not framework mastery. ### Key Features: - **State Control**: Implementing state machines and durability to ensure the unpredictability of LLMs is contained, observable, and recoverable. - **Crash Recovery**: Handling retries and edge inference to ensure the agent can recover from crashes and network drops. - **Model Routing**: Implementing model routing to ensure the model is used correctly and predictably. - **Multi-Agent Orchestration**: Building multi-agent systems to ensure the agents can work together effectively. - **Security**: Implementing security measures to protect the system from unauthorized access and attacks. - **Observability**: Ensuring the system can be observed and monitored for performance and reliability. ### Tech Stack: Python ### Why It Matters: This repository is designed to help engineers build reliable, auditable, production-grade software on top of a component that is fundamentally probabilistic. By focusing on deterministic systems, it aims to close the failure modes that arise from the inherent unpredictability of LLMs, ensuring that the system

### Overview: BlossomLM is a powerful, open-source large language model designed for local deployment, combining efficient two-mode thinking, tool use with interleaved reasoning, and image understanding. It supports agentic workflows and handles up to 262,144 tokens with speculative decoding via Multi-Token Prediction (MTP). ### Key Features: - **Efficient Two-Mode Thinking:** Supports always-on `medium` and `max` thinking, defaulting to `max` for more thorough reasoning. - **Tool Use with Interleaved Thinking:** Reasons and makes decisions before every tool call, enabling strong performance in agentic tasks. - **Image Understanding:** Understands image inputs alongside text. - **Long Context:** Handles up to 262,144 tokens (256K); use 131,072 tokens (128K) for best results. - **Faster Inference with MTP:** Supports speculative decoding via Multi-Token Prediction (MTP) in both vLLM and llama.cpp. ### Tech Stack: BlossomLM uses a custom chat template that differs from the native templates of its Qwen base models. Always use the bundled `chat_template`; do not replace it with a native Qwen chat template or combine the two templates. ### Why It Matters: BlossomLM is designed for local deployment, combining efficient two-mode thinking, tool use with interleaved reasoning, and image understanding. It supports agentic workflows and handles up to 262,144 tokens with speculative decoding via Multi-Token Prediction

Overview: searxNcrawl is an open-source MCP server that provides web search and clean Markdown crawling functionalities, built on Crawl4AI. It allows AI agents to search SearXNG and You.com in parallel and merges results, running self-hosted and without requiring an API key. Key features: - Web search: Searches configured providers in parallel and merges results. - Page crawl: Converts URLs to Markdown or JSON with repeated blocks removed and links stripped. - Site crawl: Crawls whole sites with depth and page limits, optimized for documentation. - Authenticated crawls: Captures login sessions with `crawl-capture` and crawls behind the login. - STDIO and HTTP: Connects to local MCP clients or runs an HTTP endpoint with optional Host/Origin checks and CORS configuration. - CLI and Python API: Commands `crawl`, `search`, and `crawl-capture` available on the command line or in Python. Tech stack: Python Why it matters: searxNcrawl simplifies web scraping and content management for AI agents, providing a self-hosted solution without the need for external APIs, making it ideal for developers looking to integrate web search and clean crawling into their projects.

### Overview: This repository is a curated list of awesome open source libraries that help you deploy, monitor, version, scale, and secure production agentic systems and applications. Stay updated with new production ML libraries added every month through the GitHub releases. ### Key features: - ADK (Agent Development Kit) - Google's framework for building production-ready AI agents. - Agent Lightning - Framework for building production-ready AI agents with Lightning AI. - AgentKit - Help agent developers build multi-agent networks with deterministic routing and rich tooling via MCP. - Agents - Allows users to build AI-driven server programs that can see, hear, and speak in real-time. - AgentScope - Multi-agent platform designed for building and managing AI agents. ### Tech stack: Python ### Why it matters: This repository provides a curated list of awesome open source libraries that are essential for deploying, monitoring, versioning, scaling, and securing production agentic systems and applications. Stay updated with new libraries added every month through the GitHub releases.

Overview: This project compiles all 5,352 accepted papers from ICLR 2026 into a static HTML webpage. Each paper is automatically analyzed in six dimensions, including research motivation, problem statement, phenomenon analysis, main methods, data and experiments, and main contributions. The papers are organized into 21 major categories and 126细分 categories, with a two-level directory structure. The Oral papers are highlighted with a golden badge, and the entire 224 Oral papers are marked with a chip for easy filtering. Key features: - Automatically analyzes 5,352 ICLR 2026 accepted papers in six dimensions. - Organizes papers into 21 major categories and 126细分 categories. - Highlights 224 Oral papers with a golden badge. - Provides a two-level directory structure for easy navigation. - Includes a search function with a chip for filtering papers by title or keywords. Tech stack: HTML Why it matters: This project simplifies the process of reading through 5,352 papers, making it easier for developers to quickly understand the latest research in the field without having to read through each paper individually.

### Overview: GooglePT is an AI-powered chat and search summaries tool designed to enhance your Google Search experience by summarizing relevant results in a conversational manner. Powered by the latest language models, it helps you quickly find the most pertinent information from your searches. ### Key Features: - **AI Chat Interface**: Engage with the tool through a conversational chat interface. - **Search Summaries**: Receive concise summaries of your search results. - **Integration with Google Search**: Seamlessly integrate with your existing Google Search experience. ### Tech Stack: - JavaScript ### Why It Matters: GooglePT simplifies your search experience by providing quick, relevant summaries directly from Google Search. This tool enhances your productivity and makes finding the information you need more efficient and enjoyable. Repository: KudoAI/googlegpt

Overview: AnythingLLM Mobile is an AI assistant that lives on your phone, providing real-time support, local access to models, and privacy-focused functionality. It combines the power of local models with seamless connectivity to the cloud, ensuring that your AI experience is always available and efficient. Key features: - **Local-first AI**: Runs models right on your phone, ensuring privacy and efficiency. - **Hybrid model approach**: Models can be run locally on your phone or connected to cloud providers. - **Cross-device compatibility**: Supports integration with various models and providers, including LM Studio, Desktop, and llama.cpp. Tech stack: TypeScript Why it matters: AnythingLLM Mobile offers a powerful, private AI assistant that can be accessed locally, ensuring that your AI experience is always available and efficient, without the limitations of cloud-based models.

### Overview: This project implements a novel text-to-image personalization technique based on a neural space-time representation. The approach optimizes a small neural mapper to directly represent concepts in a space-time representation, resulting in a compact yet expressive representation. The method introduces a textual bypass to improve convergence and visual fidelity, and allows for controlling the reconstruction and editability of the learned concept using a single trained model. ### Key features: - **Neural Space-Time Representation**: Represents concepts in a space-time representation composed of hundreds of vectors. - **Implicit Representation**: Concept representation is implicitly optimized through a small neural mapper. - **Textual Bypass**: Adds a textual bypass to improve convergence and visual fidelity. - **Controlled Editability**: Allows users to control the reconstruction and editability of the learned concept using a single trained model. ### Tech stack: Python ### Why it matters: This method offers a compact yet expressive representation of personalized concepts, enabling high-quality and controllable compositions without the need for fine-tuning the generative model. The ability to control the reconstruction and editability of the learned concept at inference time is a significant advantage, as it allows for fine-tuning without altering the generative model parameters.

### Overview: Nautilo is an AI-powered collaborative workspace that allows people and machines to work together seamlessly across desktop, mobile, and web. It's designed to be open source and MIT-licensed, making it accessible to everyone. ### Key features: - **AI Collaboration**: Enables real-time collaboration between humans and AI agents. - **Model API Key**: Allows users to interact with and create models. - **Model Provider Keys**: Enables access to various model providers. - **Security & Privacy**: Ensures secure and private interactions. - **Room Management**: Facilitates the creation and management of collaborative spaces. - **Genie Customization**: Allows users to personalize their AI agents with personalities, memories, and voices. - **Server Deployment**: Supports deployment on local servers or cloud platforms. - **Documentation**: Comprehensive guides and resources for operators, users, and builders. ### Tech stack: TypeScript ### Why it matters: Nautilo is a powerful tool for anyone looking to enhance their AI capabilities or collaborate with others in a dynamic and efficient environment. Its open-source nature and MIT license make it accessible and adaptable to various use cases.

### Overview: Localwriter is a LibreOffice Writer extension that enables local inference generative AI features, making it easier to add local AI-powered enhancements to your writing. ### Key Features: - **Extend Selection**: Uses a language model to predict what comes after the selected text, useful for generating stories, emails, or summaries. - **Edit Selection**: Allows users to replace the selected text with edited versions, suitable for changing tone, translating text, or semantically editing a scene. ### Tech Stack: Python ### Why It Matters: Localwriter simplifies the integration of AI-powered writing tools into LibreOffice, making it accessible to a broader audience without requiring extensive setup or technical knowledge.

Overview: A simple command-line client for the Perplexity API, allowing users to ask questions directly from the terminal. Key features: - Easy querying of the Perplexity API - Support for various language models - Optional display of token usage statistics - Optional display of citations - Colorful output formatting (with glow support) - API key handling from environment variable or command-line argument Tech stack: Python Why it matters: This tool simplifies interaction with the Perplexity API, making it accessible and user-friendly through a straightforward command-line interface. Repository: dawid-szewc/perplexity-cli Primary language: Python GitHub stars: 182

Overview: LlamaBot is a Pythonic interface designed to simplify the experimentation and integration of Large Language Models (LLMs) into Python applications. Key features: - Supports multiple LLM providers including LiteLLM, OpenAI, and Mistral. - Provides a simple `SimpleBot` class for stateless bot creation. - Integrates with Ollama, LMStudio, and other API providers for model access. - Compatible with various LLM models supported by LiteLLM. Tech stack: Python Why it matters: LlamaBot streamlines the process of experimenting with and integrating LLMs into Python projects, making it easier to build applications that utilize LLMs without needing deep knowledge of LLMs or API configurations. Repository: ericmjl/llamabot GitHub stars: 183

Overview: Paint-by-Sketch is a method that allows users to manipulate images by combining sketches and reference images, enabling fine-grained modifications and sub-part editing. Key features: - Incorporates sketches as a novel modal alongside reference images for image composition. - Uses edge-level controllability to enable desired structure and content edits. - Fine-tunes a pre-trained diffusion model to complete missing regions with reference image guidance. - Offers unique opportunities for user-driven image manipulation, enhancing the ability to obtain in-demand images. Tech stack: Python Why it matters: Paint-by-Sketch provides a powerful tool for image manipulation by leveraging the strengths of large-scale text-to-image generative models, making it suitable for a wide range of applications including art, design, and creative editing. Repository: kangyeolk/Paint-by-Sketch GitHub stars: 183 GitHub description: Stable Diffusion-based image manipulation method with a sketch and reference image README excerpt: <!-- <br> --> ### Teaser  ### Multi-backgrounds  ### Multi-references  ## Abstract >Recent remarkable improvements in large-scale text-to-image generative models have shown promising results in generating high-fidelity images. To further enhance editability and enable fine-grained generation, we introduce a multi-input-conditioned image composition model that incorporates a sketch as a novel modal, alongside a reference image. Thanks to the edge-level controllability

### Overview: The LLM Tools project is an open-source calculator designed to estimate the memory requirements for running large language models (LLMs). This tool simplifies the process of determining the necessary resources needed for inference and training, making it easier for developers to plan and optimize their LLM systems. ### Key Features: - **Memory Requirements Calculator:** Estimates the memory needed for running or training LLMs. - **Inference:** Calculates the total memory required for inference based on model size, precision, batch size, and sequence length. - **Training:** Estimates the memory required for training an LLM, considering factors like model size, precision, batch size, and sequence length. ### Tech Stack: Python ### Why It Matters: This tool is crucial for developers working with large language models as it helps in determining the necessary computational resources, ensuring efficient and effective use of hardware resources.

### Overview: 本项目开发了一种有界、自纠错的长期记忆层，用于LLM代理，通过整合连续、含矛盾的项目交互日志，形成一个可更新、可查询的记忆库。该技术旨在提高代理在处理复杂任务时的效率和准确性。 ### Key features: - **Bounded, Self-Correcting Long-Term Memory**: 蒸馏连续且含矛盾的项目交互日志，形成有界且可自更新的记忆库。 - **Candidate Pool**: 代理候选池，用于筛选和评估任务。 - **Value Dynamics**: 任务价值动态评估，确保记忆库中的信息是实时和相关的。 - **Tension Adjudication**: 背景信息的冲突解决机制，确保记忆库中的信息是准确的。 - **Causal-Replay Evaluations**: 基于因果关系的重放评估，提高记忆库的准确性和有效性。 ### Tech stack: TypeScript ### Why it matters: 本项目通过整合连续且含矛盾的项目交互日志，形成一个有界、自纠错的长期记忆层，显著提高了LLM代理在处理复杂任务时的效率和准确性。该技术在多个实际应用场景中表现优异，能够有效解决代理在面对复杂任务时的信息处理问题。 ### Repository: 1173591564/memory-manage GitHub stars: 184 GitHub description: LLM agent 的有界自纠错长期记忆层：把连续、含矛盾的

Overview: Voice Key is a privacy-first voice-to-text solution for Windows, macOS, and Linux, designed to transcribe speech directly into text without changing your workflow. Key features: - Local-first transcription with downloadable ONNX models - Two recognition modes: compact SenseVoice final transcription or bilingual X-ASR partial transcription with punctuation and English casing - Personal dictionary for realtime ASR hotwords and optional LLM-assisted corrections - Optional text cleanup using your own API key - Translation in place for selected text - Desktop-native workflow with configurable hotkeys, microphone selection, and status HUD - Local history for search, copy, inspect, and delete transcripts - English and Chinese interface Tech stack: TypeScript Why it matters: Voice Key simplifies voice input by integrating seamlessly with your existing workflow, providing real-time transcription without requiring external APIs, and offers a user-friendly interface that supports both English and Chinese. GitHub stars: 181 Repository: BuildWithAIs/voicekey

Overview: FairyTailor is a multimodal generative framework designed for children's storytelling, combining text and image generation to create coherent and creative narratives. Key features: - Co-creation between human writers and generative models - Integration of visual storytelling - Accessible through a Vue 2.6 client and a FastAPI backend - Large data management using DVC - Updated September 17, 2023 Tech stack: Python, FastAPI, HuggingFace Transformers, CLIP, NLTK, DVC Why it matters: FairyTailor offers a unique approach to storytelling by leveraging both text and images, providing a more engaging and interactive experience for children.

### Overview: DreamCube is an official implementation of the paper "DreamCube: 3D Panorama Generation via Multi-plane Synchronization". It addresses the challenge of generating 3D panoramic representations from single-view inputs using diffusion models, specifically focusing on RGB-D cubemap generation. ### Key Features: - **Multi-plane Synchronization**: Adaptation of 2D diffusion models to generate multi-plane panoramic representations. - **DreamCube Framework**: A diffusion-based framework for generating RGB-D cubemap representations from single-view inputs. - **Model Weights**: Available on HuggingFace for inference and training. ### Tech Stack: - Python ### Why It Matters: DreamCube provides a practical solution for generating high-quality 3D panoramic images and scenes from single-view inputs, making it useful in various applications such as augmented reality, virtual reality, and scene generation tasks. ### Repository: - **Repository**: Yukun-Huang/DreamCube - **GitHub Stars**: 182 - **Project Page**: [Project Page](https://yukun-huang.github.io/DreamCube/) - **Model Weights**: [Model Weights](https://huggingface.co/KevinHuang/DreamCube) - **Inference Code**: [Inference Code](https://github.com/Yukun-Huang/DreamCube) ### News: - **Release Notes**: - **2025-11-05**: More Jupyter notebooks for Multi-plane synchronization on SANA, Lotus, and VideoCrafter2. - **2025

**Overview:** This project is a comprehensive collection of RAG interview questions and answers, covering 52 retrieval augmented generation (RAG) architectures, system design scenarios, vector databases, embeddings, chunking, reranking, evaluation, and production failure modes. The repository includes 1619 questions and 52 RAG types, designed to help AI engineers, ML engineers, and GenAI/LLM developers prepare for interviews. **Key features:** - 1619 RAG interview questions and answers - 52 RAG architectures - System design scenarios - Vector databases - Embeddings - Chunking - Reranking - Evaluation - Production failure modes **Tech stack:** JavaScript **Why it matters:** This resource is essential for anyone preparing for interviews in the field of AI, machine learning, and generative language models. It provides a thorough understanding of the RAG interview process and helps candidates prepare effectively for real-world challenges they might encounter in their roles. **Repository:** athere-tech/rag-interview-system **Primary language:** JavaScript **GitHub stars:** 182 **GitHub description:** A complete collection of RAG interview questions, answers (1619 questions & 52 RAG types), system design scenarios, architecture patterns, and production-ready concepts. **GitHub description excerpt:** This repository is a comprehensive collection of RAG interview questions, answers, system design scenarios, architecture patterns, and production-ready concepts. It is designed to help AI engineers,

Overview: GEO（Generative Engine Optimization）智能内容优化平台是一款帮助品牌在AI搜索引擎中被优先、准确、可信地提及的工具，旨在提升品牌曝光、优化搜索排名、建立权威形象并提供数据驱动的内容策略。 Key features: - 关键词策略：智能关键词生成、搜索意图覆盖、数据增强、语义扩展 - 内容生成：20+平台模板、E-E-A-T内嵌、品牌植入策略、结构化输出 - 知识库（RAG）：文档管理、智能检索、来源保障 - 内容优化：E-E-A-T强化、事实密度增强、结构化数据、内容评分 - 多模型验证：DeepSeek、OpenAI、通义千问等多模型交叉验证 - AI搜索验证：接入Perplexity真实搜索引擎验证 - 竞品对比：同时验证自有品牌和竞品 - 负面防护：主动生成负面查询并监控 - 数据分析：提及率趋势、平台贡献度、ROI分析、内容独特性 - 平台同步：一键发布到GitHub、12平台 - 工作流自动化：一键完成全流程 - 资源库：优化资源推荐、品牌知识库 Tech stack: Python, Streamlit, LangChain, GitHub API Why it matters: 通过GEO平台，品牌能够显著提高在AI搜索引擎中的曝光率，优化搜索排名，建立权威形象，并通过数据驱动的内容策略提升品牌表现。