Projects

Overview: Kheish is an orchestration engine designed for long-running, tool-using AI agents, ensuring their execution is durable, survives interruptions, and is governed by named authorities. Key features: - Orchestration of AI agent runs - Separation of agent and caller - Survives interruptions (process restart, machine reboot, network blips, etc.) - Governed by named authorities (session, sidechain, lease) - Persistent agent runtime and control plane - Multi-provider routing and skills - Documentation available at docs.kheish.ai Tech stack: Rust Why it matters: Kheish ensures that AI agent runs remain durable, survive interruptions, and are governed by named authorities, making them independent of any caller, process, or human reachability. Repository: graniet/kheish Primary language: Rust GitHub stars: 217

Overview: Aimirror is a powerful tool designed to significantly speed up downloads by leveraging AI and multiple sources. It provides parallel downloads, intelligent caching, dynamic routing, and supports Docker, PyPI, CRAN, and Hugging Face for seamless acceleration. Key features: - **⚡ Parallel Downloads**: HTTP Range Range, multi-threaded concurrency, maximizing bandwidth. - **💾 Intelligent Caching**: File digest-based deduplication, LRU eviction. - **🎯 Dynamic Routing**: Small files proxy, large files parallel download. - **🔗 Multi-source Support**: PyPI, Docker Hub, CRAN, Hugging Face out of the box. - **🔌 Arbitrary Extension**: Supports any HTTP download, configured rules can accelerate by up to 200x. - **📝 Content Rewrite**: Automatically rewrite HTML/JSON links for seamless proxying. - **🎛️ Special Handling**: Customizable handlers for complex scenarios. - **🚦 Concurrent Control**: Global download concurrency limit. - **🔄 Path Rewriting**: Flexible path replacement rules. - **🔑 Cache Optimization**: Supports original URL as cache key. Why it matters: Aimirror is essential for AI engineers who face slow internet connections, providing a solution that can significantly boost download speeds and efficiency. Tech stack: Aimirror is built using Python, FastAPI, Docker, PyPI, CRAN, and Hugging Face, leveraging AI to accelerate downloads across multiple sources. GitHub stars: 216, indicating the tool's popularity and usefulness among developers. Repository: livehl/aimirror

Overview: MKGFormer is a hybrid transformer model designed for multimodal knowledge graph completion, presented at SIGIR 2022. It includes various baselines and a new multimodal analogical reasoning task at ICLR 2023. Key features: - Multi-level Fusion for Knowledge Graph Completion - Unified Multimodal KGC Framework - Novel MKG Task: Multimodal Analogical Reasoning over Knowledge Graphs Tech stack: Python Why it matters: MKGFormer offers a comprehensive solution for knowledge graph completion, demonstrating significant improvements over existing methods. Repository: zjunlp/MKGformer Primary language: Python GitHub stars: 218

### Overview: ETOS LLM Studio 是一个原生的 AI / Agent 客户端，支持 OpenAI、Anthropic Claude、Google Gemini 和本地 GGUF 模型。它提供本地 RAG、MCP、Siri Shortcuts、Agent Skills、跨会话协作、本地 RAG 记忆、实时活动与双端同步等功能。 ### Key features: - **本地模型支持**：支持导入本地 GGUF 权重并作为“本地模型”提供商使用。 - **会话管理增强**：支持会话全文检索、上下文预览、消息序号定位、文件夹分类、彩色标签、快捷筛选、批量操作、单会话跨端发送、全屏会话管理入口与单会话跨端发送。 - **多模型支持**：原生适配 OpenAI Chat、OpenAI Responses、Anthropic（Claude）和 Google（Gemini）等接口格式。 - **端侧本地模型**：支持流式输出、GGUF Jinja chat template、本地工具调用解析、思考内容解析、本地嵌入模型路由与后台 detached completion。 - **本地模型高级调参**：支持 LoRA GGUF Adapter 并调节强度，按需覆盖上下文长度、输出上限、GPU 层数、batch / ubatch、KV offload、flash attention、seed、采样链、grammar、重复惩罚与对话模板透传等参数。 - **高级请求配置**：支持自定义请求头、参数表达式、结构

Overview: This repository contains a growing collection of AI blueprint projects that run end-to-end using Jupyter notebooks, MLflow deployments, and Streamlit web apps. These projects are built using HP AI Studio and are designed to be easily sharable and reproducible. Key features: - Jupyter notebooks for end-to-end AI projects - MLflow for model deployment - Streamlit for web application development - HP AI Studio for project setup and collaboration Tech stack: Jupyter Notebook Why it matters: These projects showcase the power of Jupyter notebooks in building and deploying AI solutions, making them accessible and reusable for developers and researchers alike.

Overview: DiffuseStyleGesture is a project that leverages diffusion models to generate stylized audio-driven co-speech gestures. This entry to the GENEA Challenge 2023 showcases significant improvements in speech-gesture alignment and motion realism. Key features: - Uses diffusion models for generating stylized gestures. - Achieves high Speech-Gesture Alignment. - Supports training on multiple gesture datasets for improved gesture refinement. - Provides pre-trained models and code for easy integration into projects. Tech stack: Python Why it matters: This project demonstrates the potential of diffusion models in generating realistic and stylized gestures, making it a valuable tool for researchers and developers in the field of gesture generation and animation. Repository: YoungSeng/DiffuseStyleGesture GitHub stars: 216 GitHub description: DiffuseStyleGesture: Stylized Audio-Driven Co-Speech Gesture Generation with Diffusion Models | The DiffuseStyleGesture+ entry to the GENEA Challenge 2023 (ICMI 2023, Reproducibility Award)

### Overview: This repository is a curated list of papers on pre-training for graph neural networks (Pre-train4GNN), categorized by their published years, graph types, pre-training strategies, tuning strategies, and applications. It aims to provide a comprehensive overview of the latest research in this field. ### Key Features: - **Graph Types**: Covers various types of graphs, including Multiplex, Dynamic, Static, Multi-graphs, Hypergraphs, and KG. - **Pre-training Strategies**: Includes Contrastive, Generative, and MTL. - **Tuning Strategies**: Prompts and N/A. - **Applications**: General, Recommendation, Biology, Chemistry&Biology, and more. ### Tech Stack: The repository uses Markdown, Git, and GitHub for its organization and documentation. ### Why It Matters: This curated list is essential for researchers and developers interested in graph neural networks, providing a clear and up-to-date overview of the current research trends and applications in the field.

Overview: `pdf2md` is a Python library and command-line tool that converts PDFs into clean, structured Markdown. It extracts headings, lists, tables, code, math, and images from PDFs, providing a high-quality Markdown output. Key features: - Extracts document-wide font-size statistics for headings - Supports paragraphs, bold/italic text, inline code, super/subscript, links, lists, tables, code blocks, and math - Handles images, footnotes, table of contents, and CJK/Unicode text - Supports multi-column reading order, rotated/landscape pages, and header/footer stripping - Provides a progress bar for long conversions Tech stack: Python Why it matters: `pdf2md` is useful for developers who need to extract structured information from PDFs for documentation, reports, or other content that requires clean, readable Markdown. It simplifies the process of converting PDFs into a format that can be easily integrated into other projects or platforms. Repository: iamarunbrahma/pdf-to-markdown Primary language: Python GitHub stars: 213

### Overview: semble_rs is a Rust implementation of the code search and dependency analysis engine [semble](https://github.com/MinishLab/semble), designed specifically for AI coding agents. It provides fast, accurate code search capabilities, replacing traditional commands like `grep`, `cat`, and `ls` with a more efficient and token-friendly approach. ### Key Features: - **Fast Indexing**: Local repository indexing in 150 ms for 22 files, 10 seconds for 1,600 files. - **Token-Efficient**: `tree` reduces `ls -R` by 4×–747×, `--outline` is -47% vs full output, and `digest` reaches -98.9% on real GitHub Actions logs. - **Hybrid Retrieval**: Combines BM25 with Model2Vec embeddings, then reranks with definition, identifier-stem, and file-coherence boosts. - **Dependency Graph**: Shows imports, definitions, and impact of file changes, with optional Graphviz output. - **Build / CI Compression**: Automatically compresses build and CI output. - **Single Binary**: No Python, no daemon, no API keys. Runs on CPU. ### Tech Stack: semble_rs is built on Rust and uses static embeddings and semantic code search. ### Why It Matters: - **Token Savings**: Up to -99% token reduction compared to traditional commands. - **Efficient Indexing**: Fast and accurate code search with minimal token cost. - **Dependency Analysis**: Provides a

### Overview: This project is an official implementation of the NeurIPS 2025 paper "Iterative Self-Incentivization Empowers Large Language Models as Agentic Searchers". The code aims to enable Large Language Models (LLMs) to act as search agents, actively seeking information, selecting key knowledge, and summarizing the final answer iteratively. This approach differs from previous Retrieval-Augmented Learning (RAG) methods by incorporating reasoning techniques, making it more effective for search tasks. ### Key Features: - **ExSearch**: An agent-based search method that leverages LLMs to actively seek and summarize information. - **Iterative Self-Incentivization**: A technique that encourages LLMs to continuously improve their search performance through self-incentive. - **Key Knowledge Selection**: The ability to select the most relevant and useful knowledge from a corpus of documents. - **Agentic Search**: The model iteratively seeks information, selects key knowledge, and summarizes the final answer. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. - **ColBERT**: The retrieval model used to pair queries with top-20 documents. - **VLLM**: The language model used for the agent-based search. - **PyTorch**: The deep learning framework used for training the model. ### Why It Matters: - **Improves Search Efficiency**: By enabling LLMs to actively seek and summarize information, this project significantly enhances the efficiency and effectiveness of search tasks. - **Enhances Knowledge Selection**: The ability to select key

### Overview: Cute 是一个面向全网研究与复杂数据分析的云端智能体，能够研究公开网络、YouTube、BiliBili、X、小红书、Hacker News 等平台上的任何主题。 ### Key features: - **全域采集**：调度多个智能体并行采集网页信息、Reddit 社区讨论、X 平台动态、YouTube 字幕内容、小红书互动数据，以及由真实资金与内幕信息定价的 Polymarket 预测赔率。 - **私有文档证据链**：通过多模态 RAG 解析用户私有的 PDF、Word、PPT 与图片，把公域情报与内部知识放进同一条证据链。 - **长期记忆**：沉淀偏好、关键事实与可复用流程，让下一次任务不必从头开始。 - **分析与交付**：结合 E2B 沙箱的代码执行能力与专属数据分析 Agent，完成数据清洗、交叉验证与深度分析，最终以 GenUI 交互画布和 PDF、Word、PPT、HTML 等多格式报告自动交付。 ### Tech stack: Java ### Why it matters: Cute 通过跨平台全域检索、私有文档证据链、长期记忆等功能，解决了信息孤岛问题，为程序员群体提供了高效的信息获取工具，帮助他们及时、完整地捕捉全社区的原生技术情报，同步跟进最前沿的技术方向。

Overview: OneJev is a multimodal system designed to provide calibrated answers to typed questions about screenshots, photos, videos, and text in a single forward pass. It leverages the Qwen model for its decision-making capabilities. Key features: - Multimodal input support for text, screenshots, photos, and videos. - Calibration of probabilities for precise decision-making. - Available in 0.8B, 4B, 9B, and 27B models. Tech stack: Python Why it matters: OneJev simplifies complex decision-making processes by providing accurate and calibrated answers to questions about various media types, making it a valuable tool for applications requiring precise and reliable decision-making based on multimodal inputs. Repository: OmniJev/OneJev GitHub stars: 210 GitHub description: 🚀🚀 A multimodal System One decision model that gives calibrated answers to typed questions about screens, photos, video and text in one forward pass. GitHub description: OneJev, a Multimodal System One Decision Model. 🚀🚀 A model that gives calibrated answers to typed questions about screenshots, photos, videos, and text in a single forward pass. Available in 0.8B, 4B, 9B, and 27B models. 🚀🚀 GitHub description: OneJev, a Multimodal System One Decision Model. 🚀🚀 A model that gives calibrated answers to typed questions about screenshots, photos, videos, and text in a single forward pass. Available in 0

### Overview: AsyncDiff is a universal and plug-and-play diffusion acceleration scheme that enables model parallelism across multiple devices. Our approach divides the cumbersome noise prediction and diffusion process into parallelizable tasks, significantly reducing inference time and improving overall performance. ### Key Features: - **Parallel Noise Prediction and Diffusion**: AsyncDiff splits the noise prediction and diffusion process into parallelizable tasks, allowing for efficient parallel execution across multiple devices. - **Model Parallelism**: It enables model parallelism across multiple GPUs, effectively utilizing the computational resources of each device. - **Inference Speedup**: On various models, AsyncDiff can achieve up to 2.8x faster inference speed compared to traditional methods, with notable improvements on models like SDXL and Stable Diffusion 3 Medium. - **Support for Multiple Models**: AsyncDiff supports a wide range of diffusion models, including Stable Diffusion 3 Medium, Stable Diffusion 2.1, Stable Diffusion 1.5, Stable Diffusion x4 Upscaler, Stable Diffusion XL 1.0, Stable Diffusion XL Inpainting, and ControlNet. ### Tech Stack: AsyncDiff is implemented using Python, leveraging the PyTorch library for model inference and parallel processing. ### Why It Matters: AsyncDiff offers a significant performance improvement across various diffusion models, making it an essential tool for accelerating the inference process in applications like image generation and inpainting. Its universal design and plug-and-play nature make it a valuable addition to any diffusion model toolkit.

### Overview: MakerAI is a comprehensive AI operating system designed for Delphi developers, offering a full-fidelity API wrapper and a complete ecosystem for building intelligent systems. ### Key Features: - **Native Components**: Direct access to provider-specific components for full-fidelity interaction. - **RAG Pipelines**: Vector and graph-based reasoning with SQL-like query languages. - **Autonomous Agents**: Graph-based orchestration with checkpoints and human-in-the-loop approval. - **MCP Servers and Clients**: Expose or consume tools using the Model Context Protocol. - **ChatTools**: Bridge AI reasoning with deterministic real-world capabilities. - **Skills**: Reusable instructions in the SKILL.md format. - **A2A**: Expose agent graphs to other agents and call remote ones. - **Production Controls**: Tool-call guardrails, evals, OpenTelemetry tracing, persistent memory, and calibrated decisions with Jev. - **FMX Visual Components**: Drop-in UI for multimodal chat interfaces. - **Universal Connector**: Switch providers at runtime without changing application code. ### Tech Stack: Delphi 10.4+ (limited), full support from Delphi 12 Athens. ### Why it Matters: MakerAI provides a complete AI application ecosystem for Delphi, allowing developers to build production-grade intelligent systems entirely in Delphi. It includes native, provider-specific components, RAG pipelines, autonomous agents, MCP servers, chat tools, skills, A2A, production controls, and FMX visual components, making it a versatile solution for various AI integration needs.

Overview: This plugin integrates Azure Speech Cognitive Services into Unreal Engine, allowing developers to perform text-to-speech and speech-to-text functionality through asynchronous tasks. Key features: - Text-to-speech functionality - Speech-to-text functionality - Editor tool for generating audios as USoundWaves directly in the Engine Tech stack: Unreal Engine C++ Why it matters: This plugin simplifies the integration of Azure Speech Cognitive Services into Unreal Engine, making it easier for developers to implement text-to-speech and speech-to-text functionalities without requiring extensive Azure Cognitive Services knowledge.

### Overview: noai-watermark is a powerful tool designed to remove invisible AI watermarks and strip AI metadata from images. It uses diffusion-based image regeneration to disrupt watermark patterns and provides a way to inspect and clean up metadata. The tool is open-source and can be used with both CLI and Python APIs, making it accessible to developers and enthusiasts alike. ### Key Features: - **Watermark Removal**: Automatically removes invisible AI watermarks (SynthID, StableSignature, TreeRing) and strips AI metadata from images. - **Regeneration**: Regenerates image pixels to disrupt watermark patterns. - **Metadata Inspection**: Provides separate metadata inspection and cleanup functionalities. - **Compatibility**: Works with various AI image generators and can regenerate image pixels to remove watermarks. ### Tech Stack: - Python ### Why It Matters: noai-watermark is essential for developers and content creators who need to ensure their images are free from AI-generated watermarks, maintaining the integrity and originality of their work. It simplifies the process of verifying and removing AI-generated metadata, making it easier to share and distribute content without the risk of infringing on others' intellectual property rights. ### Repository: mertizci/noai-watermark - **Primary Language**: Python - **GitHub Stars**: 213 - **GitHub Description**: Remove invisible AI watermarks (SynthID, StableSignature, TreeRing) and strip AI metadata from images. Open-source CLI & Python toolkit. - **README Excerpt**: A screenshot of the tool's functionality is provided, demonstrating how it removes watermarks

### Overview: Access and integrate the Gemini API into your .NET applications with the Gemini SDK, designed to enable developers to utilize Google's state-of-the-art generative AI models for feature development and application creation. ### Key Features: - Connects to the Gemini API through Google AI Studio or Vertex AI. - Fully compatible with all Gemini API models and features, including recent improvements in tool usage and media generation. - Supports recent additions like improved code execution, function calling, and integrated Google search grounding. - Provides enhanced media generation capabilities, including improved image creation. ### Tech Stack: .NET and ASP.NET Core ### Why It Matters: The Gemini SDK simplifies the integration of cutting-edge AI technology into your .NET applications, enabling developers to build intelligent features and applications efficiently.

Overview: MQL is a tool designed to convert natural language queries into executable SQL queries, making it accessible even without coding knowledge. It supports PostgreSQL and can be run locally using Docker. Key features: - Converts natural language queries into SQL - Supports PostgreSQL database - Runs locally using Docker Tech stack: Python Why it matters: MQL simplifies database querying by translating natural language inputs into SQL, making it easier for non-technical users to interact with databases. Repository: shurutech/mql GitHub stars: 214

Overview: GuardianDB is a high-performance, local-first decentralized database built on Rust and Iroh. It stores and processes data, and runs business logic on the network. Key features: - Local-first & P2P: Every node holds a full replica; reads and writes are local, changes sync automatically over Iroh. - Three store types: key-value, document, and append-only event log (LWW CRDT + causal DAG). - PostgreSQL-compatible SQL: the real Postgres wire protocol: psql, TypeORM, node-postgres, and DBeaver connect unchanged. - Supabase-compatible gateway: a Kong-shaped HTTP layer over the SQL engine, REST, Auth (JWT), GraphQL, Storage and Realtime. - Vector search & RAG: pgvector-compatible HNSW index, auto-embedding, and retrieve-augment-generate right beside your data. - Guardian Compute: delegate WebAssembly tasks to the most capable peer: capacity-aware scheduling, Edge AI, and LLM inference. - Guardian Sentinel: a terminal UI to inspect, manage, and monitor a live database. - Optional ODM and collection API: TypeORM/Mongoose-style models, collections, and indexes over the document store. - Encryption: opt-in payload codecs and an encrypted keystore for confidentiality even from replica holders. Tech stack: Rust Why it matters: GuardianDB provides a decentralized, local-first database solution for apps that require peer-to-peer synchronization, offline-first operation, and high performance. It stores and processes data, runs business logic on the network, and is compatible

### Overview: CoMoSpeech is a one-step speech and singing voice synthesis method that leverages a consistency model to achieve high-quality audio synthesis with significantly improved inference speed compared to traditional diffusion models. ### Key Features: - **CoMoSpeech**: A method for speech and singing voice synthesis that uses a consistency model to distill a high-quality model for real-time inference. - **FlashSpeech**: An efficient zero-shot speech synthesizer based on the latent consistency model and adversarial training. - **CoMoSVC**: A well-designed singing voice conversion (SVC) version based on the consistency model. ### Tech Stack: - **Python** - The primary programming language used for implementing the CoMoSpeech method. ### Why It Matters: CoMoSpeech offers a significant improvement in inference speed, making speech synthesis practical and efficient. It achieves high audio quality by distilling a high-quality model from a well-designed diffusion-based teacher model, making it comparable to state-of-the-art methods in terms of both inference speed and audio quality.