Projects

### Overview: Bottleneck Transformer is an advanced visual recognition model that significantly outperforms EfficientNet and DeiT in terms of performance-computes trade-off. This implementation is built using PyTorch and is designed for efficient and effective visual recognition tasks. ### Key features: - **Bottleneck Transformer Model**: A state-of-the-art model that combines convolutional and attention mechanisms to achieve high performance. - **Convolution + Attention Architecture**: Utilizes convolutional layers for feature extraction and attention mechanisms for capturing global dependencies. - **Efficient Computation**: Designed to be highly efficient in terms of both performance and computational resources. - **Performance-Computes Trade-off**: Outperforms EfficientNet and DeiT in terms of performance-computes trade-off, making it suitable for various applications. ### Tech stack: PyTorch ### Why it matters: Bottleneck Transformer is a groundbreaking model that leverages PyTorch to achieve exceptional performance in visual recognition tasks. Its efficient architecture and high performance make it a valuable tool for researchers and practitioners looking to improve the accuracy of their models while maintaining high computational efficiency.

### Overview: Frozen — cross-agent persistent memory for coding assistants. Compressed-memory core now lives inside **caveman**. ### Key Features: - **Persistent Memory Across Sessions**: Hooks capture what happened, and the store keeps it. - **Compressed at Rest**: Deterministic caveman grammar, round-trip-guaranteed expansion for humans. - **Progressive MCP Retrieval**: `search`, `timeline`, `get_observations` — agents filter before fetching. - **Hybrid Search**: SQLite FTS5 keyword + local vector index, combined with a tunable ranker. - **Local by Default**: No network calls. Optional remote embedding providers via config. - **Web Viewer**: Read-only UI at `http://localhost:37777` for browsing sessions in human-readable form. - **Cross-IDE Installers**: Claude Code, OpenCode, Codex, GitHub Copilot, Augment Code capture observations; Cursor, Gemini CLI, Antigravity, IBM Bob are query-only (MCP search over memory captured elsewhere) — one command each. - **Privacy-Aware**: `<private>...</private>` stripped at write boundary. Path globs exclude whole directories. ### Tech Stack: TypeScript ### Why It Matters: Frozen — cross-agent persistent memory for coding assistants. Hooks fire at session boundaries, compress observations with the caveman grammar (~75% fewer prose tokens, code and paths preserved byte-for-byte), and write to local SQLite. Agents query their own history through three MCP tools. No network. No cloud.

### Overview: This plugin provides a seamless integration of Generative AI models into Unreal Engine, allowing developers to focus on game development without the complexities of integrating various AI models. ### Key Features: - **OpenAI GPT-5** - **Deepseek R1** - **Claude Opus/Sonnet** - **Gemini 3** - **Grok 4** - **Alibaba Qwen** - **Kimi** - **ElevenLabs TTS** - **Inworld** - **OpenRouter** - **GLM** - **Ollama** - **Local** - **Meshy** - **Tripo** - **Hunyuan3D** - **Rodin** - **Dashscope** - **Seedream** - **NPC AI** - **agentic** - **chat** - **3D gen** - **TTS** - **multimodal** - **image gen** ### Tech Stack: C++ ### Why It Matters: This plugin simplifies the integration of Generative AI models into Unreal Engine, making it easier for developers to focus on game development without the complexities of integrating various AI models.

### Overview: This repository is a collection of papers and codes related to Computer Vision and Artificial Intelligence (AIGC) from various conferences and workshops. It includes papers and codes from CVPR 2026, CVPR 2025, ICCV 2025, CVPR 2024, ECCV 2026, and ECCV 2024. The repository is maintained by Kobaayyy and is a valuable resource for researchers and practitioners in the field of computer vision and artificial intelligence. ### Key Features: - Papers and codes from CVPR 2026 - Papers and codes from CVPR 2025 - Papers and codes from ICCV 2025 - Papers and codes from CVPR 2024 - Papers and codes from ECCV 2026 - Papers and codes from ECCV 2024 ### Tech Stack: Unknown ### Why it Matters: This repository serves as a comprehensive collection of papers and codes related to computer vision and artificial intelligence, making it an invaluable resource for researchers and practitioners in the field. It provides a platform to share and access cutting-edge research in the latest conferences and workshops, facilitating collaboration and knowledge sharing among the community. ### GitHub Stars: 675 ### GitHub Description: A Collection of Papers and Codes for CVPR2026/CVPR2025/ICCV2025/CVPR2024/ECCV2026/ECC

### Overview: `gollm` is a Go package designed to simplify the integration of Large Language Models (LLMs) into your projects. It offers a unified, flexible interface that allows developers to interact with various LLM providers, including OpenAI, Anthropic, Groq, Ollama, and OpenRouter. This package streamlines the process of configuring and managing LLMs, making it easier to craft and deploy AI services. ### Key Features: - **Unified API for Multiple Providers:** Interact with multiple LLM providers seamlessly, including OpenAI, Anthropic, Groq, Ollama, and OpenRouter. - **Easy Provider and Model Switching:** Configure preferred providers and models with simple options. - **Flexible Configuration Options:** Customize using environment variables, code-based configuration, or configuration files. - **Advanced Prompt Engineering:** Craft sophisticated instructions to guide AI responses effectively. - **Prompt Optimizer:** Automatically refine and improve prompts for better results, with support for custom metrics and different rating systems. - **Memory Retention:** Maintain context across multiple interactions for more coherent conversations. - **Structured Output and Validation:** Ensure outputs are consistent and reliable with JSON schema generation and validation. - **Provider Comparison Tools:** Test performance across different LLM providers and models for the same task. - **High-Level AI Functions:** Use pre-built functions like `ChainOfThought` for complex reasoning tasks. - **Robust Error Handling and Retries:** Built-in retry mechanisms to handle API rate limits and transient errors. - **Extensible Architecture:** Easily expand support for new LLM providers

### Overview: HunyuanImage-2.1 is an efficient diffusion model designed for high-resolution text-to-image generation, capable of producing 2K (2048 × 2048) images with cinematic composition. It leverages an extensive dataset and structured captions to enhance text-image alignment and significantly reduces computational costs through a highly expressive VAE with a (32 × 32) spatial compression ratio. ### Key Features: - **High-Quality Generation**: Efficiently produces ultra-high-definition (2K) images with cinematic composition. - **Multilingual Support**: Provides native support for both Chinese and English prompts. - **Advanced Architecture**: Built on a multi-modal, single- and dual-stream combined DiT (Diffusion Transformer) backbone. - **Glyph-Aware Processing**: Utilizes ByT5's text rendering capabilities for improved text generation accuracy. - **Flexible Aspect Ratios**: Supports a variety of image aspect ratios (1:1, 16:9, 9:16, 4:3, 3:2). ### Tech Stack: - Python ### Why it Matters: HunyuanImage-2.1 stands out by significantly improving text-image alignment, enhancing text rendering across multiple languages, and providing high-quality, high-resolution image generation. Its efficient model architecture and advanced processing techniques make it a powerful tool for generating realistic and high-quality images from text inputs.

Overview: Denoising Diffusion Restoration Models (DDRM) is a project that leverages pre-trained diffusion models to solve general linear inverse problems efficiently and without problem-specific supervised training. Key features: - Uses pre-trained diffusion models for solving inverse problems - Efficient and unsupervised approach - Supports various types of degeneration and noise levels Tech stack: Python Why it matters: DDPMs offer a powerful framework for solving inverse problems, and by leveraging pre-trained models, DDPMs can be used to restore images without needing to train on specific datasets, making them highly versatile and efficient. Repository: bahjat-kawar/ddrm Primary language: Python GitHub stars: 674

### Overview: Surf is a zero-config, agent-agnostic CLI for controlling Chrome, built by reverse-engineering production extensions. It offers battle-tested features like smart defaults, AI without API keys, and seamless integration with various AI providers. ### Key Features: - **Agent-Agnostic**: Pure CLI commands over Unix socket, works with Claude Code, GPT, Gemini, Cursor, custom agents, shell scripts - anything that can run commands. - **Zero Config**: Install the extension, run commands. No MCP servers to configure, no relay processes, no subscriptions. - **Battle-Tested**: Built by reverse-engineering production browser extensions and methodically working through agent-hostile pages like Discord settings. Falls back gracefully when CDP fails. - **Smart Defaults**: Screenshots auto-resize to 1200px (saves tokens). Actions auto-capture screenshots (saves round-trips). Errors on restricted pages warn instead of fail. - **AI Without API Keys**: Query ChatGPT, Gemini, Perplexity, and Grok using your existing browser logins. No API keys needed. - **Network Capture**: Automatically logs all network requests while active. Filter, search, and replay API calls without manually setting up request interception. - **Free**: Works with Claude Code, GPT, Gemini, Cursor, custom agents, shell scripts - anything that can run commands. - **Whole Threads and Feeds as Records**: Via the site's own API, not just the current page. ### Tech Stack: - TypeScript ### Why It Matters: - **

Overview: fenic is a semantic DataFrame engine that turns AI-assisted exploration of structured and unstructured data into reusable, inspectable pipelines. It integrates semantic operators like `extract`, `classify`, `summarize`, and `embed` into PySpark/SQL-style operations, enabling humans and agents to work on the same pipelines. Key features: - Semantic DataFrame engine - Integrates AI operators into PySpark/SQL-style operations - Enables humans and agents to work on the same pipelines - Provides typed, inspectable, rerunnable, and callable artifacts Tech stack: Python Why it matters: fenic transforms data exploration into durable, reusable, and programmable pipelines, making it easier to author, inspect, and reuse data work. Repository: typedef-ai/fenic Primary language: Python GitHub stars: 674 GitHub description: Semantic DataFrames for humans and agents

Overview: RAGLight is a modular framework for Retrieval-Augmented Generation (RAG) that simplifies the integration of different LLMs, embeddings, and vector stores. It now includes seamless MCP integration for connecting external tools and data sources. Key features: - Modular integration of LLMs, embeddings, and vector stores - Seamless MCP integration for connecting external tools and data sources - Easy setup with no third-party runtime dependency Tech stack: Python Why it matters: RAGLight simplifies the integration process, making it easier to experiment with various LLMs, embeddings, and vector stores, and connects external tools and data sources seamlessly.

### Overview: ClearAI is a domain ontology discovery and exploration platform built on the principles of evidence-based research. It allows researchers to grow a continuously evolving domain ontology through a structured epistemic loop, ensuring that each conclusion is backed by evidence and verified through a rigorous process. ### Key Features: - **Domain Ontology (What You Get):** Your project's vocabulary, knowledge entries established through the loop, and their graphs. The ontology grows as you research, providing a retrievable knowledge structure. - **Epistemic Loop (How You Get It):** A question → ontology → candidate hypotheses → evidence → conclusion → growing ontology. Every edge is tested by evidence and independent evaluation. ### Tech Stack: JavaScript ### Why It Matters: ClearAI is essential for researchers who need to build a comprehensive and evidence-based domain ontology. It ensures that conclusions are supported by verified evidence, making the research process more reliable and transparent. ### Repository: - **Repository:** Clearailhc/clearai-dsh - **Stars:** 670 - **Description:** ClearAI: an ontology-first DSH agent preset — evidence-checked research that grows into a domain ontology. - **Primary Language:** JavaScript - **GitHub Badge:** [Trendshift Badge](https://trendshift.io/repositories/248415?utm_source=trendshift-badge&amp;utm_medium=badge&amp;utm_campaign=badge-trendshift-248415) ### Installation: ```bash dsh plugin --profile web add clearai-dsh@0.5

### Overview: 本项目提供了支持中文的SpaCy模型，旨在推动中文SpaCy模型的开发。模型目前处于测试阶段，具有部分中文特性和在线演示功能。用户需要具备基础的SpaCy知识，并提供下载和安装模型的说明。此外，项目还提供了如何从零构造模型的指导和语料库信息。 ### Key features: - 中文模型 - 在线演示 - 特性展示 - NER功能 - 系统要求 - 模型安装 - Demo代码运行 - 语料库信息 - TODO列表 - 使用组件 - 如何贡献 - 版本控制 ### Tech stack: Python 3 ### Why it matters: 推动了中文SpaCy模型的开发，提供了中文特性和在线演示功能，方便用户使用和调试。

Overview: NeuralKart is a real-time Mario Kart 64 AI project that uses Convolutional Neural Networks (CNNs) for AI, offline search, and DAGGER (Deep Architecture for Generalized Reinforcement Learning) techniques. The project is designed to run on Windows and is based on BizHawk emulator. Key features: - Real-time AI for Mario Kart 64 - Uses CNNs for AI - Implements offline search for improved performance - Utilizes DAGGER for reinforcement learning - Runs on Windows Tech stack: Lua Why it matters: This project showcases the integration of advanced AI techniques with real-time gaming, demonstrating how machine learning can enhance gameplay experience.

Overview: Stirrup is a lightweight framework designed for building agents, offering a model-centric approach that avoids rigid workflows and incorporates best practices and tools for context management and agent customization. Key features: - Code execution - Online search/web browsing - MCP client support - Document input/output - Skills system - Flexible tool execution - Human-in-the-loop - Context management - Flexible provider support - Multimodal support Tech stack: Python Why it matters: Stirrup simplifies agent development by providing a model-centric approach, integrating best practices, and offering flexibility in tool execution and customization, making it an ideal choice for developers looking to build agents efficiently and effectively. Repository: ArtificialAnalysis/Stirrup Primary language: Python GitHub stars: 667 GitHub description: The lightweight framework for building agents

### Overview: ViewComfy is an open-source tool designed to help developers create beautiful web applications using ComfyUI workflows. It supports multiple workflows in the same UI and seamlessly integrates with image, video, and text outputs. ### Key features: - **Input Editor**: Configures how each workflow input appears and exposes only necessary information. - **Mask Editor**: Easily adds masks to image inputs. - **Complex Workflow Behaviors**: Supports complex workflow behaviors like workflow routing and optional images. - **User Management and Analytics**: Built-in user management and analytics for apps deployed on ViewComfy Cloud. - **Serverless API Connection**: Runs workflows on serverless infrastructure and connects them via API to your apps. - **App Hub**: Built-in hub with all of your apps for apps deployed on ViewComfy Cloud. - **History**: Built-in user output history for apps deployed on ViewComfy Cloud. - **Web-Hosted Editor**: Allows connecting ViewComfy to ComfyUI or workflows deployed on ViewComfy cloud. - **Detailed Installation Guide**: Available for a more comprehensive setup guide. ### Tech stack: Node.js v20.18 ### Why it matters: ViewComfy simplifies the process of creating and deploying web applications using ComfyUI workflows, making it easier for developers to build visually appealing and feature-rich applications. The tool's integration with serverless infrastructure and built-in analytics help streamline the development process and improve the overall user experience.

Overview: Ekphrasis is a text processing tool designed for social network text, including tokenization, word normalization, word segmentation (for hashtags), and spell correction, using word statistics from two large corpora: Wikipedia and Twitter. Key features: - Social Tokenizer - Word Segmentation - Spell Correction - Customizable Word Segmentation and Spell Correction - Pre-Processing Pipeline Tech stack: Python Why it matters: Ekphrasis is useful for researchers and developers working on sentiment analysis, topic modeling, and other NLP tasks involving social network text data. Repository: cbaziotis/ekphrasis GitHub stars: 672 GitHub description: Ekphrasis is a text processing tool, geared towards text from social networks, such as Twitter or Facebook. It performs tokenization, word normalization, word segmentation (for splitting hashtags), and spell correction, using word statistics from 2 big corpora (English Wikipedia, Twitter - 330 million English tweets). _ekphrasis_ was developed as part of the text processing pipeline for DataStories team's submission for SemEval-2017 Task 4 (English), Sentiment Analysis in Twitter. If you use the library in your research project, please cite the paper ["DataStories at SemEval-2017 Task 4: Deep LSTM with Attention for Message-level and Topic-based Sentiment Analysis"](http://www.aclweb.org/anthology/S17-2126). Citation: ``` @InProceedings{baz

Overview: Second Brain is an agentic framework designed to act as an operating system, integrating local file intelligence, workflow automation, and language models to perform tasks and communicate across multiple modalities and messaging platforms. Key features: - Local file intelligence - Workflow automation - Language model integration Tech stack: Python Why it matters: Second Brain simplifies the integration of AI models and workflows, making it easier for developers to leverage language models and automate tasks across various platforms. Repository: henrydaum/second-brain GitHub stars: 672 GitHub description: Second Brain is an agentic framework that acts as an operating system, using local file intelligence, workflow automation, and LLMs to complete tasks and communicate over multiple modalities and messaging platforms.

### Overview: DynamicViT is a dynamic token sparsification framework that reduces FLOPs by 30% and improves throughput by 40% while maintaining accuracy within 0.5%. ### Key Features: - **Dynamic Token Sparsification**: Prunes redundant tokens in vision transformers progressively and dynamically. - **Efficient Vision Transformers**: Reduces FLOPs by 30% and improves throughput by 40%. - **Dynamic CNNs and Swin Models**: Pretrained on ImageNet with improved dynamic spatial sparsification. ### Tech Stack: PyTorch ### Why It Matters: DynamicViT offers a significant performance boost without compromising accuracy, making it a valuable addition to the toolkit for vision transformer research and applications.

Overview: This repository is a curated list of open-source AI agent platforms, covering AI coworkers and teammates, agent builders and frameworks, workflow automation platforms, browser agents, and coding agents. It includes platforms and frameworks that developers can use to build and automate AI agents. Key features: - AI coworkers and teammates - Agent builders and frameworks - Workflow automation platforms - Browser agents - Coding agents Tech stack: Astro Why it matters: This repository provides a comprehensive overview of open-source AI agent platforms, helping developers find the right tools and frameworks for their projects. Repository: Agenta-AI/awesome-ai-agent-platforms GitHub stars: 669 GitHub description: A curated list of open-source AI agent platforms: AI coworkers and teammates, agent builders and frameworks, workflow automation platforms, browser agents, and coding agents, with licenses.

### Overview: Cheetah is a compact, computationally efficient on-device speech-to-text engine developed by Picovoice. It is designed to run locally on various platforms, including Linux, macOS, Windows, Android, iOS, Chrome, Safari, Firefox, Edge, Raspberry Pi, and more. Cheetah is private, ensuring all voice processing occurs locally, and it is accurate, cross-platform, and efficient. ### Key Features: - **Private Processing**: All voice processing is done locally, ensuring privacy. - **High Accuracy**: Achieves high accuracy in speech recognition. - **Compact and Efficient**: Designed for low-power devices, it is computationally efficient. - **Cross-Platform**: Compatible with various operating systems and devices. - **Demos**: Available in Python, C, iOS, Android, Flutter, React Native, Node.js, Java, and Web (both vanilla JavaScript/HTML and React). ### Tech Stack: - **Python**: Python SDK for Cheetah. - **C**: C SDK for Cheetah. - **iOS**: iOS SDK for Cheetah. - **Android**: Android SDK for Cheetah. - **Flutter**: Flutter SDK for Cheetah. - **React Native**: React Native SDK for Cheetah. - **Node.js**: Node.js SDK for Cheetah. - **Java**: Java SDK for Cheetah. - **.Net**: .Net SDK for Cheetah. - **Web**: Web SDK for Cheetah (both vanilla JavaScript/HTML and React). ### Why It Matters: Cheetah