Projects

Overview: Pipelex is a platform that enables developers to build AI methods using their coding agents, allowing them to deploy these methods in various applications such as chatbots, webapps, or APIs. Key features: - Build AI methods with your coding agent - Deploy methods as an MCP for chatbots - Deploy methods as a webapp for people - Deploy methods via API for software Tech stack: Python Why it matters: Pipelex simplifies the process of integrating AI capabilities into software projects, making it accessible for developers without extensive AI expertise.

Overview: This repository chronicles the timeline of key events related to ChatGPT, Generative AI, and Large Language Models (LLMs), focusing on the period before and after the ChatGPT announcement. Key features: - OpenAI's Reddit deal to train AI on user posts - Dissolution of OpenAI's long-term AI risks team - International Scientific Report on AI safety - TRANSIC: Sim-to-Real Policy Transfer - Toon3D: New perspective on cartoons - Testing the reliability of AI for ecological information extraction - Many-Shot In-Context Learning in multimodal foundation models Tech stack: GitHub Why it matters: This repository serves as a valuable resource for tracking the evolution of Generative AI and LLMs, highlighting key milestones and developments in the field. Repository: hollobit/GenAI_LLM_timeline GitHub stars: 949

### Overview: KG-RAG is a task-agnostic framework that combines the explicit knowledge of a Knowledge Graph (KG) with the implicit knowledge of a Large Language Model (LLM) to empower a general-purpose LLM by incorporating an optimized domain-specific 'prompt-aware context' from a biomedical KG. ### Key features: - Extracts "prompt-aware context" from a biomedical Knowledge Graph (KG) to enhance the response to user prompts. - Empowers a general-purpose LLM by incorporating optimized domain-specific context. - Utilizes a massive biomedical Knowledge Graph (SPOKE) for context extraction, covering over 40 biomedical knowledge repositories from diverse domains. ### Tech stack: Jupyter Notebook ### Why it matters: KG-RAG enables LLMs to better understand and respond to biomedical questions by incorporating explicit and implicit knowledge from a biomedical Knowledge Graph, making it a powerful tool for tasks requiring knowledge-intensive inputs. Repository: BaranziniLab/KG_RAG GitHub stars: 947 GitHub description: Empower Large Language Models (LLM) using Knowledge Graph based Retrieval-Augmented Generation (KG-RAG) for knowledge intensive tasks README excerpt: <p align="center"> <img src="https://github.com/BaranziniLab/KG_RAG/assets/42702311/0b2f5b42-761e-4d5b-8d6f-77c8b965f017" width="450"> </p> ## Table

Overview: MAC-VO is a metrics-aware covariance approach for learning-based stereo visual odometry, presented at ICRA 2025 as the best paper award finalist. It offers faster pose graph optimization and mixed-precision inference, achieving 2x speedup compared to previous versions. Key features: - Metrics-aware covariance for improved accuracy - Fast pose graph optimization - Mixed-precision inference for speed Tech stack: Python 3.10, CUDA 12.4 Why it matters: MAC-VO significantly enhances learning-based stereo visual odometry by providing efficient and accurate pose estimation, making it a valuable contribution to the field of robotics and computer vision. Repository: MAC-VO/MAC-VO GitHub stars: 947 GitHub description: [ICRA 2025 Best Paper] MAC-VO: Metrics-aware Covariance for Learning-based Stereo Visual Odometry

Overview: This project, `calflops`, is designed to calculate FLOPs, MACs, and parameters for various neural networks, including custom models based on PyTorch. It supports models like Linear, CNN, RNN, GCN, and even large language models like BERT and LLaMA. Key features: - Calculates FLOPs, MACs, and parameters for all neural networks - Supports custom models via `torch.nn.functional.*` - Provides detailed breakdown of each submodule's performance - Supports Hugging Face platform models for inference - Automatically builds model input data for inference in meta device Tech stack: Python Why it matters: This tool simplifies the process of calculating FLOPs and parameters for neural networks, making it easier to understand the performance of each part of the model. It is particularly useful for large language models and models on the Hugging Face platform, where it can be used to compute FLOPs without downloading the entire model weights locally. Repository: MrYxJ/calculate-flops.pytorch Primary language: Python GitHub stars: 947 GitHub description: The calflops is designed to calculate FLOPs, MACs, and Parameters in all various neural networks, such as Linear, CNN, RNN, GCN, Transformer(Bert, LLaMA etc Large Language Model), even including any custom models via torch.nn.functional.* as long as based on the Pytorch implementation. Meanwhile this tool supports the printing of FLOPS, Parameter calculation value and proportion of each submodule of the model, it

### Overview: Gibson Environments is a virtual environment simulator designed to enhance the learning of perception and active behavior in embodied agents. This project aims to bridge the gap between real-world perception and robotic learning by providing a realistic simulation environment that reflects the complexity of real-world scenarios. ### Key Features: - Gibson is a realistic virtual environment simulator that mirrors the semantic complexity of real-world spaces. - It includes a baked-in mechanism for transferring to real-world scenarios, known as Goggles. - The agent is embodied and constrained by physics and space, allowing for practical learning and testing. ### Tech Stack: - C programming language ### Why It Matters: - Gibson provides a realistic and practical environment for learning perception and active behavior in embodied agents. - It addresses the challenges of transferring learning from simulations to the real world, making it a valuable tool for researchers and developers. ### Repository: Repository: StanfordVL/GibsonEnv GitHub Stars: 947 GitHub Description: Gibson Environments: Real-World Perception for Embodied Agents GitHub URL: https://github.com/StanfordVL/GibsonEnv ### Paper: ["Gibson Env: Real-World Perception for Embodied Agents"](https://arxiv.org/abs/1808.10654), in CVPR 2018 [Spotlight Oral]. ### Release: - Version: 0.3.1 - Release Date: [Insert Release Date Here] ### Database: - Includes 572 spaces and 1440 floors - Downloadable:

### Overview: This repository is a curated list of hosting providers sorted by minimal plan price, showcasing the best options for minimal investment. ### Key Features: - **Web Services Platform**: A comprehensive list of web services platforms. - **Backend-as-a-Service (BAAS)**: Providers offering scalable backend solutions. - **Containers**: Options for container-based hosting. - **Lambda**: Functions-as-a-service providers. - **Node.js**: Platforms for running Node.js applications. - **Static**: Solutions for static website hosting. - **Wordpress**: Hosting for WordPress websites. - **VPS**: Virtual Private Servers for dedicated hosting. - **Database**: Options for database hosting. - **GPU Cloud**: Providers offering GPU-based computing. - **AI LLM/Inference**: AI-powered hosting for AI coding and inference. - **Proxy services**: Services for enhancing application performance. - **Tokens**: Tokens for API access. - **Gateways**: Gateways for secure communication. - **Intelligent Model Routers**: Routers for intelligent model routing. - **Fusion Model Routers**: Fusion model routers for enhanced model management. - **Aggregators**: Aggregators for hosting multiple services. - **Resources**: Additional resources and links for further exploration. ### Tech Stack: Node.js The repository uses Node.js to manage and display the hosted information. ### Why it Matters: This list is essential for developers looking to start a project with minimal investment. It provides a straightforward comparison of hosting options, making it easier to choose the best fit for their needs. ### GitHub Stars: 94

Overview: Theia is an open-source library designed for multiview geometry and structure from motion, offering efficient, scalable, and accurate processing. Key features: - Multiview geometry algorithms - Structure from motion techniques - Modular pipeline design for easy extension - Efficient and scalable performance Tech stack: C++ Why it matters: Theia is crucial for anyone working on computer vision applications involving multiple images, providing a comprehensive solution for structure from motion and multiview geometry. Repository: sweeneychris/TheiaSfM GitHub stars: 946 GitHub description: An open source library for multiview geometry and structure from motion

Overview: Chat-UniVi is a unified visual representation model that enhances large language models by integrating image and video understanding capabilities. Key features: - Unified visual representation model - Image and video understanding - Empowers large language models Tech stack: Python Why it matters: Chat-UniVi leverages advanced visual understanding to significantly improve the performance of large language models, making them more versatile and effective in various applications.

Overview: DonSeTch is a web research tool designed for AI agents, built entirely in Rust with no keys or accounts required. It fetches, searches, and crawls the web, providing a unified experience across multiple MCP clients and as a standalone CLI. Key features: - Real Chrome TLS - Temporal stealth - Solve-and-bounce - Self-improving fetch - Keyless search - Pixel-fusion PDF Tech stack: Rust Why it matters: DonSeTch offers a fast, stealthy, and keyless web research experience, making it ideal for AI agents and developers looking to integrate web research into their applications without the need for external dependencies or keys. Repository: dondai44423/donsetch GitHub stars: 944 Primary language: Rust GitHub description: Web fetch, search, and crawl for AI agents. Built from scratch in Rust. No keys, no accounts. AGPL v3.

### Overview: Pix2Seq is an open-source codebase designed for multi-task generative modeling, specifically focusing on object detection. It leverages efficient TPUs/GPUs support and integrates both autoregressive and diffusion models for versatility. ### Key Features: - **Pix2Seq Object Detection**: Supports object detection using pre-trained checkpoints from Object365 and fine-tuned checkpoints from COCO. - **FitTransformer (FIT)**: Official implementations of FitTransformer as an encoder, diffusion decoder, and autoregressive decoder. - **Diffusion Models**: Official implementations of diffusion models like Bit Diffusion, RIN, built on top of the original Pix2Seq codebase. ### Tech Stack: Tensorflow 2 ### Why It Matters: Pix2Seq aims to be a general framework that can convert RGB pixels into semantically meaningful sequences, making it versatile for various tasks and their combinations.

### Overview: Transcription Stream is a turnkey, self-hosted offline transcription and diarization service. It includes drag-and-drop diarization and transcription via SSH, a web interface for file upload and download, and summarization with Ollama and Mistral. The service is designed for simplicity and efficiency, making it easy to integrate into workflows. ### Key Features: - **Drag and Drop Diarization and Transcription:** Process files directly via SSH. - **Web Interface:** Upload, review, and download files. - **Summarization with Ollama and Mistral:** Enhances transcription with insightful summaries. - **Meilisearch:** Provides fast full text search for transcriptions. ### Tech Stack: - Python ### Why It Matters: Transcription Stream simplifies the process of offline transcription and diarization, making it accessible and efficient. The integration with Ollama and Meilisearch adds powerful tools for summarization and full text search, respectively, enhancing the overall utility of the service. ### Repository: - **Repository:** transcriptionstream/transcriptionstream - **GitHub Stars:** 946 - **GitHub Description:** A turnkey self-hosted offline transcription and diarization service with LLm summary ### Additional Information: - **Ports:** - **SSH:** 22222 - **HTTP:** 5006 - **Ollama:** 11434 - **Meilisearch:** 7700 - **SSH Server Access:

### Overview: ODISE is an official PyTorch implementation of the paper "Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models" presented at CVPR 2023. This project aims to provide open-vocabulary panoptic segmentation capabilities using diffusion models and discriminative models. ### Key features: - **Text-to-Image Diffusion Models**: Utilizes pre-trained diffusion models to generate high-quality images from text inputs. - **Open-Vocabulary Panoptic Segmentation**: Can segment any category in the wild, not just predefined categories. - **Efficient Inference**: Fast inference times through optimized models and hardware acceleration. ### Tech stack: Python ### Why it matters: ODISE offers a powerful tool for open-vocabulary panoptic segmentation, making it highly useful for applications requiring flexibility and adaptability to new categories. Its ability to segment any category in the wild is particularly valuable in scenarios where new objects or categories may appear in the environment. Repository: **NVlabs/ODISE** GitHub stars: **943** GitHub description: Official PyTorch implementation of ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models [CVPR 2023 Highlight]

### Overview: The most advanced, fully offline client-side AI suite on Android today. ### Key Features: - **Offline AI**: Access to AI capabilities without internet connectivity. - **Kotlin**: Developed using Kotlin for seamless integration and performance. - **LLama.cpp**: Utilizes advanced language models for text generation and more. - **LiteRT Engine**: Optimized for low-power devices, providing excellent offline performance. - **Stable-Diffusion**: Enhances image generation capabilities. - **Whisper**: Improves speech-to-speech translation. - **Snapdragon NPU**: Employs the latest in mobile NPU technology for enhanced performance. - **Google Tensor G5**: Supports advanced machine learning models for improved accuracy. - **MediaTek**: Provides robust support for various hardware configurations. - **Gemini Nano**: Offers a powerful ML Kit for device-based AI. - **RAG**: Enhances retrieval-based augmented generation for more context-aware responses. - **MCP Servers**: Supports multi-device communication for better performance. - **Document Analysis**: Facilitates efficient document analysis and processing. - **Super-Resolution**: Enhances image quality for better visual representation. - **Voice Mode**: Enables speech-to-speech translation for better communication. - **SenseVoice**: Supports multilingual speech-to-speech translation. - **Supertonic**: Provides on-device text-to-speech for better voice output. - **SQLCipher**: Ensures secure data storage with AES-256 encryption. - **Biometric**: Enhances security with biometric authentication. - **Offline**: Designed for hard

### Overview: herdr-reviewr is a terminal plugin designed to review and comment on changes made by a code agent in a Herdr workspace. It provides a convenient way to review code diffs, comment on specific lines, and send feedback back to the agent, enhancing collaboration and code quality. ### Key features: - **Unified and Split Views**: Displays the diff in a single column or side-by-side split view. - **Four Diff Scopes**: Supports uncommitted, branch, last turn, and commits. - **Line Comments**: Allows commenting on specific lines or ranges of code. - **Text Selection**: Enables copying and pasting code snippets directly into the editor. - **File Viewer**: View any file within the worktree. - **Search**: Search files by name or code by content, powered by fff. - **Find in File**: Step through every match in the file. - **PR View**: Read-only view of a branch's pull request. - **Markdown Review**: Renders Markdown files as plain text. - **Themes**: Offers 20 themes, including dark and light. ### Tech stack: Rust herdr-reviewr is written in Rust, leveraging its performance and stability. ### Why it matters: herdr-reviewr streamlines the review process by providing a dedicated pane for code diffs, comments, and feedback, enhancing collaboration and improving the quality of code reviews in Herdr environments.

Overview: Codexia is a lightweight agent workstation designed to streamline the development workflow with Codex, Claude, and Keke-Agent, featuring a task scheduler, Git worktree, and remote control capabilities. Key features: - Codex integration for advanced AI code generation - Claude for natural language understanding in code - Keke-Agent for remote code execution and management - Task scheduler for efficient task management - Git worktree for seamless code repository management - Remote control for remote development sessions Tech stack: TypeScript Why it matters: Codexia simplifies the development process by combining cutting-edge AI tools with a user-friendly interface, making it an essential tool for developers looking to enhance their productivity and efficiency.

Overview: LangCorn is a powerful API server designed to streamline the deployment of LangChain models and pipelines, making it easy to serve language processing applications with FastAPI. Key features: - Easy deployment of LangChain models and pipelines - Ready to use auth functionality - High-performance FastAPI framework for serving requests - Scalable and robust solution for language processing applications - Supports custom pipelines and processing - Well-documented RESTful API endpoints - Asynchronous processing for faster response times Tech stack: Python Why it matters: LangCorn simplifies the process of integrating and deploying LangChain models and pipelines, making it an essential tool for developers looking to enhance their language processing applications with the power of FastAPI.

### Overview: Espresso is a fast, end-to-end neural speech recognition toolkit built on PyTorch and fairseq. It supports distributed training across GPUs and computing nodes, and offers various decoding approaches, including look-ahead word-based language model fusion. Espresso is designed for various speech datasets, including WSJ, LibriSpeech, and Switchboard. ### Key Features: - **Modular Architecture**: Espresso is highly modular, allowing for easy integration and customization. - **Distributed Training**: Supports distributed training across multiple GPUs and nodes. - **Fast Decoder**: Implements a parallelized decoder for efficient decoding. - **State-of-the-Art Training Recipes**: Provides state-of-the-art training recipes for WSJ, LibriSpeech, and Switchboard datasets. - **Conformer Encoder**: Supports Conformer encoder with an option to enable it in the configuration file. - **On-the-Fly Feature Extraction**: Enables feature extraction from raw waveforms using PyTorch Audio. - **Transducer Model Training**: Includes a suite of Transducer model training and decoding code. - **CTC Model Training**: Supports CTC model training and decoding. - **On-the-Fly Feature Extraction**: Enables feature extraction from raw waveforms using PyTorch Audio. - **Transformer Recipes**: Supports E2E LF-MMI and Cross-Entropy training for hybrid ASR. - **SpecAugment Support**: Supports SpecAugment for improved model robustness. ### Tech Stack: - **Python**: Version 3.8 or higher. - **PyTorch**: Version 1.10.0 or

Overview: FDEOps is a set of skills designed to help forward-deployed engineers quickly and effectively turn customer problems into working systems. It combines design thinking, first-principles thinking, and systems thinking to ensure a clear understanding of the people involved, question assumptions, and examine how changes affect the whole system. Key features: - **Strategy**: Clarify the goal and choose an architecture. - **Architecture**: Design and implement the system architecture. - **Engineering**: Build and integrate the system, and verify the result. Tech stack: JavaScript. Why it matters: FDEOps helps forward-deployed engineers efficiently turn customer problems into working systems, ensuring clear communication and effective collaboration between teams. Repository: suboss87/FDEOps Primary language: JavaScript GitHub stars: 948 GitHub description: Forward deployed engineering skills for AI coding agents. Docs: docs/README.md

Overview: Gated-SCNN is a state-of-the-art semantic segmentation model designed for high-resolution urban street scene understanding. Key features: - Gated Shape CNNs for efficient semantic segmentation - Supports Python 3 and requires PyTorch, torchvision, and other libraries - Pretrained models available for download and evaluation on the Cityscapes dataset Tech stack: Python 3, PyTorch, torchvision, numpy, scipy, scikit-image, tensorboardX, tqdm, torch-encoding, OpenCV, PyYAML Why it matters: Gated-SCNN offers improved performance and efficiency compared to traditional methods, making it suitable for real-world applications in urban street scene analysis.