Projects

### Overview: This project provides a working stack and crash-safe configurations for serving Poolside Laguna S 2.1 (NVFP4) on the NVIDIA DGX Spark (GB10). The repository includes detailed documentation, benchmarks, and troubleshooting guides to ensure smooth operation without hanging the box. ### Key Features: - **128K + DFlash Configuration:** Serves at a typical speed of about 25 to 30 tokens per second, with sustained performance up to 45 tokens on code. - **256K, No DFlash Configuration:** Provides a crash-safe configuration for deep-context work, able to handle up to 153K prefill. - **256K + DFlash Configuration:** Not viable on the GB10, leading to a hard hang on the GPU. ### Tech Stack: The repository uses a combination of Shell scripts and Docker images to ensure a smooth and crash-safe operation. ### Why It Matters: This project is essential for developers looking to deploy and run large language models on the DGX Spark without encountering common issues like hanging boxes or silent crashes. The detailed documentation and step-by-step guides ensure that users can quickly set up and troubleshoot their environments effectively. ### GitHub Stars: 19 Stars ### Repository: The repository is hosted on GitHub at [sudoingX/dgx-spark-laguna](https://github.com/sudoingX/dgx-spark-laguna).

Overview: Sutro is a tool designed to optimize and run consistent AI functions reliably, especially for teams needing grounded LLM judges, classifiers, and extractors. Key features: - Sutro Functions: Task-specific judges, classifiers, and extractors aligned to decision preferences. - Sutro Batch: Serverless async inference for high-volume AI workloads. Tech stack: Python. Why it matters: Sutro simplifies the process of building and deploying AI functions, ensuring reliable and consistent performance, making it ideal for teams working with unstructured data. Repository: sutro-sh/sutro Primary language: Python GitHub stars: 19 GitHub description: Optimize and run consistent AI Functions.

Overview: Vectorify is an AI connector for Laravel that allows you to query and explore your data in natural language, turning your Laravel app into a conversational knowledge base. Key features: - Vectorify package for Laravel - Chats page within our platform - Embed Chatbot into Laravel app - Add MCP server to ChatGPT, Claude, etc. - Call REST API endpoints Tech stack: PHP Why it matters: Vectorify simplifies AI-powered data exploration in Laravel, turning complex data querying into a seamless and intuitive feature.

### Overview: OneEmo is an advanced multimodal reasoning model designed for emotion perception, understanding, and interaction. It achieves state-of-the-art results on eight emotion tasks and is trained on the EmoWorld-130K dataset. ### Key Features: - Sentiment Analysis (MSA) - Basic Emotion Recognition (B-MER) - Open-Vocabulary Emotion Recognition (OV-MER) - Intent Recognition (MIR) - Humor & Sarcasm Understanding (MHU & MSU) - Empathic Response Generation (ERG) - Emotional Support Conversation (ESC) ### Tech Stack: Python ### Why It Matters: OneEmo is a unified model that addresses the complexities of emotion perception and understanding across various tasks, providing a comprehensive solution for emotion-related applications. ### Repository: Repository: waHAHJIAHAO/OneEmo GitHub Stars: 19 ### Updates: - **[Aug 2026]** 🥰 Online demo available at [HF Space.](https://huggingface.co/spaces/Jiaha0Hu4ng/OneEmo) - **[Aug 2026]** 🚀 Released datasets and evaluation scripts. - **[Aug 2026]** 🎉 Released Paper. - **[Jul 2026]** 🔥 Released Post-Training weights for OneEmo. - **[Jul 2026]** 💻 Inference code live! ### 💡 About OneEmo: OneEmo targets video emotion intelligence

### Overview: This project presents the official implementation of the Multi-Modal Fact Verification dataset from the De-Factify workshop at AAAI-22, showcasing the winning team's approach and results. ### Key features: - Pre-CoFactv2, an improved version of the dataset, won the first prize at the FACTIFY 2 Challenge. - The project includes a multi-modal fact verification dataset with 35,000 training examples and detailed metrics for evaluation. - The README provides comprehensive instructions for training, evaluating, and ensemble models. ### Tech stack: Python ### Why it matters: This project offers a valuable dataset and implementation for researchers and developers interested in improving fact verification techniques using multimodal data.

Overview: This project is a PyTorch implementation of the Generative Adversarial Text-to-Image Synthesis paper, utilizing a Conditional Generative Adversarial Network (CGAN) to generate images from text descriptions. Key features: - Text-to-image synthesis using a CGAN architecture - Inspired by DCGAN with text conditioning - Utilizes Caltech-UCSD Birds 200 dataset - Implements requirements for PyTorch 2.0 Tech stack: PyTorch Why it matters: This project demonstrates the application of CGANs in generating images from textual descriptions, showcasing the potential of deep learning in generating realistic images from natural language inputs. Repository: yashashwita20/text-to-image-using-GAN Primary language: Jupyter Notebook GitHub stars: 19

### Overview: This repository serves as a comprehensive guide for researchers and developers interested in Edge/Mobile Large Language Models (LLMs). It includes a wide range of papers, codebases, toolchains, and open-source frameworks, covering various aspects of the intersection between Edge AI and LLMs. ### Key Features: - Comprehensive list of papers, codebases, toolchains, and open-source frameworks. - Focus on Edge/Mobile LLMs, including frameworks, benchmarks, commercial applications, and models. - Contains detailed information on deployment frameworks, commercial products, models, and performance benchmarks. ### Tech Stack: The repository utilizes Markdown, GitHub, and other open-source technologies to ensure accessibility and maintainability. ### Why it Matters: This repository provides a valuable resource for anyone looking to explore the latest advancements in Edge/Mobile LLMs, helping researchers and developers stay updated with the most relevant and cutting-edge work in the field. ### Repository: The repository is hosted on GitHub and is currently at version 1.0. It has received 19 GitHub stars, indicating a growing interest in the content. ### Tags Convention: - **🔗 Hyper links**: Links to papers, official websites, GitHub repositories, and Huggingface models. - **💡 Highlights**: Short names, surveys, and badges to provide quick access to key information. ### Contents: - **🔨 Deployment Frameworks**: Includes open-source frameworks and benchmarks, performance benchmarks, and model export formats. - **📱 Commercial Products and Applications**: Covers commercial cases, accelerator and AI chip applications, and models. - **

### Overview: Visual Agentic Memory (VAM) is an innovative system designed for sparse, distant, and difficult-to-localize evidence scenarios. It transforms unconstrained video streams into hierarchical visual memory, enabling query-driven agentic retrieval and grounding answers through visual inspection. ### Key Features: - **Online Indexing with Adaptive Deduplication**: Identifies boundary-sensitive moments using an adaptive thresholding mechanism. - **Hierarchical Memory with Parallel Representation**: Organizes moments and events across temporal tiers for scalable storage and visual detail preservation. - **Agentic Retrieval for Grounded Reasoning**: Uses an MLLM as a retrieval agent to search, inspect, and summarize evidence for auditable reasoning over distant events. - **Training-free End-to-End Framework**: Supports streaming perception and long-horizon retrieval using pre-trained multimodal components without task-specific retraining. ### Tech Stack: Python ### Why it Matters: VAM is tailored for settings where relevant evidence is sparse, temporally distant, and challenging to localize with single-pass video reasoning. By leveraging online indexing, hierarchical memory, and agentic retrieval, VAM efficiently retrieves the most relevant evidence for downstream reasoning and verification, ensuring accountability and accuracy. ### Repository: yiliu-li/Visual-Agentic-Memory ### GitHub Stars: 18 ### Primary Language: Python ### GitHub Description: Official Implementation of "Visual Agentic Memory: Online Long Video Understanding via Online Indexing, Hierarchical Memory, and Agentic Retrieval" ### Motivation: VAM is designed to address scenarios where evidence is sparse, tempor

### Overview: Bilibili Video Research is an evidence-aware tool designed to analyze Bilibili videos, extracting insights from captions, video frames, and community context. It provides a traceable research report that separates public metadata, captions, and untrusted community opinions, ensuring a deterministic `VIDEO CONTEXT` block for each result. ### Key Features: - Extracts insights from Bilibili video captions, video frames, and community context. - Supports three modes: `language`, `vision`, and `multimodal`. - Determines the best mode based on the evidence needed for the question. - Includes a `VIDEO CONTEXT` block with public metadata and community context status. - Compatible with Codex, OpenCode, and other MCP-compatible clients. ### Tech Stack: TypeScript ### Why It Matters: This tool is essential for researchers and developers looking to extract meaningful insights from Bilibili videos, ensuring that the analysis is both accurate and contextually relevant.

### Overview: CardioLab is a machine and deep learning framework designed for the estimation and monitoring of laboratory abnormalities through Electrocardiogram (ECG) data. ### Key features: - **Estimation Task**: Predicts laboratory values from ECG features. - **Monitoring Task**: Detects abnormalities in laboratory values over a future time horizon. ### Tech stack: Python ### Why it matters: CardioLab offers a comprehensive solution for the analysis and monitoring of laboratory values from ECG data, addressing both estimation and monitoring tasks. This framework is particularly valuable in clinical settings where early detection of abnormalities can significantly improve patient outcomes.

Overview: This project outlines and demonstrates how language models can understand image, video, and text content. Key features: - Image understanding - Video understanding - Audio understanding Tech stack: Jupyter Notebook Why it matters: This notebook provides a high-level overview of techniques used to teach language models image, video, and audio understanding, showcasing open-source model examples.

### Overview: DAR Library is an open-source, fast, modular, and diversity-aware library for Multi-Agent Debate research. It provides a comprehensive collection of state-of-the-art MAD baselines, including the DAR method, making it an essential tool for researchers and developers in the field. ### Key Features: - **Diversity-Aware Retention (DAR):** A method that retains only the most diverse and informative messages, significantly reducing message volume and improving performance. - **SOTA Baselines:** Includes uncertainty-aware prompting, voting mechanisms, uncertainty-based filtering, persona, society of minds, and topology variants (sparse, centralized), all in one place. - **Modular and Extensible:** Easy to add new models, benchmarks, or filtering strategies with minimal code changes. ### Tech Stack: Python 3.10 ### Why It Matters: - **100x Faster Inference:** Native vLLM integration allows for significantly faster inference compared to existing MAD libraries, enabling more experiments to be run and experiments to be completed more quickly. - **SOTA Baselines:** Provides access to the latest research in MAD, including the DAR method, making it easier to benchmark and build on the latest research in minutes. - **Modular and Extensible:** Easy to add new models, benchmarks, or filtering strategies with minimal code changes, making the library highly adaptable and expandable. ### GitHub Stars: 18 ### Project Structure: ```plaintext DAR/ ├── README.md ├── LICENSE ├── requirements.txt ├── assets/ │ └──

Overview: RisuAI Agent is a plugin that simulates an AI agent orchestration layer on top of RisuAI, allowing your main AI model to focus on delivering a stellar, immersive performance. Key features: - Information Extraction: Distills long, sprawling chat histories into high-density, easily digestible information blocks. - Bot Reorganization: Dismantles the character card and intelligently re-categorizes entries. - Broad Prompt Compatibility: Works alongside your favorite prompts. - Director's Notes: Actively guides the plot, manages emotional shifts, and corrects logical inconsistencies. - Vector Search: Replaces rigid keyword matching with intelligent semantic retrieval, increasing relevance and density. Tech stack: JavaScript Why it matters: This plugin optimizes your AI's performance by delegating heavy lifting to specialized background AI agents, allowing your main model to focus on delivering a stellar, immersive performance.

### Overview: AVI-Bench evaluates how well Omni-Multimodal Large Language Models (Omni-MLLMs) such as Gemini, GPT-4o, Qwen-Omni, and Baichuan-Omni handle joint audio-visual reasoning. It organises evaluation around the human cognitive process — *Perception → Understanding → Reasoning* — and adds the **Primitive Sensation (PriSe)** extension to test generalisation to low-semantic, unfamiliar audio-visual inputs. ### Key features: - **14-task / 5,864-sample benchmark dataset layout** reported in the paper. - **Unified inference + answer-formatting + evaluation pipeline**. - **Reference adapter for any OpenAI-compatible API (Gemini / GPT-4o / Claude / self-hosted)**. - **Four-level AVI taxonomy** for principled cross-stage comparison. ### Tech stack: Python ### Why it matters: This repository contains the 14-task / 5,864-sample benchmark dataset layout reported in the paper, a unified inference + answer-formatting + evaluation pipeline, a reference adapter for any OpenAI-compatible API (Gemini / GPT-4o / Claude / self-hosted), and a four-level AVI taxonomy for principled cross-stage comparison.

Overview: This repository showcases a deep learning-based solution for automated drowning detection using YOLO v8 architecture. It focuses on enhancing drowning incident response systems through meticulous fine-tuning and integrated methodologies. Key features: - Datasets curated from reputable sources, ensuring comprehensive coverage of drowning scenarios. - Fine-tuning of YOLO v8 architecture exclusively for drowning instances to improve accuracy and sensitivity. - Integration of secondary datasets to refine the model's understanding of varied drowning scenarios. - Utilization of Ultralytics library for efficient model loading and image analysis. - Incorporation of coordinate calculation method for precise object localization. - Integration with pool cameras and depth estimation sensors for real-time spatial analysis. - Sample detection images illustrating the system's capabilities. Tech stack: Python, Jupyter Notebook, Ultralytics. Why it matters: This solution revolutionizes drowning incident response systems by providing swift and accurate identification of distress situations in aquatic environments, enhancing the efficiency of response systems. Repository: Hasibwajid/Automated-Drowning-Detection-YOLOV8 GitHub stars: 18 GitHub description: Automated Drowning Detection: A repository showcasing a deep learning-based solution using YOLO v8 architecture for swift and accurate identification of drowning instances in aquatic environments. Enhanced accuracy through meticulous fine-tuning and integrated methodologies. Empowering drowning incident response systems for improved efficiency.

Overview: This project introduces Self-Control, a method for controlling the behavior of large language models (LLMs) by compressing suffix gradients into a prefix controller. The goal is to enhance the safety and ethical behavior of LLMs without requiring human annotations. Key features: - Suffix Gradient - Prefix Controller Tech stack: Jupyter Notebook Why it matters: This approach offers a novel way to manage LLMs' behaviors, ensuring their outputs are controlled and safe, while maintaining efficiency and ease of use. Repository: GitHub - https://github.com/HenryCai11/LLM-Control Primary language: Jupyter Notebook GitHub stars: 18

Overview: This is an independent, clean-room AGPL-3.0 reimplementation of the halogen 0.1.3 inference engine specifically tailored for Qwen3.8-27B on AMD Strix Halo (gfx1151). It adheres to the same wire protocol and OpenAI-compatible API as the original halogen, ensuring compatibility with existing systems. Key features: - Same wire protocol as halogen's engine (`PING/INFO/CSTAT/GEN/X` on the token port) - Loads `.hgn` checkpoints (see `converter/` and `docs/halogen/CHECKPOINT-FORMAT.md`) - Greedy token streams match the original engine on identical weights Tech stack: HIP Why it matters: This reimplementation is designed for developers who need a clean-room, open-source alternative to the closed-source halogen server, ensuring compatibility and interoperability without the need for reverse engineering.

Overview: HomoScriptor is a vibrant community-driven dataset designed for fine-tuning large language models, focusing on human-written tasks categorized into language and logic tasks. Key features: - Categorized JSON Files: Organized into language and logic tasks, making it easy to explore specific linguistic domains. - Short and Long Variant Outputs: Includes both short and long variant outputs for flexibility in use cases. - Open-Source and Collaborative: Embraces community contributions and welcomes contributors for enhancing the dataset. Tech stack: Unknown Why it matters: HomoScriptor provides a curated collection of human-written datasets for fine-tuning language models, fostering innovation and advancing the field of language modeling.

### Overview: This project provides a Delphi wrapper for the Stability AI API, enabling developers to leverage advanced AI models for image, video, and 3D generation, as well as sophisticated editing and upscaling tools. ### Key Features: - **Stable Image Ultra**: Generates high-quality images from text descriptions. - **Stable Image Core**: Offers image creation and style manipulation. - **Stable Diffusion**: Supports text-to-image generation and image-to-image with prompts and masks. - **Upscale**: Provides conservative and creative upscale options. - **Edit**: Includes features like erasing, inpainting, outpainting, and background removal. - **Control**: Features such as sketching, structuring, and styling options. - **Results**: Displays generated images and videos. - **3D**: Supports 3D generation and rendering. - **Video**: Enables video editing and manipulation. - **Other Features**: Includes model list, user account, and user balance functionalities. ### Tech Stack: Delphi ### Why it Matters: This project is essential for developers looking to integrate advanced AI capabilities into their Delphi applications, offering a seamless and efficient way to create, edit, and upscale images and videos, as well as generate 3D models and videos. ### Repository: MaxiDonkey/DelphiStabilityAI ### GitHub Stars: 18 ### GitHub Description: The Stability API wrapper for Delphi utilizes advanced models developed by Stability.ai to provide capabilities for seamless image, video, and 3D generation, as well as

### Overview: Openextract is a Python library designed to extract structured data from various types of documents, images, audio, and video using large language models (LLMs). It supports multiple providers, including Anthropic, Google, and OpenAI, and allows users to specify Pydantic schemas and models for accurate data extraction. ### Key Features: - **Pydantic Schemas**: Automatically validate and extract structured data from files or models. - **Multiple Providers**: Supports Anthropic, Google, and OpenAI models. - **Flexible Inputs**: Accepts paths, URLs, and raw file content. - **Decision Models**: For schemas with boolean, literal, and enum fields, uses decision models for better accuracy. - **Decision Provider Support**: Supports OpenAI Decisions and OpenRouter Decisions for more complex schemas. ### Tech Stack: - Python 3.12+ - Pydantic - OpenAI, Anthropic, Google models ### Why It Matters: - **Accuracy and Reliability**: Utilizes advanced LLMs to extract structured data accurately. - **Versatility**: Handles various types of media files and supports multiple providers. - **Ease of Use**: Simplifies data extraction through Pydantic schemas and decision models. ### Repository: - **GitHub Repository**: [Mellow-Artificial-Intelligence/openextract](https://github.com/Mellow-Artificial-Intelligence/openextract) - **GitHub Stars**: 18 - **Primary Language**: Python ### Installation: - **Install via Pip**: `pip install 'openextract[all]'`