Projects

### Overview: OhMyBuddy is a self-hosted web app that transforms the Pi coding agent into a personal AI assistant accessible from any device. It offers a dedicated Buddy with a personalized interface and a workspace for project work, all while keeping your data and settings safe across multiple platforms. ### Key Features: - **Personal Assistant Buddy:** A Buddy remembers your preferences, keeps track of your conversations, and handles daily tasks and reminders. - **Full Agent Workspace:** A workspace for project work, including threads, streaming tool calls, diffs, subagents, side chats, and session trees. - **Real Work Done:** Buddies can read and write files, run commands, use web tools, and start subagents and scheduled tasks on your machine. - **Model Management:** Use any provider supported by Pi, including Anthropic, OpenAI, Google Gemini, GitHub Copilot, OpenRouter, DeepSeek, Qwen, Kimi, MiniMax, xAI, and Mistral, with an API key, OAuth, or your own OpenAI-compatible endpoint. - **Machine Security:** Your models, keys, and machine remain on your server, ensuring data privacy and security. ### Tech Stack: JavaScript ### Why It Matters: - **Personalization:** Keeps track of your preferences and conversations, making interactions feel more natural. - **Real Work Done:** Offers a comprehensive workspace for project work, ensuring tasks are completed efficiently. - **Model Flexibility:** Supports various providers, allowing for a wide range of AI models. - **Machine Security:** Ensures your models, keys, and machine remain

### Overview: This repository contains the official implementation of a paper titled "Latent Diffusion Prior Enhanced Deep Unfolding for Snapshot Spectral Compressive Imaging" at the 2024 European Conference on Computer Vision (ECCV). The project focuses on improving the reconstruction quality and computational efficiency of snapshot spectral compressive imaging using a generative model called the latent diffusion model (LDM) and a Trident Transformer (TT) for deep unfolding. ### Key features: - **Latent Diffusion Model (LDM)**: Generates degradation-free prior to enhance the regression-based deep unfolding method. - **Trident Transformer (TT)**: Extracts correlations among prior knowledge, spatial, and spectral features to integrate knowledge priors in the deep unfolding denoiser. - **Generative Model Integration**: Combines physics-driven deep unfolding with generative LDM to improve reconstruction quality and computational efficiency. ### Tech stack: Python ### Why it matters: This approach is significant because it introduces a novel method to integrate generative models with traditional deep unfolding techniques, leading to improved reconstruction quality and efficiency in snapshot spectral compressive imaging.

### Overview: This repository contains tested image prompt recipes for the FLUX.1 Kontext Pro model, including example renders, aspect ratios, model ID (`flux-kontext-pro`), and cost per image. The recipes are designed to be effective and efficient, with specific rules for naming materials, setting lighting, framing, and choosing the right ratio and lens. ### Key features: - **Prompt Recipes**: 8 tested recipes with example renders, aspect ratios, and the model ID (`flux-kontext-pro`). - **Cost Estimation**: Estimated cost of the whole set: $0.256. - **Model Compatibility**: Compatible with OpenAI's API endpoint `https://api.apimart.ai/v1`. - **Pricing**: $0.032 per image, with flexible pricing and no subscription required. - **Asynchronous Nature**: Tasks are batched and retries are common, not a custom integration. ### Why it matters: - **Cost Efficiency**: The recipes are designed to be effective and efficient, ensuring that the cost per image is minimized. - **Material Focus**: The recipes emphasize the importance of naming materials rather than adjectives, setting the light explicitly, and pinning the framing. - **Lighting and Composition**: The recipes provide clear instructions on setting the light, framing, and choosing the right ratio and lens, ensuring the output is tailored to the desired effect. ### Tech stack: Python ### GitHub stars: 37

### Overview: Veo 3.1 Lite API (`veo-3.1-lite`) provides model IDs, pricing details, and parameters for the OpenAI-compatible APIMart endpoint. It offers a cost-effective alternative to higher-tier Veo models, with a flat rate of $0.064 per second at 720P resolution, suitable for both text-to-video and image-to-video conversions. ### Key Features: - **Model ID and Pricing**: Provides model IDs and pricing details for the OpenAI-compatible APIMart endpoint. - **Parameter Options**: Supports parameters such as `model`, `prompt`, `duration`, `resolution`, `size`, and `image_urls`. - **Pricing Model**: Offers a per-generated-second pricing model, with no subscription required and no free quota. ### Tech Stack: - **Python**: The primary programming language used for developing and implementing the API. ### Why It Matters: - **Cost Efficiency**: Provides a lower-cost alternative to higher-tier Veo models, making it suitable for teams on a budget. - **Flexible Pricing**: Offers a flat rate of $0.064 per second, with no subscription required, making it easy to scale up or down as needed. - **Real-Time Pricing**: Reports the exact amount billed per call, ensuring accurate cost tracking. ### Repository: - **Repository**: `abkktyaqnmkmi/veo-3-1-lite-api` - **GitHub Stars**: 35 - **GitHub Description**: Veo 3.1 Lite API (`veo-

### Overview: Multimodal Language Models (MLLMs) have shown significant potential in handling complex tasks involving both text and images. However, their effectiveness is often limited by the knowledge they acquire during training, which restricts their practical utility. This project introduces Reflective LLaVA (ReflectiVA), a novel method that enhances the adaptability of MLLMs by integrating external knowledge sources. By using reflective tokens, ReflectiVA dynamically determines when external knowledge is needed and predicts the relevance of information retrieved from an external database, enabling MLLMs to manage external knowledge while maintaining fluency and performance on tasks where external knowledge is not required. ### Key Features: - **Reflective Tokens**: Dynamically determine the need for external knowledge and predict the relevance of information. - **External Knowledge Retrieval**: Predictively retrieves and utilizes external knowledge from databases. - **Fluency Preservation**: Maintains the fluency of the output while managing external knowledge effectively. ### Tech Stack: - Python ### Why It Matters: This project addresses the limitations of current MLLMs by enhancing their adaptability and practical utility. By integrating external knowledge sources, ReflectiVA enables MLLMs to perform tasks that require external knowledge without compromising their performance on tasks where external knowledge is not needed. This makes ReflectiVA a valuable tool for applications that benefit from the integration of both text and image data. ### Repository: - **Repository**: aimagelab/ReflectiVA - **GitHub Stars**: 57 - **Primary Language**: Python - **GitHub Description**: [CVPR

### Overview: This repository provides a complete pipeline for the BraTS 2020 multimodal brain tumor segmentation challenge, utilizing a 3D U-net model trained on the BraTS 2020 dataset. The repository includes a ready-to-use pipeline for segmentation tasks and is also extensible for other applications such as survival and uncertainty tasks. It supports the BraTS 2020 dataset and can be used with previous versions as well. ### Key features: - **Model Training**: A 3D U-net model is trained using the BraTS 2020 dataset. - **Data Preparation**: The dataset is split into training, validation, and test sets. - **Model Saving**: The best model is saved to `isensee_2017_model.h5`. - **Prediction**: Prediction files are generated for validation and test sets. ### Tech stack: - Python 3.5 or higher - TensorFlow GPU version 1.15 - Required packages: pytables, SimpleITK, nilearn, nibabel - Optional package: nipype for preprocessing (n4itk bias correction) ### Why it matters: This repository is a valuable resource for researchers and practitioners working on brain tumor segmentation tasks, providing a comprehensive solution for training, validating, and testing models on the BraTS 2020 dataset. ### GitHub stars: 57 ### Repository: - Repository: akhanss/BraTS-2020 - Primary language: Python - GitHub stars: 57

Overview: This repository is a comprehensive survey of reasoning and agentic systems in time series, focusing on large language models. It categorizes research papers and non-research contributions along two levels: reasoning topology and primary objective. Key features: - Research papers for direct reasoning - Research papers for linear chain reasoning - Research papers for branch-structured reasoning - Non-research papers Tech stack: Python, Markdown, and GitHub Pages. Why it matters: This survey provides a structured overview of the field, helping researchers and practitioners understand the current state and future directions in time series reasoning and agentic systems using large language models.

### Overview: ARFlow is a data-sharing layer designed to democratize and accelerate Augmented Reality (AR) research and development. It provides a tool for developers to easily share, store, and visualize AR device data, making it accessible and user-friendly. ### Key features: - **Data Sharing Layer**: Enables seamless data exchange between AR devices. - **Server Installation**: Simple installation via `pip` for both Python and C#. - **Client Setup**: Prebuilt Android and iOS apps available for easy deployment. - **Developer-Friendly Interface**: Streamlined setup process for both server and client. - **Documentation and Examples**: Comprehensive guides and examples available for reference. ### Tech stack: C# and Python ### Why it matters: ARFlow aims to make AR research and development more accessible and efficient by providing a user-friendly data-sharing platform. It streamlines the process of data sharing, storage, and visualization, allowing researchers and developers to focus on their core work rather than the technicalities of setting up and managing AR systems.

Overview: This project allows developers to bypass the middleman in ComfyUI, enabling direct API calls without any account required. It removes the proxy, allowing users to see the true cost of their API calls and bypasses vendor terms of service. Key features: - Direct API routing - No account needed - Local asset storage - 30+ providers supported - Simple UI configuration - Coverage inspector - Graceful fallback Tech stack: Python Why it matters: This project addresses the issue of API laundering, where third-party APIs are wrapped behind a proxy that manipulates pricing and authentication, leading to price opacity, vendor terms of service violations, and data routing issues. It ensures users can compare their costs directly to the actual vendor rates, bypassing the middleman's arbitrary pricing and regulatory risks.

### Overview: ImgX Studio is a self-hosted GPT Image WebUI designed for teams requiring product shots, campaign visuals, AI creative exploration, and multilingual image workflows. It offers a modern, customizable interface and supports multiple languages for seamless localization. ### Key Features: - **GPT Image / OpenAI-compatible API Integration**: Supports GPT Image 2, 2026-04-21, and GPT-1 models. - **Reference-Aware Generation**: Upload up to 4 reference images, each up to 10MB in size. - **Creative Iteration Board**: Select a generated image, set it as the active source, and generate the next branch. - **Remix Recipes**: Quickly apply variation, commercial polish, hero upscale, or local redraw instructions. - **Output Control**: Choose from smart, square, landscape, or portrait formats with customizable dimensions. - **Production Formats**: Export images in PNG, JPEG, or WEBP formats with auto, opaque, or transparent backgrounds. - **Privacy-Aware Key Handling**: Store connection settings locally or securely on the server. - **Localized Product Experience**: Supports 9 languages, including language switchers, locale-aware metadata, and error messages. ### Tech Stack: TypeScript ### Why It Matters: ImgX Studio provides a polished, self-hosted GPT Image workspace that streamlines the creative process from idea to production-ready visuals, making it the fastest path from API access to a usable product.

### Overview: VPGen is a novel framework for text-to-image generation that leverages visual programming to interpret and generate images step-by-step. This project, detailed in the NeurIPS 2023 paper, aims to make text-to-image generation more accessible and interpretable. ### Key features: - **Interpretable Step-by-Step Generation**: VPGen breaks down the text-to-image generation process into three distinct steps: object/count generation, layout generation, and image generation. - **LM-based Object/Count Generation**: Uses a language model (LM) to generate objects and counts. - **Layout-to-Image Module**: Utilizes a layout-to-image module to transform predicted layouts into images. - **Fine-tuned on Public Datasets**: Finetuned Vicuna 13B on text-layout pair annotations on datasets like Flickr30K, MS COCO, and PaintSkills. ### Tech stack: - **Jupyter Notebook**: The primary language used for development and experimentation. ### Why it matters: VPGen provides a more interpretable and step-by-step approach to text-to-image generation, making it easier to understand and debug the generation process. This framework could significantly enhance the accessibility and usability of text-to-image generation tools, particularly in fields like art, design, and education.

Overview: MIRAGE is an official repository for the paper "MIRAGE: A multimodal foundation model and benchmark for comprehensive retinal OCT image analysis," published in npj Digital Medicine. It is designed to perform a wide range of tasks such as disease staging, diagnosis, and layer and lesion segmentation on retinal OCT and SLO images. Key features: - MIRAGE is a multimodal foundation model. - It is based on the MultiMAE architecture and is pretrained using a multi-task learning strategy. - It is available in two sizes: MIRAGE-Base and MIRAGE-Large. - It is evaluated on a comprehensive benchmark consisting of 19 tasks from 14 publicly available datasets and two private datasets. Tech stack: Python Why it matters: MIRAGE significantly outperforms state-of-the-art foundation models across all task types, making it a valuable resource for researchers and developers in the field of retinal OCT image analysis.

### Overview: EmbeddingGemma2 is a powerful tool for local multimodal search, combining text, code, images, video, and audio into a shared embedding space. This list covers resources and applications supporting the second generation of EmbeddingGemma2. ### Key features: - **Popular Applications**: DigUp, Google AI Edge Gallery, MemPalace. - **Community Applications**: GitHub repositories and datasets. - **Fine-tuning and Decision Models**: Resources for fine-tuning and decision-making models. - **Decision Application Examples**: Practical examples of how EmbeddingGemma2 is used in decision-making processes. - **Start Here**: A guide to getting started with EmbeddingGemma2. - **Model at a Glance**: A brief overview of the model architecture and its capabilities. - **Guides and Examples**: Detailed guides and examples for various use cases. - **Models and Runtimes**: Information on available models and runtimes. - **Inference Frameworks and Servers**: Resources for deploying and running the models. - **Practical Notes**: Tips and best practices for using EmbeddingGemma2 effectively. - **Evaluation**: Metrics and evaluations of the model's performance. ### Tech Stack: Python ### Why it Matters: EmbeddingGemma2 provides a unified approach to multimodal search, making it easier to find and retrieve information across different types of data. This tool is particularly useful in applications where integrating different sources of data is crucial, such as local file search, decision-making models, and fine-tuning existing models.

Overview: Lightning Rod Python SDK is a powerful tool for obtaining calibrated probabilities for future predictions using an OpenAI-compatible API. It is trusted by high-stakes organizations and processes billions of tokens daily. Key features: - Predicts future events with calibrated probabilities - Supports OpenAI-compatible clients - Handles billions of tokens and 100k+ inference requests per day - Provides detailed documentation and an API key guide Tech stack: Python Why it matters: Lightning Rod Python SDK offers a reliable and efficient way to obtain future predictions, trusted by major financial institutions, and is designed to handle large volumes of data and frequent requests. Repository: lightning-rod-labs/lightningrod-python-sdk Primary language: Python GitHub stars: 57

Overview: MMFakeBench is an official dataset and code repository for evaluating the performance of Large Language Models (LLMs) in detecting misinformation, specifically focusing on mixed-source multimodal misinformation detection. This benchmark aims to assess the ability of models to identify and differentiate between manipulated text and images. Key features: - Contains annotated samples with mixed-source type, including textual veracity distortion and visual veracity distortion. - Provides detailed annotations for each sample, including text, image paths, sources, ground truth answers, and misinformation types. - Offers a structured dataset format that is easily accessible for researchers and developers to evaluate and improve their LLM models. Tech stack: Python Why it matters: MMFakeBench serves as a crucial benchmark for evaluating the capabilities of Large Language Models (LLMs) in detecting misinformation, particularly in mixed-source multimodal scenarios. By providing a standardized dataset and annotations, it helps researchers and developers to objectively assess and improve the performance of their models, thereby advancing the field of misinformation detection and verification in the context of LLMs.

Overview: VoiceCue is an AI-powered tool that analyzes voice recordings to identify sentiments, tags, entities, and actions, making it easier to process and understand audio data. Key features: - Voice recognition - General stats - Sentiment analysis - Word cloud generation - Entity name recognition - Activity tracking - Interactive transcript - Speaker detection - Cue word usage - Custom search - Waveform preview - Audio controls - Drag and drop support - Upload MP3 files - Progress loaders - Fully responsive design Tech stack: NextJS, Deepgram, sentiment, react-tagcloud, react-tabs, react-drag-drop-files, wavesurfer.js, GitHub, Vercel, ESLint, prettier, MIT license. Why it matters: VoiceCue simplifies the analysis and processing of voice recordings, making it a valuable tool for anyone looking to extract meaningful insights from audio data efficiently.

Overview: Aozora Forge is a low VRAM SDXL fine-tuning toolkit designed for 12GB GPUs, offering a memory-thrifty approach to training high-quality AI models. Key features: - Adaptive gradient checkpointing - Stratified parameter freezing - Quantized latent caching Tech stack: HTML Why it matters: This toolkit enables serious model adaptation without requiring datacenter-class hardware, making it ideal for developers and artists on modest GPUs.

Overview: This project demonstrates how to build a content-based article recommender using Pinecone's similarity search, showcasing the integration of TypeScript with a real-world example. Key features: - Content-based article recommendation - Utilizes Pinecone's similarity search for efficient retrieval - Implements a simple recommendation engine based on user behavior and article content Tech stack: TypeScript Why it matters: This project provides a practical example of how to leverage Pinecone's similarity search for content-based recommendation, demonstrating the effectiveness of content-based filtering in recommendation systems. Repository: pinecone-io/recommender-example-typescript GitHub stars: 57

Overview: VoskJs is a Node.js toolkit designed for developers to integrate Vosk's offline speech recognition engine, offering sentence-based and streaming-based transcript APIs, a command-line utility, and a demo HTTP transcript server. It simplifies the process of speech recognition in various applications, including single-user and multi-user scenarios. Key features: - Sentence-based transcript API - Streaming-based transcript API (draft) - Command line utility `voskjs` - Demo HTTP transcript server `voskjshttp` Tech stack: Node.js Why it matters: VoskJs simplifies the integration of Vosk's speech recognition engine into Node.js applications, making it easier for developers to implement speech recognition functionality in their projects, especially in scenarios requiring low latency and efficient processing.

Overview: This project is a solution to the TensorFlow Speech Recognition Challenge, achieving a rank of 63 on the private leaderboard. It includes a Jupyter Notebook-based approach using various models and techniques for speech recognition. Key features: - A variant of Convolutional LSTM - LSTM-L - C-RNN - GRU-L - Resnet Tech stack: Jupyter Notebook Why it matters: This project demonstrates the effectiveness of ensemble models and advanced deep learning techniques in speech recognition, achieving a significant performance improvement over the competition. Repository: subho406/TF-Speech-Recognition-Challenge-Solution Primary language: Jupyter Notebook GitHub stars: 57 GitHub description: Source code of the model used in Tensorflow Speech Recognition Challenge (https://www.kaggle.com/c/tensorflow-speech-recognition-challenge). The solution ranked in top 5%.