Projects

Overview: This repository presents "Efficient Remote Sensing with Harmonized Transfer Learning and Modality Alignment" (HarMA), a novel framework that addresses key challenges in remote sensing image-text retrieval. HarMA integrates Harmonized Transfer Learning and Modality Alignment to achieve state-of-the-art performance without relying on external data or tricks, making it a versatile and resource-efficient solution for remote sensing applications. Key features: - **Harmonized Transfer Learning and Modality Alignment (HarMA)**: A unified perspective on multimodal transfer learning to enhance task performance, modality alignment, and single-modality uniform alignment. - **Hierarchical Multimodal Adapter**: Inspired by human brain's information processing, this adapter integrates shared mini-adapters for fine-grained semantic alignment. - **Parameter-Efficient Fine-Tuning**: Reduces training overhead while achieving state-of-the-art performance on popular multimodal retrieval tasks. Tech stack: Python Why it matters: HarMA offers a versatile and resource-efficient solution for remote sensing applications, enhancing vision and language representations in remote sensing tasks. Repository: seekerhuang/HarMA Primary language: Python GitHub stars: 64 GitHub description: [ICLRW 2024] Efficient Remote Sensing with Harmonized Transfer Learning and Modality Alignment

### Overview: LLM Code (Open Claude Code) is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster through natural language commands. It integrates directly with your development environment, allowing you to edit files, navigate directories, and interact with an AI assistant. ### Key Features: - 🤖 Interactive AI coding assistant - 📁 File and directory operations - ✏️ File editing and appending - 🔍 Codebase context understanding - ⚙️ Configurable settings - 🌈 Colored terminal output ### Tech Stack: Python ### Why it Matters: This tool aims to provide a similar interactive coding experience to Anthropic's Claude Code, while being open-source and customizable. It is designed to help developers speed up their workflow by providing quick access to commands and context understanding. ### Repository: syscalldev/LLMCode ### GitHub Stars: 64 ### GitHub Description: LLM Code (Open Claude Code) is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster through natural language commands.

### Overview: This repository is part of a position paper titled "Quo Vadis, World Modeling? Towards Interactive World Proxies for Continually Improving Agents." It serves as a curated list of works related to agentic world modeling, focusing on the transition from passive world models to interactive world proxies that provide agents with the necessary information to improve their actions. ### Key Features: - **Dynamics Proxy**: Returns the future state of the world. - **Spatial Proxy**: Provides a rendered view of the environment. - **Execution Proxy**: Returns the execution result of an action. - **Memory / Experience Proxy**: Retrieves past experiences or memories. - **Skill Proxy**: Verifies the correctness of an action or skill. - **Reward / Verification Proxy**: Provides feedback on the success of a plan or action. ### Tech Stack: HTML ### Why it Matters: This repository aims to bridge the gap between traditional world models and interactive world proxies, emphasizing the importance of continuous improvement and agent-centric interfaces in enhancing agent performance and learning capabilities. ### GitHub Stars: 64 ### Repository: worldbench/awesome-agentic-world-model ### Primary Language: HTML ### GitHub Description: 🔥 Quo Vadis, World Modeling? Towards Interactive World Proxies for Continually Improving Agents ### Table of Contents: - Background - Classical World Models & Model-Based RL - Continual Improvement & Interaction Cost - The Six Proxy Functions - Dynamics Proxy - Spatial Proxy - Execution Proxy - Memory, Experience & Skill Proxy - Reward / Verification Proxy

Overview: This project focuses on a frame-by-frame closed-form update method for mask-based adaptive maximum variance distortion (MVDR) beamforming, enhancing speech enhancement. Key features: - Frame-by-frame closed-form update - Mask-based adaptive MVDR beamforming - Improved speech enhancement performance Tech stack: Python Why it matters: This method offers a significant improvement in speech enhancement by efficiently updating the beamforming process in real-time, making it highly effective for applications requiring high-quality speech recognition and noise reduction.

### Overview: Open Character AI is an open-source alternative to Character.AI, offering interactive persona chat portals and customizable AI characters powered by Next.js, Prisma, Supabase, NextAuth, and Stripe. ### Key Features: - **Interactive Chat Studio**: Explore preset personas, create custom characters, and adjust LLM parameters per chat. - **Dashboard & Character Builder**: Choose from categories, create personas with customizable details, and manage chat history. - **Stripe Credit Billing**: Manage credit plans and pay-as-you-go options. - **Google Authentication & Persistence**: Securely authenticate users and manage credit balances. ### Tech Stack: - Next.js (App Router) - Prisma ORM - Supabase - NextAuth - Stripe ### Why It Matters: - **Production-Ready SaaS**: Built with Google OAuth and Stripe Checkout workflows. - **High-Fidelity Chat UI**: Features a premium dark mode, message bubbles, typing indicators, custom avatars, and a parameter console. - **Dynamic LLM Tuning**: Adjusts temperature, max tokens, and system prompt per chat thread. - **Persistent Chat History**: Securely saves all conversations, chat history, and creations in a PostgreSQL database. ### Tech Stack: - Next.js (App Router) - Prisma ORM - Supabase - NextAuth - Stripe ### Why Use Open Character AI? - **Production-Ready SaaS**: Complete with Google OAuth and Stripe Checkout workflows. - **Glassmorphic Chat UI**: Features a premium dark mode, message bubbles, typing indicators, custom

### Overview: This repository implements time series diffusion in the frequency domain, focusing on the application of diffusion models in the frequency domain for time series data. ### Key Features: - **Time Series Diffusion in the Frequency Domain**: The primary focus is on applying diffusion models in the frequency domain for time series data. - **Hydra Override Syntax**: Allows for hyperparameter modification to retrain models as described in the paper. - **Datasets**: Supports datasets such as ECG, MIMIC III, NASA, NASDAQ, and USDroughts. - **Model Training**: Simple command-line interface for training models in the time domain. - **Model Sampling**: Command to sample from trained models, with the ability to reproduce results from the paper. ### Tech Stack: - **Python 3.10** or newer - **conda** for environment management - **freqdiff** package for implementing time series diffusion in the frequency domain ### Why It Matters: This repository is significant because it provides a practical implementation of time series diffusion in the frequency domain, offering a solution for analyzing and processing time series data in a novel way. The ability to train and sample from models trained in the frequency domain can be particularly useful in fields such as climate science, financial analysis, and more, where time series data is prevalent and complex. ### Repository: JonathanCrabbe/FourierDiffusion ### Primary Language: Jupyter Notebook ### GitHub Stars: 63 ### GitHub Description: This repository implements time series diffusion in the frequency domain. ### README Excerpt: This repository implements

### Overview: Decision Forge is a tool designed to compile organizational policy into verifiable decision training data, ensuring AI agents are trustworthy. ### Key Features: - **Decision Context Reconstruction**: Reconstructs each decision moment with explicit candidate actions, grouped into families. - **Judgment Standard Compilation**: Compiles rules with verified source quotations, goes through business review, and labels every example. - **Execution Receipt Ingestion**: Ingests runtime events and reviewed outcomes as feedback for the next dataset version. ### Tech Stack: Python ### Why It Matters: Decision Forge turns organizational judgment into verifiable data, ensuring AI agents are trustworthy and aligned with an organization's specific sense of right.

Overview: PlayJev is a 0.8B JEV-like model trained to play ten classic browser games from raw pixel inputs. It achieves impressive performance with a latency of 43 ms per move. Key features: - Plays ten classic browser games - Uses raw pixel inputs for gameplay - Achieves 0.57 vs teacher performance - Trained with 2.2M frames Tech stack: JavaScript Why it matters: PlayJev demonstrates the potential of multimodal models in game AI, showcasing real-time decision-making in a highly competitive environment. Repository: OmniJev/PlayJev GitHub stars: 62 GitHub description: 🚀🚀 A 0.8B JEV-like multimodal model playing GUI games directly from raw pixels.

Overview: This extension allows developers to create simple Discord bots that can recognize and respond to speech, making it easier to integrate speech recognition functionality into their bot projects. Key features: - Speech recognition for Discord bots - Integration with Discord.js library - Ready-to-use example bot in the examples directory Tech stack: TypeScript Why it matters: This extension simplifies the process of creating bots that can understand and respond to spoken commands, making it a valuable tool for developers looking to enhance their bot's capabilities.

Overview: This project profile for the "AI Engineering Primer" repository is designed to provide a comprehensive overview of the course, its key features, the technology stack used, and its significance in the field of AI engineering and AI design patterns. Key features: - AI Engineering Core Concepts - Agentic AI - Design Patterns - RAG - Retrieval Augmented Generation - Projects | Building Agentic AI Workflows from scratch - Building Multi-agent workflow in LangChain & LangGraph - Agentic AI Framework - Reading Material & Code-Labs Tech stack: The course utilizes Python, PyTorch, LangChain, Hugging Face, and Google Gemini. Why it matters: This course is essential for developers looking to bridge the gap between theoretical Machine Learning and AI Engineering concepts, and to understand and implement agentic workflows and design patterns in production-ready AI engineering workflows and models.

Overview: IPDiff is an official implementation of the ICLR 2024 paper "Protein-Ligand Interaction Prior for Binding-aware 3D Molecule Diffusion Models". This project aims to enhance the understanding and prediction of protein-ligand interactions in molecular diffusion models. Key features: - Pretrained IPNet models for efficient inference - Data preparation following TargetDiff methodology - Environment setup with conda - Pretrained IPDiff model download and installation - Training script for IPDiff model - Sampling script for generating split data - Evaluation script for model performance - Metrics calculation script Tech stack: Python Why it matters: This project provides a practical solution for researchers and developers working on protein-ligand interaction modeling, offering a pre-trained model and tools for efficient data preparation and model evaluation, making it a valuable resource in the field of computational chemistry and molecular biology. Repository: YangLing0818/IPDiff Primary language: Python GitHub stars: 63

Overview: This repository contains 8 tested recipes for Wan 2.7 Video Edit prompts, including example renders, duration, and the cost per image on the APIMart route. Key features: - 8 tested Wan 2.7 Video Edit prompts with example renders, duration, and cost on the APIMart route. - 300+ models accessible via the Wan 2.7 Video Edit (`wan2.7-videoedit`) endpoint. - Subsequent pricing is based on a single API key, with no subscription or prepaid plans. - Charges are reported on completed tasks and refunded for failed ones. - Designed as a non-blocking API call, requiring only the `task_id` to be polled. Tech stack: Python Why it matters: This repository provides a concise and practical guide for developers looking to use Wan 2.7 Video Edit prompts effectively, with clear instructions on how to structure prompts and the cost implications of each recipe.

Overview: Dana is an agent-native programming language designed to capture, retain, and multiply critical operating knowledge that exists in people's heads. This project aims to address the $3.1 trillion problem by enabling enterprises to leverage cognitive ontology to optimize decision-making and expertise sharing. Key features: - Captures and retains knowledge from human operators. - Understands how processes connect and why decisions are made. - Provides instant access to domain expertise across the enterprise. - Enables knowledge to evolve and multiply as expertise is shared. Tech stack: Python Why it matters: Dana offers a novel solution to the traditional challenges of knowledge capture, retention, and multiplication, making it a critical tool for enterprises looking to optimize their decision-making processes and enhance their operational efficiency. Repository: aitomatic/dana Primary language: Python GitHub stars: 63 GitHub description: Domain-Aware Neurosymbolic Agent (Dana), an agent-native programming language README excerpt: <div align="center"> <img src="https://raw.githubusercontent.com/aitomatic/dana/release/docs/.archive/0804/images/dana-logo.jpg" alt="Dana Logo" width="80"> </div> Repository: aitomatic/dana Primary language: Python GitHub stars: 63 GitHub description: Domain-Aware Neurosymbolic Agent (Dana), an agent-native programming language

Overview: Grok Imagine Image 2.0 API provides high-fidelity image generation at a cost per image by resolution and quality, with a single base URL and flexible pricing. It supports OpenAI-compatible endpoints and offers a $1 minimum charge per task. Key features: - One key, entire catalog: Reach Grok Imagine Image 2.0 and over 300 other models with a single API endpoint. - Pay as you go: No subscription required; top up from $1 and spend on calls. - Charge back in response: The cost is reported in the response for every completed task. - Async by design: Tasks can be submitted and monitored via the `task_id`. Tech stack: Python Why it matters: This API simplifies access to a wide range of image models, offering flexible pricing and a single endpoint for teams looking to integrate with OpenAI-compatible services. GitHub stars: 63 Repository: apimartkjp84/grok-imagine-image-2.0-api-ko

Overview: mlx-speech is an open-source Python library for text-to-speech (TTS) and automatic speech recognition (ASR) on Apple Silicon Macs. It supports voice cloning, audio editing, sound effects, and live transcription, all running locally in MLX. Key features: - Text-to-speech (TTS) - Speech-to-text (ASR) - Voice cloning - Audio editing - Sound effects - Live transcription Tech stack: Pure-MLX framework with MLX inference. Why it matters: mlx-speech simplifies the process of creating high-quality voice cloning and speech synthesis effects on Apple Silicon Macs, making it easier for developers to integrate these features into their projects without the need for complex setups or external services.

### Overview: Dabarqus is a stand-alone application that implements a complete RAG solution, designed to be easy to use and integrate with existing applications. ### Key Features: - Ingest documents, databases, and APIs - LLM-style prompting - REST API for comprehensive control - Multiple semantic indexes (memory banks) ### Tech Stack: C++ with built-in vector search capabilities ### Why It Matters: Dabarqus simplifies the deployment and maintenance of RAG solutions, eliminating the need for Python, vector databases, and licensing challenges. It provides a straightforward, all-in-one solution with a REST API for easy integration, making it ideal for developers and companies looking to integrate RAG solutions into their existing systems.

### Overview: TransVW is an innovative method for self-supervised learning that leverages the semantic patterns in anatomical structures across medical images. This approach automatically extracts robust visual words that encode rich semantics about human anatomy, enabling CNNs to learn generic semantic representations without manual labeling. TransVW is a fully autodidactic scheme, serving as an add-on strategy to existing self-supervised methods and yielding pre-trained models that require minimal annotation effort for various 3D medical applications. ### Key Features: - **Transferable Visual Words (TransVW)**: Sophisticated, consistent, and recurring anatomical patterns are extracted from unlabeled medical images, serving as strong yet free supervision signals for CNNs. - **Fully Autodidactic Scheme**: No manual labeling required, making it an efficient add-on strategy. - **Semantics-Enriched Representation**: Learns generic semantic representations through self-supervision, enhancing the performance of existing self-supervised methods. - **Pre-trained Models**: Yields pre-trained models that require minimal annotation effort for various 3D medical applications through transfer learning. ### Tech Stack: Python ### Why It Matters: TransVW offers superior transfer learning performance, accelerates the training process, and reduces the annotation cost, making it a valuable tool for medical image analysis and self-supervised learning.

Overview: iGemini is a powerful Apple Shortcut that seamlessly integrates Google Gemini into your Apple ecosystem, allowing you to interact with AI capabilities naturally from Siri, the Share Sheet, or your Apple Watch without switching apps or opening browser tabs. Key features: - Integrates Google Gemini with Apple's ecosystem - Sends queries to Gemini and returns intelligent, contextual answers - Lightweight interface within Shortcuts environment - Supports multiple languages: English, Italian, German, Spanish Tech stack: Swift, Apple Shortcuts, Google Gemini API Why it matters: iGemini simplifies the integration of advanced AI capabilities into your daily workflows, making it easier to access Gemini's insights directly from your Apple devices without the need for switching apps or opening browser tabs.

### Overview: This comprehensive guide is designed for engineers building AI systems at the application layer. It covers essential concepts and mechanics of applied AI engineering, including Retrieval-Augmented Generation (RAG), embeddings, vector search, agents, Model Context Protocol (MCP), and more. The material is presented in a structured format, addressing why each concept is necessary, how it actually works, and what goes wrong in production environments. ### Key Features: - **Retrieval-Augmented Generation (RAG)**: Explains how to build systems that are correct, fast, cheap, and safe. - **Embeddings & Search**: Discusses how to use embeddings for similarity metrics and vector databases. - **Vector Databases**: Covers choosing a store, metadata filtering, and scaling operations. - **RAG Pipeline**: Details the retrieval and aggregation process. - **Agents**: Explains the agent loop, tool calling, and reliability in multi-agent systems. - **MCP**: Introduces the Model Context Protocol for building MCP servers. - **Production Engineering**: Addresses scalability, cost, latency, and failure modes in LLM pipelines. ### Tech Stack: - **MDX** (Markdown Documentation) ### Why It Matters: - This guide is not a machine-learning course. It focuses on the practical aspects of building AI systems, addressing the specific pain points and challenges that arise in real-world applications. ### Repository: - **Repository**: karthikreddy-7/ai-engineering-playbook - **GitHub Stars**: 63 - **GitHub Description**: A zero-to-1

### Overview: Perchance T2I Desktop is a powerful, cross-platform desktop application that integrates with the Perchance API to generate high-quality images from text descriptions. Built with Tauri, Rust, and Vue 3, it offers a lightning-fast and premium user experience. ### Key Features: - **Powerful Generation**: Full integration with Perchance AI for diverse image generation. - **Hardware Upscaling**: Intelligent image upscaling using WebSR for crisp, high-resolution results. - **Style Management**: Extensive Style Picker with categorized prompt templates, and a Style Manager for custom styles. - **Advanced History**: Automatically saves every generation, sorted by Date, Seed, or Likes, and a Favorite system for tracking best results. - **Interactive Preview**: High-performance image viewer with smooth zoom and pan. - **Modern Aesthetics**: Sleek dark mode interface built with PrimeVue and custom CSS. - **Internationalization**: Support for multiple languages with prompt translation indicators. ### Tech Stack: - **Framework**: Tauri - **Frontend**: Vue.js 3 (Composition API) - **State Management**: Pinia - **UI Components**: PrimeVue - **Styling**: Vanilla CSS & PrimeVue Passthrough (PT) - **Language**: Rust & TypeScript ### Why It Matters: Perchance T2I Desktop provides a premium user experience for generating high-quality images from text descriptions, making it a valuable tool for anyone looking to enhance their creative projects or experiments with AI-generated images.