Projects

Overview: This project implements a real-time image and video object detection classifier using pretrained YOLOv3 models. Key features: - Real-time object detection on images and videos - Pretrained YOLOv3 models from the official YOLOv3 paper - Option to perform classification on webcam Tech stack: Python Why it matters: This project provides a practical solution for real-time object detection, making it useful for applications requiring immediate analysis of images and videos, such as security systems, autonomous vehicles, or surveillance cameras.

### Overview: Mind Map Wizard is an AI-powered text-to-mind map generator that allows users to create comprehensive visual overviews of any topic in seconds. The project is open-source and integrates with Ollama for local AI support, enabling users to generate mind maps directly in their browsers. ### Key Features: - 🤖 AI Mind Map Generation: Automatically generates a mind map based on the input text. - 🔒 Local AI support (Ollama): Provides quick and accurate AI-generated summaries. - 🌐 Web Search: Access to web resources for deeper insights. - 🖼️ Image Inserting: Easily add images to enhance the mind map. - ✏️ Direct Mind Map Editing: Modify the mind map directly. - ✍️ AI Editing Assistant: Enhances the mind map with AI-generated summaries. - 🕹️ Drag and Drop Reordering: Easily rearrange the mind map structure. - 🎨 Style Customization: Customize the mind map's appearance. - ☑️ Checkboxes: Add checkable nodes for quick navigation. - 💾 Multiple Export Formats: Export the mind map in various formats. - 📝 AI Summaries: Generate AI-written summaries. - 📚 Generation History: Track and review the mind map's history. - ➕ Mind Map Expansion: Automatically expand mind map branches. - 🌈 Branch Coloring: Customize the color of mind map branches. - ↩️ Redoing & Undoing: Undo and redo actions easily. - 🔍 Zoom & Pan: Adjust the zoom level and pan the

### Overview: PromptAgent is a novel automatic prompt optimization method designed to craft prompts equivalent in quality to those handcrafted by experts. It leverages the principles of Monte Carlo Tree Search (MCTS) to study prompt optimization, providing a principled framework for understanding and improving prompt optimization. ### Key Features: - **Expert-Level Prompt Optimization**: Craft prompts that are as effective as those created by experts. - **Unified Sampling and Rewarding**: Combines prompt sampling with MCTS to find high-reward paths. - **Versatile Models**: Supports various language models, including OpenAI, Hugging Face, and VLLM. - **Easy Installation**: Requires minimal configuration and can be run directly from the command line. ### Tech Stack: Python ### Why It Matters: PromptAgent offers a novel approach to prompt optimization, providing a principled framework for studying and improving prompt crafting. By focusing on expert-level prompts, it addresses a challenging yet untapped area in the field of prompt optimization, making it a valuable tool for researchers and practitioners in the field of language models and natural language processing. ### Repository: - **Repository**: maitrix-org/PromptAgent - **Stars**: 358 - **Description**: Official repo for "PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization". - **News**: - **Aug. 04, 2024**: Support for VLLM LLMs, speeding up local inference. - **Jun. 27, 2024**: Integrates LLM Reason

Overview: Fondant is a production-ready data processing framework designed for collaborative dataset building, making it easy and shareable. Key features: - Initialize datasets - Apply various operations - Share operations with others - Track dataset versions Tech stack: Python Why it matters: Fondant simplifies the process of data processing and sharing, enabling teams to collaborate effectively and efficiently on large datasets.

Overview: Companion is an Obsidian plugin that enhances your note-taking and personal knowledge management with AI-powered autocomplete, including support for ChatGPT and Ollama models. Key features: - AI-powered autocomplete for note-taking and knowledge management - Suggestions for completing words, phrases, and sentences based on context - OpenAI's GPT-3 and GPT-3.5, including ChatGPT, and locally hosted Ollama models - Unobtrusive ghost text suggestions that can be accepted or ignored - Supports more models planned in the future Tech stack: TypeScript Why it matters: Companion streamlines your note-taking process by enhancing productivity and creativity, making it easier to write more efficiently and effectively.

Overview: Agentica is a cutting-edge AI Agent framework designed for multi-session collaboration, persistent memory, and skill-driven self-evolution. It offers a single model across CLI, Web, Desktop, and Python SDK, achieving high accuracy and efficiency with minimal token usage. Key features: - Multi-session collaboration across CLI, Web, Desktop, and Python SDK - Persistent memory for seamless data access and session continuity - Skill-driven self-evolution for continuous improvement and adaptability - Ultra-efficient token usage with minimal token consumption Tech stack: Python Why it matters: Agentica simplifies AI development by providing a unified model across multiple platforms, ensuring high accuracy and efficiency with minimal resource usage. Its multi-session collaboration and persistent memory features make it highly suitable for complex projects requiring real-time collaboration and data persistence. Repository: shibing624/agentica GitHub stars: 358 GitHub description: Agentica is a framework that brings together a single model across CLI, Web, Desktop, and Python SDK, achieving high accuracy and efficiency with minimal token usage. It supports multi-session collaboration, persistent memory, and skill-driven self-evolution. The framework is designed for complex projects requiring real-time collaboration and data persistence.

Overview: FIBO is a state-of-the-art, open-source text-to-image model designed for professional workflows, built specifically for generating images from structured JSON captions up to 1,000+ words. This model ensures precise control over lighting, composition, color, and camera settings, fostering native disentanglement and strong prompt adherence. Key features: - VLM guided JSON-native prompting - Iterative controlled generation - Disentangled control - Enterprise-grade, 100% licensed data - Strong prompt adherence - Built for production, including API endpoints and ComfyUI nodes Tech stack: Python Why it matters: FIBO sets a new standard for controllability, predictability, and disentanglement, making it ideal for professional applications requiring precise image generation. Repository: Bria-AI/FIBO Primary language: Python GitHub stars: 356 GitHub description: FIBO is a SOTA, first open-source, JSON-native text-to-image model built for controllable, predictable, and legally safe image generation.

### Overview: Data-Centric AI is a methodology that focuses on the dataset as the core of the AI solution, rather than the model. This approach emphasizes the importance of understanding and profiling the data before developing any AI models. ### Key Features: - **Data Profiling Libraries:** - YData Profiling: Supports both Pandas and Spark DataFrames, offering fast and straightforward data understanding. - SweetViz: An open-source Python library for generating high-density visualizations in EDA. - DataPrep.EDA: An EDA tool for understanding Pandas/Dask DataFrames with just a few lines of code. - AutoViz: Automatically visualizes datasets with a single line of code. - Lux: Facilitates fast and easy data exploration by automating the visualization and data analysis process. - Great Expectations: Helps data teams eliminate pipeline debt through data testing, documentation, and profiling. - D-Tale: An open-source auto-visualization library that simplifies viewing and analyzing Pandas data structures. - Data Profiler: A Python library for making data analysis, monitoring, and sensitive data detection easy. - whylogs: An open-source standard for data logging, using data profiling techniques to create whylogs profiles for monitoring and observability. ### Tech Stack: - Jupyter Notebook ### Why It Matters: This repository serves as a curated list of resources for understanding and implementing Data-Centric AI. It includes libraries, tools, and tutorials that help in developing and profiling datasets effectively. By leveraging these resources, developers can enhance

### Overview: Generative AI Application Builder on AWS (GAAB) is a web-based management dashboard designed for developers to deploy, experiment with, and compare different combinations of Large Language Models (LLMs) for generative AI use cases. The solution integrates with Amazon Bedrock and its included LLMs, such as Amazon Titan, and pre-built connectors for third-party LLMs, making it accessible for novice to experienced users. ### Key Features: - **Rapid Experimentation:** Users can quickly experiment with different LLM use cases without requiring deep AI expertise. - **Modular Architecture:** The solution uses nested Amazon CloudFormation stacks for extendable and modularized architecture. - **Enterprise Ready:** Designed for companies to tackle real-world business problems. - **LLM Comparison & Experimentation:** Supports multi-LLM comparison and experimentation with metric tracking using Amazon CloudWatch dashboards. - **Model Providers & Use Cases:** Integrates with Amazon Bedrock and Amazon SageMaker AI, offering a growing list of model providers and Gen AI use cases. ### Tech Stack: TypeScript ### Why it Matters: GAAB simplifies the process of developing, rapidly experimenting with, and deploying generative AI applications, making it accessible for developers without deep AI expertise. It enables the productionization of different Gen AI use cases, integrating with Amazon Bedrock and SageMaker AI, and provides a scalable and modular solution for enterprise-level applications.

### Overview: Hallo-Live is a real-time text-driven joint audio-video avatar generation framework that achieves 20.38 FPS with 0.94 second latency on two NVIDIA H200 GPUs, preserving strong lip-sync accuracy, visual fidelity, and speech quality. ### Key Features: - **Causal Dual-Stream DiT Model**: Adapts a pretrained dual-stream DiT to the streaming setting using cross-modal future-expanding block-causal mask. - **Autoregressive Self-Rollout**: Performs autoregressive self-rollout with the audio-video KV cache to optimize the generated trajectory. - **Stream-Inference**: Maintains the audio-video KV cache for streaming inference. ### Tech Stack: Python ### Why it Matters: Hallo-Live demonstrates the potential of real-time text-driven audio-video avatar generation, showcasing its capability to generate synchronized avatar video and speech in a streaming manner with strong performance and accuracy.

### Overview: A PyTorch implementation of the Recurrent Variational Autoencoder (RVAE) model, designed to generate sequential data. ### Key features: - Implements Samuel Bowman's "Generating Sentences from a Continuous Space" with Kim's "Character-Aware Neural Language Models" embedding. - Supports sampling of generated sequences. - Includes training scripts for word embeddings and model training. ### Tech stack: Python ### Why it matters: This model is useful for generating coherent sequences of text, making it valuable for applications requiring natural language processing tasks such as text generation, summarization, and dialogue systems.

### Overview: EcoLogits tracks the energy consumption and environmental impacts of using generative AI models through APIs. ### Key features: - Tracks energy consumption and GHG emissions. - Integrates with various generative AI providers. - Supports multiple providers including Anthropic, Cohere, Google GenAI, Hugging Face Hub, MistralAI, and OpenAI. - Provides detailed environmental impact reports. ### Tech stack: Python ### Why it matters: EcoLogits helps users understand and mitigate the environmental impact of using generative AI models, promoting sustainability and responsible AI development.

Overview: tldrstory is a semantic search application designed for headlines and text content related to stories, utilizing zero-shot labeling and a txtai index for text similarity search. It includes a customizable Streamlit application and FastAPI backend service, making it suitable for dynamic categorization and analysis of processed data. Key features: - Semantic search for headlines and text content - Zero-shot labeling for categorization - Text similarity search with txtai index - Customizable Streamlit application and FastAPI backend - Supports RSS and Reddit API, with additional data sources configurable Tech stack: Python Why it matters: tldrstory simplifies the process of understanding complex news and headlines by providing a semantic search tool that can categorize and analyze text content, making it easier for developers and analysts to quickly grasp the essence of news articles and headlines.

Overview: ORBIT is a self-hosted AI gateway that allows you to connect your data and tools to local or hosted models through a single API. It supports hybrid retrieval and reranking, OpenAI-compatible model selection, and automatic provider fallbacks. Key features: - Hybrid retrieval and reranking - OpenAI-compatible model selection - Automatic model selection and provider fallbacks - Connects to local or hosted models - Deployed on your own infrastructure - Supports SQL, NoSQL, vector stores, and APIs - Connects to MCP tools, routes requests to specialized skills, and coordinates agents through A2A or async messaging - Processes documents, images, voice experiences, and generates images, audio, and video - Provides structured choices, yes/no answers, and scores for routing, triage, and moderation - Self-hosted on-premises or in a private cloud, including air-gapped deployments with local models - Security and operations built in, including SSO, role-based access, PII detection, moderation, quotas, audit logs, and cost estimates, plus an admin UI for day-to-day management - Real-time decisions with structured choices, yes/no answers, and probabilities for routing, triage, and moderation Tech stack: Python Why it matters: ORBIT is a powerful tool for connecting your data and tools to local or hosted models, making it easier to access and analyze your data in real-time. Its self-hosted infrastructure allows for flexibility and security, while its comprehensive features make it a valuable asset for any organization looking to enhance

### Overview Upgini is a Python library designed to simplify the process of finding and adding relevant features to your machine learning (ML) and artificial intelligence (AI) pipelines. It integrates with hundreds of public and premium external data sources, including open and commercial language models (LLMs), graph neural networks (GNNs), and recurrent neural networks (RNNs), to automatically generate optimal ML features. This approach significantly boosts the accuracy of ML models without the need for extensive hyperparameter tuning. ### Key Features - **Automatically Find Relevant Features**: Upgini identifies only the features that improve model accuracy, ensuring no wasted time on irrelevant data. - **Feature Generation**: Automatically generates features from external data sources using advanced techniques like LLMs, RNNs, and GNNs. - **Key Augmentation**: Automatically enhances search keys from all connected data sources to improve model performance. ### Tech Stack - Python ### Why It Matters Upgini simplifies the process of integrating external data sources into ML pipelines, making it easier for data scientists to leverage high-quality data that can significantly improve model accuracy. This tool democratizes access to external data, enabling more widespread adoption of external features in ML and AI models.

### Overview: This repository contains a collection of research papers and code snippets related to Large Language Model (LLM)-guided Multimodal Learning. It focuses on the advancements in the field of multimodal learning, particularly how LLMs can be used to enhance the understanding and interaction with visual, textual, and audio data. ### Key Features: - **Research Trends in LLM-guided Multimodal Learning**: This repository serves as a curated collection of papers and code snippets that explore the latest research in the field of LLM-guided multimodal learning. - **Backbones**: The repository includes models and backbones such as LLaMA, Alpaca, Vicuna, and others, which are used in various research papers. - **Learning Techniques**: It covers techniques such as full fine-tuning, parameter-efficient tuning (Adapter, LoRA), and in-context learning. - **Examples of LLM-guided Multimodal Models**: The repository provides examples of models like OpenFlamingo, MiniGPT-4, Otter, InstructBILP, and BLIVA. - **Evaluation on Multimodal LLM**: It includes evaluation methods like MultiInstruct, POPE, AttackVLM, and others to assess the performance of multimodal LLMs. ### Tech Stack: The primary programming language used in the repository is not specified. ### Why It Matters: This repository is valuable for researchers and developers interested in the latest advancements in multimodal learning and LLM-guided applications. It provides a comprehensive overview of research trends, models, and evaluation techniques,

Overview: SG-Nav is an online 3D scene graph prompting framework designed for zero-shot object navigation using LLMs. It enables seamless navigation across various scenes and categories without the need for training. Key features: - Online 3D scene graph prompting - Zero-shot object-goal navigation - Direct application to different scenes and categories - Accepted to NeurIPS 2024, CoRL 2025, and CVPR 2025 Tech stack: Jupyter Notebook Why it matters: SG-Nav simplifies the process of object navigation by leveraging pre-trained language models, making it accessible and effective for a wide range of applications without the need for extensive training.

Overview: This project is an official implementation of Pseudo Numerical Methods for Diffusion Models on Manifolds (PNDM), presented at ICLR 2022. It provides a comprehensive framework for generating images using DDIM-like models, including PNDM, DDIM, PF, and iDDPM. Key features: - Pseudo Numerical Methods for Diffusion Models on Manifolds (PNDM) - Denoising Diffusion Implicit Models (DDIM) - Score-Based Generative Modeling through Stochastic Differential Equations (PF) - Improved Denoising Diffusion Probabilistic Models (iDDPM) Tech stack: Python Why it matters: This project offers a versatile framework for generating images using various diffusion models, making it a valuable resource for researchers and practitioners in the field of generative modeling and diffusion-based image synthesis.

### Overview: Inversion-Free Image Editing with Natural Language ### Key features: - Inversion-Free Image Editing - Natural Language Control - High-Quality Image Generation ### Tech stack: Python ### Why it matters: This project introduces a novel approach to image editing that eliminates the need for image inversion, making it more efficient and natural. The method uses natural language to control the editing process, offering a more intuitive and user-friendly experience. The project is based on the latest versions of PyTorch and Hugging Face's Diffusers, ensuring state-of-the-art performance and compatibility with modern tools. Repository: sled-group/InfEdit Primary language: Python GitHub stars: 357 GitHub description: [CVPR 2024] Official implementation, Inversion-Free Image Editing with Natural Language README excerpt:

Overview: This repository contains the PyTorch implementation of a position-aware attention model for relation extraction, based on the paper "Position-aware Attention and Supervised Data Improve Slot Filling" by Zhang et al. (2017). Key features: - Position-aware attention model for relation extraction - PyTorch implementation - Supports both position-aware attention and LSTM models - Pretrained GloVe vectors for word embedding - Training scripts for both position-aware attention and LSTM models - Evaluation on the TACRED dataset - Model checkpoints and logs saved in `./saved_models/00` - Ensemble script provided for combining multiple models Tech stack: Python, PyTorch, GloVe vectors Why it matters: This model and implementation are useful for researchers and developers looking to improve relation extraction tasks, specifically in the context of the TACRED dataset. The position-aware attention model enhances the accuracy of relation extraction by considering the positional information of entities, while the LSTM model provides a flexible and powerful alternative for relation extraction tasks.