Projects

### Overview: AutoPrompt is an automated method designed to create prompts for a diverse set of Natural Language Processing (NLP) tasks, specifically focusing on sentiment analysis, natural language inference, fact retrieval, and relation extraction. It demonstrates that masked language models (MLMs) have an innate ability to perform these tasks effectively. ### Key Features: - **Gradient-Guided Search**: AutoPrompt uses gradient-guided search to generate prompts. - **Diverse Tasks**: Supports sentiment analysis, natural language inference, fact retrieval, and relation extraction. - **MLM Integration**: Utilizes masked language models (BERT and RoBERTa) to create prompts. - **Prompt Templates**: Provides templates for different tasks, including fact retrieval and relation extraction. ### Tech Stack: AutoPrompt is built using Python, with the following dependencies: - `conda`: For creating and activating a conda environment. - `pip`: For installing packages. - `spacy`: For downloading the English language model. ### Why It Matters: AutoPrompt showcases that MLMs can perform tasks beyond their primary purpose of language modeling, highlighting their potential for various NLP applications. This method opens up new possibilities for automated prompt generation and task adaptation.

Overview: Diffuman4D is a cutting-edge project that enables high-fidelity free-viewpoint rendering of human performances from sparse-view videos. It achieves this by utilizing spatio-temporal diffusion models to synthesize consistent 4D views from limited input data. Key features: - 4D Consistent Human View Synthesis - Sparse-View Video Reconstruction - Spatio-Temporal Diffusion Models Tech stack: Python Why it matters: Diffuman4D provides a powerful tool for researchers and developers interested in creating immersive and realistic 4D human performance videos from limited video data, offering significant advancements in the field of video synthesis and rendering. Repository: zju3dv/Diffuman4D Primary language: Python GitHub stars: 642 GitHub description: [ICCV 2025] Diffuman4D: 4D Consistent Human View Synthesis from Sparse-View Videos with Spatio-Temporal Diffusion Models

Overview: This repository hosts a suite of specialized agents designed to power your brainstorming sessions, bringing a unique perspective tailored to generate creative, targeted solutions for diverse challenges. Key features: - Structured Brainstorming Methods - Real-time streaming responses from AI - Beautiful UI with smooth animations and transitions - Secure API key management - Session history tracking - Structured output with organized idea trees - Dark mode support (coming soon) Tech stack: TypeScript, Next.js, Tailwind CSS, Zustand, OpenAI GPT models with streaming support. Why it matters: This repository offers a powerful tool for transforming your ideas with structured AI-powered brainstorming methods, making the creative process more efficient and effective.

Overview: BotLibre is an open-source platform designed for artificial intelligence, natural language processing, chat bots, virtual agents, social media automation, and live chat automation. Key features: - A web platform for developing and hosting bots for web, mobile, and social media. - An AI/NLP engine library written in Java. - JUnit test cases for the AI/NLP engine. - SDKs for Android, iOS, and web platforms. Tech stack: Java Why it matters: BotLibre provides a comprehensive solution for developers looking to integrate AI and automation into their applications, making it an essential tool for businesses and developers alike. Repository: BotLibre/BotLibre Primary language: Java GitHub stars: 641

Overview: Overture is an open-source, locally running web interface that maps out the execution plan of any AI coding agent as an interactive flowchart before the agent begins writing code. Key features: - Interactive visual flowchart - Real-time execution tracking - Approve the plan before code is written Tech stack: TypeScript Why it matters: Overture provides a solution to the common problem of working with AI coding agents where the agent writes code immediately without any visibility into its thought process. By rendering the plan as a flowchart, users can see dependencies, context requirements, and alternative approaches before proceeding, ensuring that the agent only writes the necessary code and avoids costly mistakes.

### Overview This repository provides an implementation of QA-MDT, a state-of-the-art model for text-to-music generation. The code and methods are based on the following repositories: - [AudioLDM](https://github.com/haoheliu/AudioLDM-training-finetuning) - [PixArt-alpha](https://github.com/PixArt-alpha/PixArt-alpha) - [MDT](https://github.com/lichang0928/QA-MDT) ### Key Features - **Text-to-Music Generation (TTM):** Generate music from text inputs. - **Quality-Aware Masked Diffusion Transformer (QA-MDT):** A model that extends QA-MDT to generate long-range music sequences. - **Zero-Shot Extension:** Successfully extended the model to generate inf-long music sequences without any fine-tuning. ### Tech Stack - Python - PyTorch - Hugging Face Transformers ### Why It Matters QA-MDT is a significant advancement in the field of text-to-music generation, offering improved quality and performance compared to existing models. It has been successfully applied in various applications, including inf-long music generation and zero-shot extension. ### GitHub Stars - **GitHub Stars:** 640 ### Repository - **Repository:** ivcylc/OpenMusic - **Primary Language:** Python - **GitHub Stars:** 640 ### Demo - **Demo:** Try our demo at [QA-MDT](https://qa-mdt.github.io/) or [ModelScope](https://

Overview: SEED-LLaMA is an official implementation of SEED-LLaMA, a multimodal language model that excels in both comprehension and generation. It has been accepted by ICLR 2024 and is powered by the CV Center and Tencent AI Lab. Key features: - Multi-turn in-context multimodal generation - Compositional emergent abilities - Supports multi-granularity comprehension and generation Tech stack: Python Why it matters: SEED-LLaMA is a significant advancement in the field of multimodal language models, offering unique capabilities for both comprehension and generation. Its acceptance by ICLR 2024 highlights its importance in the research community. Repository: AILab-CVC/SEED GitHub stars: 640 GitHub description: Official implementation of SEED-LLaMA (ICLR 2024)

### Overview: This repository contains a collection of end-to-end generative AI projects focused on LLM models, RAG, AI agents, AI chatbots, and multi-modal applications. The projects are designed to showcase the capabilities of various LLM models and their integration with deployment tools like Streamlit and HuggingSpaces. ### Key Features: - **Multi-PDFs 📚ChatApp AI Agent 🤖**: A chat application that integrates Langchain, Google Gemini Pro, and FAISS vector database for seamless streamlit deployment. - **🖼️Image to Speech GenAI Tool Using LLM 🌟♨️**: Converts image context into audio stories using HuggingFace models and Google Gemini Pro. - **Youtube Video Transcribe Summarizer LLM App**: Converts video transcripts into comprehensive study materials using Google Gemini Pro. - **End to end RAG LLM App**: Guides users through the process of building a RAG LLM app using LLama2 and LLaMAindex. - **Resume ATS Tracking LLM Project**: Optimizes the recruitment process by integrating an advanced Applicant Tracking System with Google Gemini Pro. - **End To End Text To SQL LLM App Along With Querying SQL Database**: Converts natural language queries into SQL commands and queries SQL databases directly. ### Tech Stack: `F/w:` Langchain, `Model`: Google Gemini Pro, `Vector DB`: FAISS, `Deployment`: Streamlit, `Model`: HuggingFace Models, `Vector DB`: Hugging Spaces. ### Why it Matters: These projects demonstrate the versatility and integration capabilities of

### Overview: ChatLLM Web is a private local model studio, AI chat, and agent workspace powered by WebGPU and WebLLM, designed for seamless integration with modern web browsers. ### Key features: - **Model Studio**: Offers a curated catalog of 20 chat, coding, reasoning, vision, and tool-capable models. - **Local Agent Workspace**: Enables running native WebLLM functions with three Hermes models. - **Local Model Studio**: Keeps conversations, selected files, tool results, and artifacts on the device. - **WebGPU Inference**: Supports WebGPU for inference tasks. - **Model Catalog**: Provides 65 logical models backed by 163 official WebLLM records. - **Advanced Catalog**: Allows device features to select a compatible quantization. - **Model Management**: Supports downloading, loading, switching, unloading, retrying, and deleting models without refreshing the page. - **File Management**: Allows attachment of TXT, Markdown, JSON, and common code files as direct local context. - **Model Sources**: Supports importing advanced MLC manifests from approved HTTPS model sources. - **PWA Integration**: Installs the PWA and reuses cached models after the first successful load. - **Localization**: Supports English or Simplified Chinese, light or dark themes. ### Tech stack: TypeScript ### Why it matters: ChatLLM Web combines AI chat, a local agent workspace, and a model studio into a seamless browser-native product, leveraging WebGPU for inference and WebLLM for model execution. The curated catalog of models ensures a high

### Overview: KoELECTRA is a Korean-specific ELECTRA model pretrained on 34GB of Korean text, featuring Wordpiece tokenization and compatibility with TensorFlow. ### Key features: - **KoELECTRA-Base**: Base model for Korean, trained on 34GB of Korean text. - **KoELECTRA-Small**: Smaller model for faster inference. - **Wordpiece Tokenization**: Uses Wordpiece tokenization for better performance. - **Compatibility with TensorFlow**: Deployable without requiring additional libraries. ### Tech stack: Python ### Why it matters: KoELECTRA simplifies the process of using pre-trained models by providing a Korean-specific ELECTRA model and compatibility with TensorFlow, making it easier to integrate into existing projects.

Overview: Liquid is a unified multi-modal generation model that seamlessly integrates visual comprehension and generation, eliminating the need for external models. Key features: - Seamless integration of visual comprehension and generation - Single large language model (LLM) for scalable and unified generation - Supports text-to-image and text-to-text tasks Tech stack: Python Why it matters: Liquid simplifies the integration of visual understanding with language models, making it easier to create unified and scalable solutions for multimodal tasks.

Overview: Sonus is a lightweight, offline speech-to-text library for Node.js, designed to easily integrate voice user interfaces into hardware and software projects. Key features: - Offline hotword detection - Supports Google Cloud Speech, Alexa Voice Services, Wit.ai, Microsoft Cognitive Services, and Houndify - Simple and dead-simple installation - Compatible with Linux, macOS, and most major distros of Windows Tech stack: JavaScript Why it matters: Sonus simplifies the integration of voice recognition into projects, making it accessible for developers without requiring extensive knowledge of machine learning or cloud services. Repository: evancohen/sonus GitHub stars: 639

Overview: Haiku.RAG is an Agentic RAG system designed for local and self-hosted document search. It leverages LanceDB for indexing, Pydantic AI models, and Docling parsing, offering hybrid retrieval, reranking, and multimodal capabilities. Key features: - Ingests PDFs, office documents, HTML, Markdown, and images using Docling. - Supports hybrid vector and full-text retrieval, cross-encoders, Jina, and more. - Provides section-aware context expansion and image search with multimodal embedders. - Offers the RAG capability for answering questions with citations to page numbers and headings. - Integrates with Python API, Pydantic AI, and a MCP server for Claude Code, Codex, and Claude Desktop. - Operates with the `haiku-ingester` service, tags, rollback, and health checks. Tech stack: Python, LanceDB, Pydantic AI, Docling, Ollama, Jina, and OpenRouter. Why it matters: Haiku.RAG simplifies document search and retrieval, making it accessible and efficient for local use cases, while offering advanced features like multimodal and section-aware context expansion, and the ability to integrate with various AI models and platforms. Repository: ggozad/haiku.rag Primary language: Python GitHub stars: 636 GitHub description: Agentic RAG for local and self-hosted document search: hybrid retrieval, reranking and multimodal RAG on embedded LanceDB, with Docling parsing and an MCP server.

Overview: This repository contains the official implementation of the paper "Practical Stereo Matching via Cascaded Recurrent Network with Adaptive Correlation" by Jiankun Li et al., presented at CVPR 2022. The project is based on MegEngine and includes datasets, dependencies, and inference/evaluation scripts. Key features: - Implementation of the cascaded recurrent network for stereo matching - Use of adaptive correlation for improved accuracy - Support for various public datasets including SceneFlow, Sintel, Middlebury, ETH3D, KITTI 2012/2015, Falling Things, InStereo2K, and HR-VS - Pretrained MegEngine model for inference and training - Docker container available for quick execution Tech stack: Python, MegEngine, OpenCV, NumPy, Pillow, TensorboardX Why it matters: This project provides a practical and effective solution for stereo matching, showcasing the capabilities of MegEngine and the cascaded recurrent network approach.

### Overview: AINL is a compact, graph-canonical, AI-native programming system designed for teams building AI workflows. It simplifies complex AI tasks by enabling multiple steps, state management, tool integration, and improved workflow efficiency. ### Key Features: - **Graph-Canonical Workflow**: AINL uses a graph-first deterministic IR, ensuring efficient and predictable execution. - **Compact Syntax**: Simplified syntax for concise and readable code. - **State Management**: Built-in support for state and memory management. - **Tool Integration**: Seamless integration with various AI tools and platforms. - **Reusability**: Encourages reusability and modularity in AI workflows. ### Tech Stack: AINL is built using Python, leveraging the capabilities of the programming language to create a robust and efficient AI-native workflow system. ### Why It Matters: AINL is designed to streamline the development and deployment of AI workflows, making it easier for teams to create and maintain AI-driven applications without the need for long, repetitive prompts. It addresses the challenges of maintaining a low dependency on LLMs and provides a more structured and efficient approach to AI development.

Overview: QMedia is an open-source AI content search engine designed for content creators. It supports the extraction of text, images, and short videos, enabling efficient analysis and integration of scattered information. Key features: - Content cards: Display image/text and video content in the form of cards. - Multimodal content RAG: Search for image/text and short video materials. - Pure local multimodal models: Implementing models for image/text and video content extraction in a local environment for offline use. Tech stack: TypeScript, Next.js, TailwindCSS, Shadcn/UI, Python, LlamaIndex. Why it matters: QMedia provides a platform for content creators to share and exchange ideas on AI content creation, making it an essential tool for anyone involved in the field of AI content generation.

Overview: This project presents a winning solution for the MICCAI 2017 Robotic Instrument Segmentation Sub-Challenge, improving upon existing results using advanced neural network architectures. Key features: - Improved U-Net architecture with pre-trained encoders - Enhanced performance in binary and multi-class robotic instrument segmentation - Superior results for automatic instrument segmentation Tech stack: Jupyter Notebook Why it matters: This project demonstrates the effectiveness of deep learning techniques in robotic surgery, providing a solid foundation for future advancements in instrument tracking and pose estimation.

### Overview: This repository is a curated list of top foundation and multimodal models, featuring models that can be used for zero-shot object detection, depth estimation, and efficient segment anything. These models are designed to process multiple modalities and are widely used in various applications. ### Key Features: - **YOLO-World**: A real-time open-vocabulary object detection model. - **Depth Anything**: A depth estimation model. - **EfficientSAM**: A model that leverages masked image pretraining for efficient segment anything. - **Qwen-VL-Plus**: A model that combines Qwen with VL-Plus for improved performance. ### Tech Stack: The repository is written in Python and utilizes various libraries and frameworks such as Hugging Face Spaces, Colab, and others. ### Why It Matters: These models are crucial in the field of computer vision and artificial intelligence, providing powerful tools for tasks such as object detection, depth estimation, and segmentation. They are widely used in research and practical applications, making them essential for developers looking to improve their models and applications.

Overview: This repository is a curated collection of papers, surveys, benchmarks, frameworks, and blog posts focused on machine unlearning in large language models. It serves as a resource for researchers and practitioners interested in the field. Key features: - Papers: 2026, 2025, 2024, 2023, 2022, 2021 - Surveys and Position Papers: 19 - Frameworks: 3 - Blog Posts: 2 Tech stack: Unknown Why it matters: This repository provides a comprehensive overview of existing research in machine unlearning, helping users stay updated with the latest developments in the field.

### Overview: Convert any face into a sticker using the `face-to-sticker` model. This project leverages ComfyUI and Replicate for seamless integration and accessibility. ### Key Features: - **Model Compatibility**: Works with Replicate and ComfyUI. - **Custom Nodes**: Requires specific custom nodes for optimal performance. - **GitHub Stars**: 635, indicating high user interest and engagement. ### Tech Stack: - Python ### Why It Matters: This project is designed to easily convert any face into a sticker, making it a valuable tool for creators and enthusiasts alike. The inclusion of specific custom nodes ensures that the project performs at its best, making it a reliable choice for those looking to enhance their projects.