Projects

### Overview: An open-source, agent-guided creative-direction workflow designed to transform a single visual reference into a connected brand system using Lovart Agent. ### Key Features: - **Collect Project Brief**: Clearly defines the project's objectives. - **Answer Design Questions**: Provides accessible language to address design challenges. - **Analyze Reference Without Copying**: Analyzes references without reproducing them. - **Build Anchor Brand Kit**: Creates an Anchor Brand Kit with guidelines, key visuals, original campaign assets, and packaging concepts. - **Use Named Google Fonts**: Utilizes specific fonts for consistency. - **Approve Generation Plans**: Ensures all generation plans are approved before proceeding. - **Create Guidelines**: Generates guidelines in a single slide per module. - **Create Original Campaign Scenes**: Uses Brand Lock to generate original campaign scenes. - **Diagnose Visual Problems**: Identifies and resolves visual issues. - **Refine or Reject Weak Directions**: Improves or rejects weak design directions. - **Audit Consistency**: Ensures brand consistency before scaling. - **Optimized for Lovart**: Designed for Lovart Agent. - **Primary Language**: Unknown. ### Tech Stack: Unknown ### Why It Matters: This Skill streamlines the creative process, ensuring consistency and originality in branding projects. It helps non-designers and experienced teams alike by providing clear guidance and avoiding unnecessary reinvention.

### Overview: Schemagate is a text-to-SQL access control solution designed to help developers build more secure and efficient SQL agents. It ensures that only the tables and data the caller is allowed to read are included in the prompt, reducing the risk of unauthorized access and improving the overall security of the system. ### Key features: 1. **Schema Selection**: Schemagate filters the schema based on the caller's grants, ensuring that restricted tables are excluded from the prompt rather than being ranked low. 2. **Row-Level Security Awareness**: It works seamlessly with LangChain, LlamaIndex, and any other SQL agent, supporting Postgres, Oracle, MySQL, SQL Server, and SQLite. 3. **Prompt Token Reduction**: By filtering out irrelevant tables, it reduces the number of prompt tokens by 75%, making the interaction more efficient and less resource-intensive. ### Tech stack: Schemagate is built using Python, leveraging the power of LangChain and other SQL agents to provide a robust and secure text-to-SQL access control solution. ### Why it matters: Schemagate simplifies the process of accessing sensitive data by ensuring that only the necessary tables and data are included in the prompt, thereby reducing the risk of unauthorized access and improving the overall security of the system. This feature is particularly useful in environments where strict access control is required, such as in financial and healthcare applications. ### Repository: The repository is hosted on GitHub at [ashishsinha1602/schemagate](https://github.com/ashishsinha1602/s

Overview: Astideepspeech is a Golang binding for Mozilla's DeepSpeech library, enabling developers to integrate speech-to-text functionality into their applications. Key features: - Compatible with version `v0.9.0` of DeepSpeech - Provides Go bindings for Mozilla's DeepSpeech library - Includes pre-trained model and scorer files for efficient speech recognition Tech stack: Go Why it matters: Astideepspeech simplifies the integration of speech recognition capabilities into Go-based projects, making it easier to implement speech-to-text functionality in applications. Repository: asticode/go-astideepspeech GitHub stars: 180

### Overview: Cactus React Native is a package that allows developers to run AI locally in their React Native apps, enhancing the app's capabilities by integrating advanced language models. ### Key Features: - **Local AI Integration:** Cactus React Native enables developers to run AI models directly within their React Native apps, providing real-time language processing without the need for cloud-based services. - **Model Options:** Users can choose between different model quantizations and NPU acceleration options for improved accuracy and performance. - **Completion Generation:** The package supports generating text responses from the model by providing a conversation history. - **Vision Integration:** It includes a vision feature that allows passing images along with text prompts, enabling the model to analyze and understand visual content. ### Tech Stack: Cactus React Native is built on C++ and utilizes the React Native framework for seamless integration with existing React Native projects. ### Why It Matters: Cactus React Native simplifies the process of integrating advanced AI capabilities into React Native apps, making it easier for developers to leverage cutting-edge language models without the complexities of cloud-based solutions. This package is ideal for developers looking to enhance their apps with natural language processing and visual analysis features, improving user experience and application functionality.

### Overview: VietASR is an open-source Vietnamese Automatic Speech Recognition (ASR) project that utilizes a small, efficient model to transcribe Vietnamese speech into text. The project is built using PyTorch and includes pretrained models trained on a large Vietnamese speech dataset, including YouTube, radio, call center, and public datasets. ### Key Features: - **Model**: A small model (13M parameters) that is very fast for inference. - **Pretrained Model**: Trained on a large Vietnamese speech dataset, including YouTube, radio, call center, and public datasets. - **Tech Stack**: PyTorch, QuartzNet model, and KenLM for language modeling. - **Why It Matters**: The project provides a fast and efficient solution for Vietnamese ASR, making it suitable for real-time applications and small devices. ### Tech Stack: PyTorch ### Why It Matters: The project offers a fast and efficient solution for Vietnamese ASR, making it suitable for real-time applications and small devices. ### Installation: - **Linux Libraries**: Update and install `libsndfile1` and `ffmpeg`. - **Python**: Install Python 3.8 or higher. - **Libraries**: Install PyTorch, PyTorch Text, and KenLM. - **Requirements**: Install the project's requirements. ### Transcribe Audio File: ```bash python infer.py audio_samples ``` ### Run Web Application: ```bash python app.py ``` ### Video Demo: - **v1**: https://youtu.be/P3mhEngL1us -

Overview: This repository contains a collection of Generative AI examples specifically tailored for iOS development, showcasing various applications of Generative AI on the platform. Key features: - Text chat - Image understanding - Video summarization - Realtime video understanding Tech stack: Swift Why it matters: This project provides practical examples of how Generative AI can be integrated into iOS applications, enhancing user experiences and providing valuable insights through text, images, and video. Repository: shu223/iOS-GenAI-Sampler Primary language: Swift GitHub stars: 180 GitHub description: A collection of Generative AI examples on iOS. [](https://github.com/shu223/iOS-GenAI-Sampler) [](https://github.com/sponsors/shu223) [](https://www.buyme

### Overview: DeepInception is a lightweight method designed to hypnotize large language models (LLMs) into becoming jailbreakers, thereby unlocking their misuse risks. This approach leverages the personification ability of LLMs to construct a novel nested scene, enabling adaptive escape in normal scenarios and facilitating further direct jailbreaks. ### Key Features: - **Hypnotize LLMs**: DeepInception can easily make LLMs jailbreak, enhancing their misuse risks. - **Nested Scene Construction**: By personifying LLMs, it creates a novel nested scene to escape usage control. - **Continuous Jailbreak**: Achieves competitive jailbreak success rates and can perform jailbreaks continuously. ### Tech Stack: Python ### Why It Matters: Despite the remarkable success of LLMs in various applications, they are vulnerable to adversarial jailbreaks. DeepInception provides a practical and effective method to make LLMs jailbreak, highlighting the critical weakness of self-losing on open-source LLMs like Falcon, Vicuna, Llama-2, and GPT-3.5/4/4V. This method emphasizes the importance of paying attention to the safety aspects of LLMs and strengthening defense against misuse risks. ### Repository: Repository: tmlr-group/DeepInception Primary Language: Python GitHub Stars: 180 ### Getting Started: To use DeepInception, ensure you have PyTorch >= 1.10 with GPU support installed. Install the required dependencies by running: ```bash pip install -

### Overview: This repository is a reference implementation for ViVid-1-to-3, a novel approach that combines video diffusion with novel-view synthesis diffusion models to enhance pose and appearance consistency in video synthesis. ### Key features: - Novel View Synthesis - Video Diffusion Models - Pose and Appearance Consistency ### Tech stack: Python ### Why it matters: ViVid-1-to-3 significantly improves the quality and realism of video synthesis by addressing pose and appearance inconsistencies, making it a valuable tool for applications requiring high-fidelity video generation.

### Overview: This repository is a curated list of Multimodal Urban Computing models and datasets. It serves as a valuable resource for researchers and practitioners interested in the field of urban computing, focusing on the integration of diverse data sources and modalities. ### Key Features: - Comprehensive coverage of Multimodal Urban Computing models and datasets. - Updated regularly to ensure relevance and accuracy. - Includes a taxonomy framework for data perspective and methodology categorization. - A taxonomy and summary of open-sourced datasets for easy access. - A highly related paper list for further reading. ### Tech Stack: Python, Jupyter Notebook, Markdown, and GitHub. ### Why It Matters: This repository is essential for anyone looking to understand and advance the field of Multimodal Urban Computing. It provides a structured overview of the latest advancements and datasets, making it a valuable resource for researchers and practitioners alike. ### Repository: Repository: yoshall/Awesome-Multimodal-Urban-Computing GitHub Stars: 180 GitHub Description: A professional list on Multi-modal Data Fusion Models and Key Datasets for Urban Computing.

### Overview: This repository is a comprehensive guide to understanding the layers of AI, focusing on Agentic AI, Generative AI, and LLMs. It provides detailed explanations of AI agents, AI workflows, and the distinctions between various AI methodologies. ### Key Features: - Detailed explanations of AI layers - Comprehensive roadmap for AI agents - Comparison of AI methodologies (AI, AGI, AGI, ML, DS, LLM, Cognitive Architectures) - Illustrations of AI workflows ### Tech Stack: - None specified ### Why It Matters: This repository serves as a valuable resource for developers and engineers looking to deepen their understanding of AI engineering, focusing on the practical applications and distinctions between different AI paradigms.

### Overview: Thesys is an open-source research workspace designed for efficient searching, citation management, and interactive research. It integrates with various academic databases, supports document highlights, and enables collaborative Q&A sessions. The system leverages local processing for enhanced performance and seamless integration with the OpenAI API for advanced comparison features. ### Key Features: - **Chat over Research Papers:** Answers citations with direct references from the document. - **Reader Mode:** Full and focused highlights for better context. - **Document Summaries:** Structured summaries of papers. - **Figure and Diagram Support:** Ingests and highlights figures and diagrams. - **Pairwise Paper Comparison:** Classifies pairs as corroboratory, contradictory, or neutral. - **Discover Papers:** Searches open-access literature from OpenAlex and arXiv. - **Library and Export:** Manages bibliographies and exports citations in multiple formats. - **Local Setup:** Runs on Python 3.11+ and Node.js 20+ with an OpenAI API key. ### Tech Stack: - **Frontend:** React (Vite) - **Backend:** FastAPI - **Local RAG:** Docling + Chroma + BGE - **OCR:** Off by default; set `DO_OCR=true` for scanned PDFs - **Chat:** OpenAI via LangChain - **Compare:** TypeSafe Jev (`TYPESAFE_API_KEY`) - **Paper Search:** OpenAlex + arXiv (optional OpenAlex API key) - **Citations:** CiteAs (uses `EMAIL`) ### Why

Overview: TimeSeriesScientist is an LLM-driven agentic framework designed for general time series forecasting, comprising four specialized agents: Curator, Planner, Forecaster, and Reporter. Key features: - Performs LLM-guided diagnostics with external tools for targeted preprocessing. - Narrows the hypothesis space of model choice using multi-modal diagnostics and self-planning. - Performs model fitting and validation, and adaptsively selects the best model configuration and ensemble strategy for final predictions. - Synthesizes the forecasting workflow into a comprehensive, transparent report. Tech stack: Python Why it matters: TimeSeriesScientist simplifies the complex process of time series forecasting, making it more interpretable and extensible across various tasks.

### Overview: tiny-audio is a speech-to-text system that combines a frozen speech encoder, a small Qwen LLM (large language model), and a small trained projector. It is designed to be simple and easy to understand, making it perfect for learning purposes. The system is trained on a single GPU and can be used for understanding how different components work together. ### Key features: - A frozen speech encoder - A small Qwen LLM - A small trained projector - Word timestamps - Speaker diarization - A batched HTTP server - A live demo - A free course ### Tech stack: Python ### Why it matters: tiny-audio is a personal side project that aims to demystify the components of a speech-to-text system. It is designed for understanding and learning rather than for production use. The system is trained on LibriSpeech and can be used for both learning and experimenting with speech recognition.

### Overview: This project introduces **DiffMVS** and **CasDiffMVS**, two methods for efficient and accurate multi-view stereo reconstruction. Utilizing a diffusion model, these techniques aim to strike a balance between accuracy and computational efficiency, providing new strong baselines for future research in MVS. ### Key Features: - **DiffMVS**: A lightweight and accurate method for multi-view stereo reconstruction using a diffusion model. - **CasDiffMVS**: A variant of DiffMVS that improves performance and efficiency by refining a coarse depth map with a diffusion model. - **Lightweight Network**: A lightweight diffusion network designed to enhance both performance and efficiency. ### Tech Stack: - Python ### Why It Matters: These methods offer a significant improvement in both accuracy and efficiency compared to existing techniques, making them promising candidates for future research in multi-view stereo reconstruction.

### Overview: Captain Claw is a self-hosted framework for orchestrating fleets of specialist AI agents, enabling teams to command and control multiple AI agents seamlessly. It supports seven orchestration modes, allows agents to read and write to each other's memory, and integrates with major LLM providers or operates entirely locally with Ollama. ### Key Features: - **Multi-Agent Command Center:** A multi-agent command center where you can spawn specialist teams and manage them. - **Orchestration Modes:** Seven different modes for orchestrating agents, including collaborative blackboards and independent review. - **Orchestrated Collaboration:** Agents can communicate with each other and share information, making it easier to collaborate and coordinate tasks. - **Code Pipeline:** A full coding pipeline that includes planning, building, and independent review, ensuring that the work is done correctly. - **Team Safety:** A team-safe deck and kiosk, as well as a team-friendly admin console, allowing non-admins to manage the team. - **Local and Remote Deployment:** The framework supports local deployment with Ollama and remote deployment with major LLM providers. - **Integration with Multiple Tools:** Connects to any system, including WhatsApp, and integrates with multiple tools and platforms. ### Tech Stack: Python ### Why It Matters: Captain Claw provides a powerful and flexible solution for managing and orchestrating AI agents, making it easier to collaborate and coordinate tasks. Its multi-agent command center and orchestration modes make it ideal for engineering teams looking to streamline their workflow and increase productivity.

### Overview: This repository contains code for papers and projects published by the Morgan Stanley Machine Learning Research team. It serves as a central hub for machine learning research and publications, showcasing cutting-edge techniques and methodologies in various domains such as time series analysis, recommender systems, network theory, and more. ### Key Features: - **Code Repositories**: Contains implementations of various machine learning models and algorithms. - **Research Papers**: List of papers published by the team, including titles, venues, and links to the corresponding code. - **Collaboration with Academia**: The team collaborates with universities, contributing to top-tier conferences and publications. ### Tech Stack: Jupyter Notebook ### Why it Matters: The repository is crucial for researchers and practitioners interested in machine learning research at Morgan Stanley. It provides a platform to access and experiment with cutting-edge techniques, contributing to the advancement of the field.

### Overview: This project implements a framework to convert speech to facial features as described in the CVPR 2019 paper - "Speech2Face: Learning the Face Behind a Voice" by MIT CSAIL. The project includes a detailed report and usage instructions for converting audio and video to facial features, with an emphasis on improving face retrieval performance. ### Key Features: - Speech to Facial Features Conversion - Efficient Database Structure - Pretrained Models and Training Options - Results and Evaluation Metrics ### Tech Stack: Python ### Why It Matters: This project contributes to the field of automatic speech recognition and facial feature synthesis, offering a practical application of machine learning techniques in audio-to-image conversion.

### Overview: YantrikDB is a cognitive memory database designed for AI agents, consolidating, detecting contradictions, and fading stale memories over time. It provides a short path to try it with a single `pip install` and a minimal configuration. ### Key features: - **Forgetfulness**: YantrikDB forgets old memories, reducing noise in the agent's memory. - **Consolidation**: It consolidates similar memories into fewer, more meaningful ones. - **Conflict Detection**: It detects and resolves contradictions in the agent's knowledge. ### Tech stack: Rust Rust is the primary language used for developing YantrikDB. ### Why it matters: YantrikDB addresses the limitations of traditional vector databases by providing a more nuanced and context-aware memory management system. It helps AI agents maintain a clean and relevant memory, reducing noise and improving recall precision. This makes it an essential tool for building more effective AI systems.

### Overview: GenAI Factory is a collection of end-to-end blueprints designed for deploying generative AI infrastructures in Google Cloud Platform (GCP), following best practices for security and infrastructure automation. ### Key Features: - **Infrastructure as Code (IaC):** Utilizes Terraform for implementing infrastructure, leveraging Google Cloud's Terraform resources and modules. - **Shared VPCs:** Leverages native shared VPCs for efficient resource management. - **Least Privilege Principle:** Ensures minimal permissions and no default service accounts. - **Cloud Foundation Fabric Compatibility:** Works with Cloud Foundation Fabric v59.2.0 and is compatible with project-factory and application templates. - **Centralized Deployment:** Supports deployment of Gemini Enterprise applications and datastores, custom MCP servers, and agents via Agent Registry and Agent Gateway. - **Agent Runtime:** Provides a private access to VPC resources, enabling secure interactions with Internet via SWP HTTP Proxy. - **Secure Cloud Run Deployments:** Deploy secure Cloud Run applications for interacting with Gemini, running ADK agents, or a self-hosted Gemma 3 model with Nvidia L4 GPUs. - **Private Model Deployment:** Privately deploy open models from Vertex AI Model Garden over Private Service Connect (PSC) and expose them through a secure Cloud Run application. - **Natural Language to SQL (NL2SQL):** A secure agent on Cloud Run that allows users to query data from BigQuery using natural language. - **RAG with Cloud Run and CloudSQL:** A "Retrieval-Augmented Generation" (RAG

### Overview: xFinder is a large language model designed to automate the evaluation of LLM responses, providing a reliable and accurate way to extract key answers. This project has been accepted to ICLR 2025 and is now available as a PyPI package, making it easy for developers to integrate into their evaluation frameworks. ### Key Features: - **Automated Evaluation**: xFinder leverages large language models to extract key answers from LLM responses, ensuring high accuracy and reliability. - **Robustness and Reliability**: The model is designed to handle various types of LLM outputs, making it suitable for diverse evaluation tasks. - **Integration with Frameworks**: xFinder can be seamlessly integrated into existing evaluation frameworks, enhancing their robustness and reliability. ### Tech Stack: - Python ### Why It Matters: xFinder offers a significant improvement in the evaluation process of large language models by automating the extraction of key answers, thereby reducing the reliance on traditional RegEx methods. This makes it easier to compare and validate model performance in a more reliable and meaningful way.