MultiModalSyntheticDataGenerator

Inspiration

One of the key challenge to accelerate innovation and rapid adoption of AI in enterprises is generating multi modal synthetic data. This project implements an agent that creates muliti modal synthetic data, tabular data and images. This is implemented as multi agent, each agent taking care of one modality.

What it does

Generates multi modal synthetic data that can be used as test data for software development or used in data augmentation for ML applications

How we built it

For building tabular data, used a third party tool sdv.dev. For images, tried out both NVIDIA/Hugging Face pipeline and Gemini model

Challenges we ran into

I am exploring how to pass the objects back and forth especially images. I think persistent service will solve, but on the exploration path.

Accomplishments that we're proud of

Understood the architecture for ADK and how to built an agent. Very confident of designing and implementing multi agents. Implemented an end to end multi agent in collab environment

What we learned

Necessary tool set up for interactive environment such as collab, the run time architecture and how to approach implementing multi agents

What's next for MultiModalSyntheticDataGenerator

A lot, deployment in Vertex so it can be enterprise grade, add more modalities and most importantly connect them and explore how this can be used in healthcare

Built With

• adk
• collab
• python