Multi-Agent Hyper-Segmentation Engine

A multi-agent AI system that crafts unique, hyper-personalized ads for every single user in real-time, powered by Google ADK.

Comment

Inspiration

The digital advertising space is saturated with generic, one-size-fits-all content. We saw a massive opportunity to challenge this norm. Our inspiration was to create a system that doesn't just segment users into broad categories, but achieves true 1-to-1 hyper-personalization. We asked ourselves: "What if every single ad a user sees is created uniquely for them, in that exact moment?" This question drove us to design an intelligent, autonomous system that could make advertising personal again.

What it does

Our project, the Multi-Agent Hyper-Segmentation Engine, is a fully autonomous system that crafts unique, hyper-personalized ads in real-time. Here's how it works:

  1. Data Ingestion: When a user interacts with our demo e-commerce site, their actions (clicks, add-to-carts, etc.) are streamed into Firestore and BigQuery.
  2. Agent Activation: Triggered by a Cloud Scheduler (or manually for the demo), a Cloud Run service activates our multi-agent team, built with the Google Agent Development Kit (ADK).
  3. The Agent Team:
    • Analyst Agent: Queries BigQuery to analyze the user's latest behavior and determines their segment (e.g., "User is in Istanbul and interested in electronics").
    • Creative Agent: Uses the strategy to generate a brand-new, personalized ad image via a generative AI model.
  4. Delivery: The new ad is saved to Cloud Storage, and the user's profile in Firestore is updated with the new ad URL. The next time the user refreshes the page, they see an ad made just for them.

How we built it

This project is built almost entirely on the Google Cloud ecosystem, leveraging the strengths of each service to create a seamless pipeline:

  • Core Logic: Google Cloud Run hosts our serverless backend, providing the scalable, event-driven environment needed to run our agent team on demand.
  • Intelligence: The Google Agent Development Kit (ADK) was the foundation for creating our specialized Analyst, Strategy, and Creative agents.
  • Data & Analytics: Google BigQuery handles the large-scale analysis of user event data, while Google Firestore stores real-time user profiles and the URLs for their personalized ads.
  • Asset Management: Google Cloud Storage is used to store the millions of potential ad creatives generated by our system.
  • Frontend: A demo e-commerce website was built with Next.js and React to showcase the system in action.

Challenges we ran into

Building a multi-agent system from scratch came with its own set of unique challenges. Orchestrating the communication and data handoff between the Analyst, Strategy, and Creative agents required careful design to ensure the pipeline was robust and error-free. Furthermore, prompt engineering for the Creative Agent was a significant task; we had to fine-tune our prompts to ensure the generated ads were not only relevant but also high-quality and on-brand.

Accomplishments that we're proud of

We are incredibly proud of creating a fully autonomous, end-to-end pipeline that demonstrates true hyper-personalization. Seeing the system correctly identify a user's context (like their location) and generate a geographically relevant ad (like the Bosphorus ad for Istanbul) was a huge success. This project proves that a multi-agent system, powered by Google Cloud, can solve complex, real-world marketing challenges effectively.

What we learned

This hackathon was an incredible learning experience. We gained deep, practical knowledge of the Google ADK and learned how to structure and deploy a sophisticated multi-agent system on Cloud Run. Integrating multiple Google Cloud services like BigQuery, Firestore, and Cloud Storage into a single, cohesive application taught us a great deal about building scalable, cloud-native solutions.

What's next

The potential for this engine is limitless. In the future, we plan to expand the agent team with new roles, such as an A/B Testing Agent to optimize ad performance and a Budgeting Agent to manage campaign spending. We also aim to integrate with real-time bidding platforms to deploy our hyper-personalized ads across the web.

Built With

Share this project:

Updates