1 of 10

IDEATE

Multi-Agent AI Instructional Design System

Team: Blake Hewitt, Craig Huang, Nidhi Sharma, Rhys Jones, Sanjana Venkatesh, Tarun Singh B

Mentor: Nancy Belmont (MARi)

2 of 10

Context and Background

Most training and educational materials exist as static PDFs, slides, or notes, making them time-consuming to update and difficult to personalize.
Existing AI tools generate content but lack quality control, grounding, and learner personalization, often producing unreliable or generic output.
Organizations need a faster, scalable way to transform raw materials into accurate, engaging, adaptive courseware.

3 of 10

IDEATE: Goals and Target Users

IDEATE automatically converts uploaded educational materials into structured, interactive learning modules, including lessons and quizzes. Uses a multi-agent AI workflow and integrates Human-in-the-loop (HITL) review to ensure accuracy, consistency, and personalization.

Designed primarily for students, enabling them to turn their static course material into personalized, self-paced learning modules.

4 of 10

IDEATE: Functions and Features

Main Functions

Upload syllabus/notes → automatic parsing and blueprint creation
Lesson and quiz generation tailored to learner preferences
Evaluation agent ensures correctness, clarity, and alignment
Revision agent improves content iteratively

Unique Features

Personalization First: Adapts lessons and quizzes to learning style, tone, complexity, and user context — the core feature that sets our system apart.
Multi-Agent Workflow: Blueprint → Lesson → Quiz → Evaluation → Revision
Event-Driven Cloud Architecture: Scalable and modular
Added prompt engineering techniques to mitigate hallucinations and copyrighting

Sanjana

Everything begins when a user uploads a syllabus or set of notes for a subject they’re studying. Before that, on the homepage, users can set their preferred learning styles, complexity level, and tone. These preferences are stored in their profile and automatically referenced whenever they upload a document. Once the file is uploaded, the system parses it and generates a structured blueprint. From there, our agents create full lessons and quizzes once the blueprint is approved.� To maintain high-quality output, an evaluation agent checks each lesson for accuracy, clarity, and alignment with learning objectives. If issues are found, a revision agent automatically improves the content, making it stronger with each iteration.

One of the most important aspect of our system—and what truly makes it unique—is its level of personalization. The platform adapts the style, tone, and difficulty to each individual user, creating a much more effective and meaningful learning experience.

Supporting this is a multi-agent workflow that moves seamlessly from blueprint creation to lesson generation, quiz generation, evaluation, and revision. The system runs on an event-driven cloud architecture that keeps everything scalable and modular, and we use targeted prompt-engineering techniques to reduce hallucinations and avoid copyright issues, ensuring that the content remains reliable and safe

5 of 10

Development: IDE and Tools

Development Tools

Deployment Tools

Nidhi�On this slide, I’ll go over the core tech stack we used throughout the project.�For development, we primarily worked with VS Code, Python, React, and Docker. Python allowed us to build each AI agent cleanly and consistently, while React gave us the flexibility to design a frontend that updates in real time as the pipeline progresses. Docker was essential for containerizing our services so they behaved the same in every environment.

What really shaped our system, though, is that the entire project was built completely on Google Cloud. We relied on several key tools to support our event-driven architecture.

Google Cloud Storage handled all user uploads, like PDFs and slides, and made them easily accessible to each agent.
Cloud Run deployed each agent as a separate microservice, giving us automatic scaling and clean isolation between components.
Firestore acted as our lightweight state manager, allowing the frontend to instantly reflect where a user’s workflow was in the multi-agent pipeline.
And Eventarc helped connect everything by triggering agents at the right moments, which kept the workflow moving smoothly without manual coordination.

Together, these tools made it possible for us to build a modular, scalable, and reliable multi-agent system entirely in the cloud.

6 of 10

Technical Details & Design Choices

System Architecture

Event-driven, multi-agent pipeline:�File Upload → Parsing → Blueprint Agent → Lesson Agent → Quiz Agent → Evaluation Agent → Revision Agent
Modular microservice architecture allows each agent to be improved independently.

Design Choices

Chunked Parsing: Using Unstructured to create labeled text chunks for grounding → reduces hallucinations.
JSON/Markdown Prompting: Guarantees schema compliance for blueprints, lessons, and quizzes.
Evaluation + Revision Loop: Multi-agent scoring system improves clarity and correctness.
Firestore as State Manager: Lightweight, real-time orchestration visibility for the frontend.
GCS for Large Files: Avoids Firestore storage limits and improves retrieval performance.

Why Multi-Agent?

Allows separation of concerns (blueprint ≠ content ≠ quizzes ≠ evaluation).
Enables iterative refinement, improving quality more than single-step generation.
Enhances system reliability and makes debugging easier.

7 of 10

Implementation Adversity

Pipeline timing & asynchronous behavior

Lesson page was loading before evaluations were complete, lesson page was fetched too early
Initial cross-service integration issues

Handling copyright constraints meant adjusting our prompts to understand these constraints and use open-source resources

8 of 10

Lessons Learned

Deploying a Google Cloud Native Project
Utilizing AI Libraries/Tools
Project Development Lifecycle

9 of 10

Future Work

Incorporate H5P-style interactive content to enrich lesson delivery.
Implement deeper reference validation to prevent further hallucinations
Experiment with speed of production to see if entire workflow can be sped up resulting in quicker content generation.

10 of 10

Demo