Quality assurance · Production

Day of AI Australia scales AI literacy simulation to 330,000+ students using Google Cloud

The problem

Day of AI Australia needed to give students the ability to direct generative AI bots as part of an AI literacy game, but this posed a considerable risk: without strict controls, bots could inadvertently generate hate speech or age-inappropriate content in violation of Australian privacy laws and education standards. The lean engineering team also faced the challenge using fragmented tools from multiple providers that proved inefficient.

First attempt

Relying on fragmented tools from multiple providers for billing, inference, and security monitoring proved inefficient for the lean non-profit team.

Workflow diagram · grounded in source
1
Student deploys bot
trigger
“when a student deploys a bot—configuring its personality and tactics—the system initiates a dynamic workflow on Google Cloud that generates unique content based on the unfolding election narrative”
2
Vertex AI processes bot interactions
ai_action
“Every 10 minutes, the system processes bot interactions, utilizing Vertex AI to manage inference requests”
3
Gemma generates social posts
ai_action
“open Gemma models generate hundreds of thousands of unique social media posts”
4
Shield Gemma validates content
validation
“every post passes through Shield Gemma, which evaluates content against strict policies, blocking harmful output in real time”
5
Posts reach Barn Wall feed
output
“Before reaching the student-facing "Barn Wall" feed, every post passes through Shield Gemma”
6
Cloud SQL stores game state
integration
“Cloud SQL manages encrypted game states and teacher logs without ever storing student PII, ensuring strict compliance with Australian education standards”
Reported outcome

The platform powered the 'Win the Farm' competition without a single performance issue, generated more than 120,000 unique social media posts without a safety incident, blocked 100% of harmful content, and reached more than 330,000 students including those in remote schools with limited hardware.

Reported metrics
Students supported to dateover 330,000
Harmful content blocked100%
safe AI social posts generated120,000+
Deployment timeunder four months
Show all 7 reported metrics
students supported to dateover 330,000
harmful content blocked100%
safe AI social posts generated120,000+
deployment timeunder four months
concurrent student teams supported350+
performance issues during competitionwithout a single performance issue
safety incidents during post generationwithout a safety incident
Reported stack
Compute EngineVertex AIGemmaShield GemmaCloud SQLModel Garden on Vertex AIGemini in Vertex AI
Source
https://cloud.google.com/customers/unswdayofaiaustralia
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The platform powered the 'Win the Farm' competition without a single performance issue, generated more than 120,000 unique social media posts without a safety incident, blocked 100% of harmful content, and reached mor…

What tools did this team use?

Compute Engine, Vertex AI, Gemma, Shield Gemma, Cloud SQL, Model Garden on Vertex AI, Gemini in Vertex AI.

What results were reported?

Students supported to date: over 330,000; Harmful content blocked: 100%; safe AI social posts generated: 120,000+; Deployment time: under four months (source-reported, not independently verified).

What failed first in this deployment?

Relying on fragmented tools from multiple providers for billing, inference, and security monitoring proved inefficient for the lean non-profit team.

How is this quality assurance AI workflow structured?

Student deploys bot → Vertex AI processes bot interactions → Gemma generates social posts → Shield Gemma validates content → Posts reach Barn Wall feed → Cloud SQL stores game state.

Related quality assurance cases