Q&A

Lessons Learned Building a GenAI-Powered App

• By David Ramel
• 09/18/2024

Sometimes, complex technical achievements are best explained through one example. That's the approach Mete Atamel, Developer Advocate at Google, is taking as he makes the rounds detailing the capabilities of Vertex AI and associated tooling on the Google Cloud Platform.

Atamel will be sharing his insights in an upcoming session titled "Lessons Learned Building a GenAI Powered App" at the big, multi-conference Live! 360 developer/IT event being held in Orlando starting Nov. 17.

In his intermediate-level session, Atamel will walk attendees through the process of building an interactive trivia quiz app that leverages generative AI to generate quiz questions and images on an infinite number of topics and languages.

VisualStudioMagazine: What inspired you to present on this topic?
Atamel: I always like to present topics on how a new technology helps unblock existing use cases or create new use cases with concrete examples. In this case, generative AI unblocked a trivia quiz app that's been sitting idle for a few years and I wanted to tell the story to others.

Can you give a brief overview of the GenAI-powered interactive trivia quiz app you built?
It is an interactive trivia quiz application that people play trivia quiz games against each other.

The initial version of the app was static with pre-set topics and was originally implemented by my co-worker Marc Cohen a few years ago. Last year, we expanded the app to generate quiz questions and images on "infinite" number of topics and languages using generative AI.

What were the main challenges you faced while developing this app?
The main challenge we faced with the app initially was the fact that it was static with a pre-set list of topics and questions on those topics and everything was in one language (English). Generative AI helped greatly to expand the quiz app to any topic and any language. However, it introduced other problems due to flakiness and inconsistencies of Large Language Model (LLM) outputs. Also, LLMs tend to hallucinate and that causes issues in a quiz app when you need accuracy. Lastly, it's hard to quantify the quality of LLM outputs but you need that if you're going to improve your prompts for better quizzes.

How did Generative AI services specifically help in overcoming these challenges?
With Generative AI, we were able to generate quizzes on any topic, in any language and even generate images for quizzes. This was a game changer. Later, we also used LLMs to measure accuracy of quizzes with moderate success.

Can you discuss some specific GenAI services used in your app and their roles?
We used Vertex AI on Google Cloud as our Generative AI platform. On Vertex AI, we initially used the PaLm model and later moved to various Gemini models for quiz generation. We also use the ImageGen model to generate images. All of these models were consumed by services running on Cloud Run, the serverless container service of Google Cloud and backed by Firebase NoSQL database.

How will attendees participate in the live demo?
I will generate a quiz on a certain topic and share a QR code. Attendees will scan that QR code and simply join the game!

What resources would you recommend for those looking to deepen their understanding of building GenAI-powered apps and prepare for your session?
My session does not require prior knowledge of Gemini or Vertex AI but participants can take a look at the Vertex AI documentation to get started: https://cloud.google.com/vertex-ai

Note: Those wishing to attend the conference can save hundreds of dollars by registering early, according to the event's pricing page. "Save $400 when you register by the Super Early Bird deadline of Sept. 27!" said the organizer of the event, which is presented by the parent company of Visual Studio Magazine.