The problem
It started on a long flight. I was playing the in-flight trivia game and noticed the same questions coming back. They were hard-coded, a fixed list someone wrote once. I started wondering: what if a language model could generate questions endlessly, so the game never ran dry?
Doom Trivia is the answer. Questions come from an AI generation pipeline organized around a taxonomy of 40 topics and 2,000 subtopics, so the model explores the long tail of knowledge instead of repeating the same famous facts.
What it does
- Swipe to answer. Drag the question card to an answer corner.
- Six game modes. Simple, Tough, Time Attack, Flags Easy, Flags Hard, and Country Quiz.
- Streaks with a mascot. A lantern character reacts as your streak grows.
- Day and night themes.
- Play your way. Guest play or Google/email sign-in, with save and load.

How it works
The question pipeline starts from a taxonomy of 40 topics split into 2,000 subtopics. Generating per subtopic pushes the model away from the handful of famous facts it would otherwise repeat. Curated questions are stored in Supabase and served to the Flutter app.
How generated questions are checked
The redesign
The first build was redesigned into a hand-directed woodland world.

Earlier buildCurrent build

What was hard and how I solved it
Results
- curated questions shipped (321 history, 183 science, 133 flags)
- 637
- curated questions shipped (321 history, 183 science, 133 flags)
- topics and subtopics in the taxonomy
- 40 / 2,000
- topics and subtopics in the taxonomy
- passing tests (reported 2026-09-13)
- 51
- passing tests (reported 2026-09-13)
- game modes
- 6
- game modes

Stack
- Flutter
- Supabase
- LLM question-generation pipeline
