Every game Tempora Studio publishes passes through several language models, each picked for the job it does best. A unified, model-agnostic API lets us route each step to the right model, compare quality and cost on real output, and fall back automatically so the 06:00 edition never misses.
Pipeline stepModel classWhy
Extract entities, facts and dates from articlesfast · low-costHigh volume, structured output; runs on every article in the feed.
Write clues and questions in the house voicefrontier · creativeWordplay, tone and difficulty control decide whether a puzzle is fun.
Fact-check each clue against its sourcereasoning · independent providerA different model family catches errors the author model would repeat.
Rebuild puzzles for another languagemultilingualNative-quality Turkish and English wordplay, not literal translation.
Moderate player names and feedbacksmall · safety-tunedKeeps leaderboards and comments brand-safe at near-zero cost.
Editorial insight summarieslong-contextReads a week of play data and coverage to brief the newsroom.
Voice the daily quiz as an audio editiontext-to-speechListen-and-answer quizzes for smart speakers, podcasts and readers who prefer audio.
Grounded, never invented
Every clue cites the publisher's own article. No source, no question.
Deterministic where it matters
Models write; code validates grids, answers and schedules before anything ships.
Provider-independent
Swap or combine models as quality and pricing change, with no rewrite and no lock-in.