The idea
Text adventures live or die on how fast the story responds to you. A three-second wait after every choice kills immersion. AI Dungeon Master was built around low-latency streaming narration first - everything else was designed around keeping that fast.
Narration
Story text streams token-by-token over Server-Sent Events, powered by the Groq API running Llama 3.3 70B - Groq's inference speed is what makes the narration feel closer to a live dungeon master typing in real time than a chatbot generating a paragraph and dumping it all at once.
Seeing the real world
The more unusual piece: the game can use your camera to recognize physical objects around you via Llama 4 Scout 17B, then weave whatever it sees into the fantasy narrative. Point it at a coffee mug mid-game and the story can turn it into an ancient relic - the world reacts to what's actually in front of you, not just what you type.
Making it visual and sayable
Each action generates a matching AI scene image through Hugging Face's FLUX.1-schnell, so the adventure isn't just text - it has a visual record as you go. Voice input via the Web Speech API means you can speak your actions instead of typing them, which matters more than it sounds like for keeping the pace of a live session.
Taking it with you
At the end of a session, the whole playthrough - narration and generated scenes - exports as a storybook PDF via jsPDF, so a run through the game becomes something you actually keep, not something that disappears when you close the tab.
What I'd build next
Persistent world state across sessions - right now each playthrough is self-contained; carrying consequences and relationships forward into a new session is the obvious next layer.