NovelViz
An AI reading companion that only ever knows what you've read. Ask questions, explore characters, generate imagery from the book in your hands — every chapter, alive, with zero spoilers.
The problem
Halfway through a novel, you forget who a minor character is, or want to picture a place the author describes. So you search the internet — and the internet knows the ending. One glance at a wiki and three hundred pages of tension are gone.
General chatbots are no safer: ask an LLM about chapter twelve and it will happily volunteer what happens in chapter thirty. Spoiler safety can't be a politeness request — it has to be an architectural guarantee.
The core idea: context gating
NovelViz tracks your reading position and only ever feeds the model text up to your bookmark. The AI isn't asked to avoid spoilers — it simply has never seen them. That's the same discipline as scoping data access in a production pipeline: don't trust the model to behave, design the system so it can't misbehave.
Everything downstream inherits the guarantee: answers, character profiles, and generated imagery are all built exclusively from the pages you've already turned.
What you can do
Question the book
"Who is this again?" "What did I miss about the locket?" — answered from your pages only, like a friend reading over your shoulder who's exactly as far in as you are.
Generate imagery
Turn the scene you just read into an image — characters, places and moments rendered from the book's own descriptions, up to your bookmark and no further.
A living library
A public-domain catalogue ingested and enriched chapter by chapter, plus a public gallery of scenes other readers have brought to life — spoiler-gated to each visitor.
Built like a product
Auth, quotas, AI-failure protection, moderation and cost controls — the unglamorous 80% that separates a working product from a demo.
Under the hood
A Next.js app with position-aware context assembly feeding Claude and OpenAI models, image generation via fal.ai, a Prisma / Neon Postgres data layer with a full ingestion pipeline for public-domain texts, Clerk auth and Cloudinary media storage.