The idea
Good exam practice needs a steady supply of good questions, and there are never enough. SparkRoute turns the questions that already exist — in official test papers and recorded walkthroughs — into a growing, structured bank, then generates new ones from it, each checked for correctness and explained in plain language for both student and parent.
How it works
The pipeline starts with ingestion. A video stage detects question boundaries in screen-recorded tests by frame-differencing, then uses Claude Vision to pull the stem and options and auto-crop figures; a PDF stage scans papers, extracts questions, and tags concept and difficulty. From there, generation embeds the question stems, retrieves similar ones, drafts new candidates with Claude, and runs them past a critic model that filters and de-duplicates — so a generated question is verified before a student ever sees it. The seed bank covers reading, maths, and thinking-skills questions drawn from official papers.
Architecture
A Turborepo monorepo: a Next.js web app with role-based student, parent, and admin views, a Python ingestion-and-generation pipeline, and Supabase for Postgres, auth, and storage. Claude does the vision and generation work, with embeddings for retrieval and ffmpeg behind the video stage.
Status
Live in staging, with the full ingestion-and-generation pipeline validated end to end. The product surfaces for students and parents are coming together on top of it.