Skip to Main Content

Calvin

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Calvin — Demo video

[CHECK LAST LINE FOR JUDGE ACCESS CODE] Kinetext turns a page a teacher already trusts into a short, interactive lesson they can actually use in class, and shows the checks behind every part of it. I'm a teacher. Good source material is everywhere (NASA, NOAA, USGS, NIST, MedlinePlus), but turning a page into something students do, not just read, takes hours. AI can do it in a minute, but a teacher can't hand students something that might be wrong. Kinetext reads a licensed page with Meta's Muse Spark, writes a summary whose key points light up the sentence they came from, and builds 2 to 3 interactive parts: label the diagram, sort, sequence, quiz, formula explorer, timelines, simulations and more. Then deterministic code checks every quote, number and formula against the page. When a check fails, the part goes back to Muse with that check and Muse revises it; a part that still fails is left out instead of shipped wrong. For diagrams, Muse Spark locates each labelled part on the page's own figure and SAM 3.1 outlines it to confirm. A teacher reviews each part and publishes the lesson to their shelf. It runs on Muse Spark 1.3 (Meta Model API) and Muse Spark Open (open weights on Modal). In a pre-registered evaluation of 194 lessons per model, Spark 1.3 built a working lesson 189 times and Spark Open 178 times. Teachers can also paste their own page and, in beta, their own diagram. Run it: clone the repo, npm install, npm start (Node 24). Judge access code for live builds (capped at USD 50 in total): . Without the code, "Explore read-only" shows every recorded lesson, the Demo and Results. Access code below: aDIQXgXOUBf6ktMSW2vlsIV-gBXQquaz

Team