Paper-Time
An AI tutor that interrupts itself: arXiv papers become narrated, animated explainers.
Research papers are the highest-signal, lowest-throughput learning material humans produce. Existing AI tools summarise papers; they don’t teach them. You skim a summary and pretend you understood the notation.
A 7-agent pipeline ingests an arXiv paper, plans visualizations, and generates Manim animation code. Four independent validators (AST, spatial bounds, narration alignment, import smoke) gate every animation before it renders, regenerating on failure, so compute is never spent on code that won’t run or overflows the frame. Equations are preserved end-to-end (ar5iv HTML → Manim MathTex → KaTeX), never paraphrased.
Built solo at the GOSIM Agentic Hackathon (STATION F, Paris). Renders a fresh paper end-to-end in ~2 minutes and ships a cached “Attention Is All You Need” run that plays instantly. Model-agnostic: one config string per agent swaps the whole stack to a different provider.
The generator is the cheap part; the verifier is the product. An LLM whose output is gated by checks it can’t talk its way past is a system you can ship. The persona layer and voice barge-in are next.