Curriculum

Beyond text

Multimodal is 34 percent (47 companies), which the report notes is broader than several topics that get far more course time, and it is a named gap: nearly absent across the field, with only Product Faculty's Sense layer touching it and no

1. 11.01

   Documents are not text

   Choose between native vision and an OCR-then-text pipeline for documents, and separate a layout or parsing failure from a real reasoning failure so you stop blaming the model for the wrong layer.
2. 11.02

   Voice

   Decide between a cascaded transcribe-generate-speak stack and a native speech-to-speech model, price the audio bill that runs an order of magnitude above text, and design a spoken turn, its barge-in, and its repair, with no screen to fall back on.
3. 11.03

   Image and video generation

   Reason about diffusion step count, temporal consistency, and character drift as product levers, and treat provenance, recording consent, and likeness rights as duties the modality imposes before you ship it.
4. 11.04

   Ship a language

   Commit to one non-English locale, build its own golden set instead of a machine-translated English one, and retest the LLM judge and safety classifier that quietly degrade outside English.
5. 11.05

   Build a multimodal golden set

   Assemble the golden set that scores audio on audio and documents on layout, seeding the real degraded distribution, so the text evals from your core suite stop passing failures they were never able to see.
6. 11.06

   INTEGRATION CASE

   When you add voice or another language, you can write a readiness pack with native cases and local grader checks.