AI · Extended learning · 2026-10-10

Multimodal apps: text, images, audio, and async workflows

Understand input governance, asynchronous state, and cost controls for vision, speech, image generation, and video tasks.

Concept notes

Multimodal is not about stuffing images, audio, and text into a single interface at random. Every input type has governance concerns: images may contain faces, locations, or copyrighted material; Speech to Text (STT) and Text to Speech (TTS) can expose identity clues or create misleading audio; and generated images or video are usually asynchronous tasks that need status polling, failure handling, and cost notices. A learning repository can show API call shapes, but a production system must still handle consent, retention, auditing, content safety, and model availability changes. A static site should not embed secrets or pretend to offer live generation.

Tradeoffs and safety exercise

Paper/local safety exercise: write a process card for “upload a product image and generate a description.” Include the pre-upload notice, file-size limit, malware scanning, vision analysis, human review, result retention, and delete request handling. Then write a state machine for an async video task: queued, running, succeeded, failed, expired. This exercise helps you treat multimodal systems as product workflows instead of one magical call.

Completion checklist

• Can list the privacy and copyright risks of multimodal inputs.
• Can explain why async task state must be shown explicitly.
• Can state that this site does not run model calls or generation demos.