Ask HN: Feasibility of AI that converts user workflows into local ML apps?

Today, most teams building ML models serve them through web APIs. That works well, but I’ve been wondering about a different path:

Suppose I provide an app that records a user’s screen (with consent).

The user narrates what they’re trying to accomplish in their workflow.

AI analyses this + the narration and suggests what ML model(s) would be needed.

Finally, it packages everything as a lightweight, local, installable app (EXE/DMG), so the user doesn’t need servers, Docker, or APIs at all.

For example: a cashier or agent might be processing transactions. By narrating “I want to flag suspicious ones,” the system could map that to an anomaly detection model, and produce a small local app that does just that.

In other words, “compile” ML workflows directly into end-user software, skipping the typical web-API path.

I'm thinking 'shallow ML' for a start, applications that could be served by decision trees or random forests, for example.The plan is to 'unroll' into C.

Has anyone seen attempts in this direction, or reasons why it might be a dead end?