AI

How large language models show up inside Galaxy Vast products: which features use them, and what keeps the output trustworthy.

Large Language Models Multimodal Prompt Engineering RAG On-device Inference

Applications

Where models do the work

Four places a large model is already carrying weight inside the product family.

Galaxy Notes can condense a long entry or answer a question about what you wrote last month — retrieval over your own material, not the open web.

Galaxy Scan leans on vision models to find page edges, correct perspective and tidy recognised text before a PDF is ever exported.

Galaxy Ledger drafts a category and an amount from a photo of a receipt; every correction you make feeds the evaluation set behind the next tweak.

Across the fleet the rule holds: models suggest, people decide. Nothing writes, deletes or spends without a visible, reversible step.

Responsible AI

Wherever a model touches something that belongs to you, these six are the commitments behind it.

Held at the same time, on every feature

None of this is a committee. It is five habits small enough that a company our size can actually keep them.

One named owner

Every AI feature has a person whose name is on it — the one who decides what the model may do, and who answers when it does something it should not have.
Written down before the build

Before code exists we note what could go wrong: who is affected, what data is touched, and what the worst realistic output would look like.
A way to report it

Bad output can be flagged from inside the app. Reports reach the feature owner, get a root cause, and the case joins the evaluation set so it cannot quietly come back.
Marked where it appears

Generated text and generated images say so in the interface. Nothing here pretends to be something a person made.
Rules we did not write

Store policies and platform guidance move faster than we do. We re-read them each release and pull a feature back when it no longer fits.