Built on Apple Foundation Models

IncidentScribe runs on Apple’s Foundation Models — the on-device language models that ship as part of Apple Intelligence on Apple Silicon. There’s no API call, no inference server, no cloud round-trip. The same model that powers system features does the extraction and drafting, locally, on your Mac. It’s one of the few shipping production apps built end to end on the framework.

Why on-device is the whole point

For IncidentScribe’s buyer — SREs under standing cloud-LLM bans — on-device isn’t a nice-to-have, it’s the only thing that makes the product usable at all. A model that runs locally means incident data never has to leave the machine, which means the App Sandbox can enforce that it doesn’t.

Designing around a token ceiling

On-device models trade raw context size for speed and privacy. The Foundation Models context ceiling is the single most load-bearing constraint in IncidentScribe’s architecture — you can’t throw a 5,000-line incident channel at it in one shot. So the product doesn’t try. It runs a five-stage pipeline:

1. Chunk — pure Swift, natural-boundary slicing of the raw input.
2. Extract — the model emits typed timeline events against a constrained schema, per slice.
3. Validate — pure Swift; every event traced back to its source, fabrications dropped.
4. Reconcile — pure Swift; per-slice timelines merged into one ordered master.
5. Draft — the model writes prose against the structured timeline, never the raw text.

Three of the five stages are deterministic code. The model is used only where judgement is genuinely needed — extraction and drafting — and is constrained at both ends by typed schemas. That’s how a compact on-device model produces output reliable enough to sign off.

Schema-typed in, schema-typed out

Extraction and drafting both use Apple’s @Generable structured-output support: the model returns typed records, not loose text. That’s what makes the timeline mergeable and the citation chain possible. The architecture is built around the model’s strengths and walled off from its weaknesses.

Open in Mac App Store