Infer defect location from photos and dictation

Auto-fill defect-level location on comments (e.g. utility room, roof, master bedroom) by combining photo context and spoken narrative, so inspectors stop double-entering location that is already obvious from media or dictation.
Keep scope on the defect’s real-world location, not the report’s section / subsection tree—that stays a separate concern.

May 15 feature sync (Chris): Server-side gap in Report Writer: photos already carry some location context, and inspectors say the location aloud, but the comment location field is still managed manually.
Desired inputs: (1) locations tied to photos, (2) what the inspector says into the phone—example: “water stain, master bedroom” should land the comment location as master bedroom.
Decision needed: confidence thresholds, override UX when inference is wrong, and privacy / retention for audio snippets.

Decision needed: concrete repo / branch / package where Report Writer comment, photo, and audio models live in this workspace snapshot those paths were not found under obvious names—coordinate with owners of ATT-1672 (server schemas) and the mobile MVP repo.

Multimodal pipeline (image metadata + ASR transcript + LLM or rules) and persistence onto the comment location field—Decision needed whether inference runs sync on save or async job with spinner.

Dictation capture quality, background noise, and offline queue behavior—Decision needed which SDK path the Report Writer app uses.

attik-frontend client report surfaces (e.g. defect presentation under src/app/client/reports/) are consumer UIs for finished reports, not the Report Writer authoring app—cite only if product wants labeling consistency between authored location and published report.

Sibling planning: ATT-1672 (move Report Writer schemas to backend).
Client report presentation tree (naming alignment only): attik-frontend/src/app/client/reports/ (e.g. components/StylesAndMaterialsCard.tsx lives alongside other report UI).

Attik