Context. Regulated drought-subsidence building-damage claims, assessed with a computer-vision plus vision-LLM assistant that keeps the human expert in the loop.

Problem. Field experts manually inspect large seasonal volumes of building photographs (~150k per season) and hand-fill a standardised damage schedule — slow, subjective and hard to scale.

Outcome. The AI proposes a structured, pre-filled assessment in minutes; the expert validates or corrects and always makes the final call. The platform runs on Azure Container Apps: it can keep multiple warm replicas for high-availability operation, or switch to a cost-optimised scale-to-zero profile through the durable async gateway. The design prioritises recall — never missing a real crack — because a false positive is dismissed in seconds while a missed crack means an unpaid claim.