Trustworthy AI for public infrastructure inspection

A multimodal inspection platform that turns field photos, notes, and voice observations into structured, evidence-backed reports while tracking how infrastructure changes across inspections.

Inspection capacity, not willingness, is usually the real bottleneck in keeping infrastructure safe. Field engineers synthesize scattered evidence, photos, handwritten or voice notes, into a formal report, often hours or days after the visit, from memory, against a blank template. That process is slow and hard to verify after the fact.

The core interaction is a session, not a chat: an inspector collects evidence throughout a visit, then triggers a single Generate Report call. Gemma 4 reasons over the full evidence set at once, rather than captioning photos in isolation, and returns a schema-constrained report: severity, defects, a recommendation, and a self-reported uncertainty flag when evidence is ambiguous.

When an asset has prior inspections, Gemma is offered one tool, get_site_history, and decides for itself whether comparing against them would strengthen the assessment, producing an improving, stable, or deteriorating trend alongside a severity-over-time chart.

Evidence-to-defect grounding uses small integer photo indices rather than database IDs, asking Gemma to reproduce opaque identifiers inside nested JSON measurably hurt reliability. Voice notes are transcribed by Gemini 3.6 Flash, not Gemma, since Gemma 4's audio input is only enabled on edge-deployment variants not exposed through the hosted API this project uses. A slower, more careful model was chosen over a faster one after testing showed the faster model hallucinating details it was never given.