AI Drone Roof Inspection Report — Findings & Repair Scope
Pricing
from $20.00 / 1,000 photo analyzed (hosted, fast tier)s
Turn drone and ground roof photos into a per-building inspection report: defect findings per photo, condition grade, repair scope, materials and labor from your price list, a Markdown report and contact sheet. Rules mode or Claude vision. Pay per photo and per building.
Roof photos to inspect, drone or ground, one per line, in the order you want them numbered (the first line is Photo 1): public http(s) image URLs, Apify key-value-store references in the form kvs:/// (use kvs://default/ for this run's own store), or data: URLs with inline base64 images. JPEG, PNG, WebP, GIF and TIFF are supported; HEIC usually cannot be decoded on the platform, convert it to JPEG first. EXIF time, GPS and camera are read from each file. The three prefilled photos (public-domain U.S. government photos from Wikimedia Commons) are the small, fast set used by Apify's daily automated test.
Tell the Actor which photos belong to which building, as a JSON array of objects: label (e.g. "Building 12" or "Clubhouse"), photos (photo numbers and ranges like "1-7, 158"), stories (integer, used for the height charge) and units (free text, e.g. "12-17"). Photos not covered by a label are grouped automatically. Without labels, photos are grouped by GPS distance when they carry GPS, otherwise by gaps in the capture time. Example: [{"label": "Building 1", "photos": "1-12", "stories": 2, "units": "101-108"}].
Leave empty for the default (the Console prefills rules so Apify's daily automated test never calls a model; clear it or pick a mode): Claude vision on the Actor's own key (hosted) when the Actor has its model key configured, otherwise the free-of-model rules analyzer. Set rules for a model-free run (EXIF index, grouping into buildings, contact sheet, empty findings; no key), or byok to use your own Anthropic API key (tokens billed by Anthropic to you). In model modes a model problem never fails the run: the photo gets the rules output, analysisMethod: "rules-fallback" with an analysisWarning, and is charged the rules price.
Your Anthropic API key, used only when Analyzer is byok. Stored encrypted as a secret and sent only to api.anthropic.com. Get one at console.anthropic.com.
fast (default): Claude Haiku 5.5, cheapest and quick, fine for clear drone overviews and close-ups. balanced: Claude Sonnet 5.5, better on subtle defects (granule loss, lifted tabs, flashing details). best: Claude Opus 5.5 for the hardest photo sets, available only with your own key (analyzer byok); a hosted run with best is rejected before any charge. In hosted mode the tier sets the per-photo and per-building prices ($0.02 + $0.10 for fast, $0.08 + $0.30 for balanced); in byok mode the tier only changes what Anthropic bills you.
Extra guidance for the model, in plain language: the roof system ("architectural asphalt shingles over plywood", "TPO membrane on the flat sections"), what to ignore ("skip the solar panels", "photos 40-45 are the parking lot"), or local terminology ("we call the cap shingles 'ridge'"). Ignored by the rules analyzer.
Your own prices for the deterministic estimate, as JSON: materials [{name, packSize, perUnit, appliesTo, priceUsd}] where perUnit is packs per unit of the repair action named in appliesTo (shingle-zone-repair, cap-repair, flashing-repair, boot-or-vent-replacement…), rounded up to whole packs per building; laborBundles [{hours, priceUsd}] (the largest bundle whose hours do not exceed the estimated hours, never below the smallest is used); heightCharges [{stories, priceUsd}]; laborRounding 'down' (default: largest bundle at or below the estimated hours, as in the reference report) or 'up' (smallest bundle that covers them); hoursPerZone (default 0.4); zoneSf (repair zone size in square feet, default 25); currency. Leave empty to use the built-in EXAMPLE price list, which is a placeholder and not market pricing: replace it with yours before quoting. All quantities are photo estimates and are field-verified on the roof.
Header details for the report, as JSON: name, email, phone, website, inspector, propertyName, propertyAddress, inspectionDate. All optional; missing fields are left out of the report.
Upper bound on photos processed in this run; photos beyond it are skipped and not charged. Use it together with the run's maximum total charge to control spend.
How many photos of the same building are sent to the model in one call (model modes). Larger batches give the model more context per call; smaller batches cost less per call when the spending limit is tight.
Photos with GPS whose positions are within this distance of a building's photos are grouped into the same building. Used only for photos not covered by Building labels.
For photos without GPS, a gap in capture time longer than this starts a new building. Used only for photos not covered by Building labels.
Photos downloaded and decoded, and model calls made, in parallel. Large photos need memory: keep it at 2 on 1 GB.
Photos larger than this are rejected with a TOO_LARGE error item instead of being downloaded. Raise it up to 50 MB for full-resolution drone photos.
Maximum time to download one photo before it is recorded as a TIMEOUT error item.
Also write CONTACT_SHEET to the key-value store: a JPEG montage of the first 100 photos labeled with their photo numbers, handy for the field crew.
