ComplyGuard AI
Live on the Shopify App Store
A compliance suite that audits a Shopify store against EU AI Act Article 50, grades it, and writes the fixes back to the storefront.
Highlights
- Scans images, apps and policies, then grades the store from A to F.
- Falls back across four AI providers so one outage doesn't stop a scan.
- Caches every verdict, so a rescan costs nothing for unchanged media.
- Three surfaces: embedded admin, theme extension, checkout extension.
The problem
Article 50 of the EU AI Act requires any store selling into the European Union to declare AI-generated media and AI interactions. Most merchants cannot answer the basic questions it raises: which product images came out of an AI tool, which installed apps use AI behind the scenes, and what their privacy policy is missing. Finding out by hand means opening every product, every app listing and every policy page — and doing it again after the next catalog update.
What I built
ComplyGuard AI audits a store against Article 50 and fixes what it can. It scans the product catalog, the installed app list and the store's legal policies, grades the store from A to F, and files every problem as a ticket with a severity level and the evidence behind it. One click then writes the fixes back: AI-generated labels into image alt text, disclosure clauses into the privacy policy and terms. On the storefront, a theme extension adds a transparency badge and a cookie consent banner, and a checkout extension shows the disclosure before the order is placed. Merchants can download a PDF certificate as proof for regulators.
Challenges
A scan cannot finish inside a web request
Scanning a catalog means an API call for every uncached image, which on a few thousand products runs far past any request timeout — while the merchant sits inside Shopify's admin iframe watching a spinner. I moved the scan into a background runner that writes its phase, progress and image count straight to the database, and gave the interface a polling endpoint to read from. The merchant can close the tab and come back; the scan keeps running and the progress is still there.
Image analysis costs money on every call
Each uncached image is a paid API request, and merchants rescan often, so the naive version would bill for the same image again on every scan. I added a cache keyed on shop plus image URL, so a rescan reads the earlier verdict instead of calling a provider. The cache also records what the model actually saw — whether it found people, text or AI artefacts — so a verdict can be explained later rather than taken on trust.
One AI provider is a single point of failure
Relying on a single model meant any outage stopped every scan, and models do not always fail loudly: they return a confident wrong answer. I built a provider table ordered by priority across OpenAI, Claude, Gemini and Flux, and the runner walks down it until one returns something usable. Every call records its errors and last failure, so a provider that starts degrading can be demoted without a deploy.
Auto-fix writes to a live store
Rewriting another business's alt text and legal policies is not something to do quietly. Every fix is recorded against the issue that caused it, with a timestamp and a note of what changed, and merchants can acknowledge an issue instead of fixing it when the finding does not apply to them. Nothing is removed — fixes only add disclosure text — so the worst case is a label a merchant did not want rather than lost content.
Outcome
Approved and published on the Shopify App Store on 22 July 2026, after passing Shopify's review for OAuth, mandatory GDPR webhooks and recurring billing. Built solo in a month, across three separate runtimes: an embedded Remix admin, a Liquid theme extension and a React checkout extension. The app is new and still early in distribution; the proof so far is the review itself, which is a third party checking the parts of the build that merchants never see.