Adora

Adora flags dropshipping storefronts that target Israeli shoppers. It's a browser extension — Chrome and Safari, desktop and iPhone — backed by a pipeline that scrapes the Meta Ad Library every night, analyzes the sites behind the ads, and maintains a risk database the extension queries as you browse.

A lot of "Israeli" shops advertised on Facebook and Instagram are AliExpress dropshippers with big markups, fake countdown timers and no real support. They churn domains fast enough that a static blocklist is useless, so Adora finds them where they actually acquire customers: the ads themselves.

How it works

1. Scrape — nightly cron jobs query the Ad Library's GraphQL endpoint for Hebrew sale keywords and collect advertisers and their landing URLs. Plain HTTP, no browser automation.
2. Analyze — Playwright loads each new site and Gemini 2.5 Flash classifies it (dropship / legit / service / uncertain) with a 0.0–1.0 risk score and evidence.
3. Price match — products on risky sites are matched against AliExpress listings to show the actual markup.
4. Serve — anything scoring ≥ 0.6 lands in risk_db. The extension calls the FastAPI /check endpoint and shows a warning banner with the price comparison.

The extension ships with a ~22K trusted-domain whitelist baked in at build time, so it never phones home for known-safe sites. Unknown domains are checked once and cached locally. Signed-in users can vote on verdicts, which feeds back into the database.

Architecture diagrams and the full pipeline write-up live in docs/.

Repo layout

backend/app/       FastAPI app — API routes, Gemini scoring, DB access
backend/scripts/   cron jobs: scraper, batch analysis, price matching, summaries
backend/data/      whitelist files
extension/         Browser extension (MV3) — Chrome and Safari/iOS builds
docs/              architecture, scraping pipeline, user system, iOS/Safari notes,
                   project poster and presentation

Stack

Python 3.11 / FastAPI / PostgreSQL backend, Gemini 2.5 Flash for site analysis, Playwright for page scraping. The extension is Manifest V3 with a React popup and a vanilla JS service worker, built with Vite. The same source builds as a Safari Web Extension, wrapped as an iOS app so it runs in Safari on iPhone — capabilities that differ between the two browsers are resolved at runtime rather than forked into separate builds. See docs/ios_safari_extension.md.

Deployment

Everything runs on a single small cloud VM: the API as a systemd service behind nginx and a Cloudflare tunnel, with the scrape → analyze → price-match schedule driven by cron. Scrape jobs are staggered an hour apart to stay under Meta's rate limits, and analysis runs are timed so Playwright and the scraper never compete for the VM's memory. Details in docs/scraping_pipeline.md.

Project Materials

Authentication and account design is documented in docs/user_system_public.md. The project poster (docs/Adora_Poster.jpg) and presentation (docs/AdoraPresentation.pptx) summarise the product and architecture for review.