How vlora.deals works

What this is

vlora.deals is a meta-search engine for online shops that deliver to Vlorë, Albania. Instead of checking a dozen shop websites one by one, you can search or browse across all of them here in one place.

Where the data comes from

Every product listed here was fetched directly from the shop's own public website — product pages, sitemaps, and public store APIs. Prices, stock status, and images are read as-is, refreshed on a nightly crawl. We don't sell anything ourselves and we're not affiliated with any of the listed shops; every "Open at shop" link takes you straight to the original listing to actually buy.

Coverage isn't perfect. Some shops publish a full product catalog we can index completely; others (notably food/grocery delivery via Wolt) only expose a small sample of their real menu through their public API, so you won't always find every item a shop actually carries — only what its own site or app makes visible to us. Very large shops can also take several nights to fully index after being added.

Which AI models, and why

Most shops list products in Albanian. Titles shown in German, English, French, Italian, and Ukrainian are machine-translated, not reviewed by a human. Quality is generally good but not perfect — if you spot a translation that's confusing, wrong, or missing, we'd genuinely like to know. Untranslated titles show a small "translation pending" note until they're processed.

This is a non-commercial hobby project, so cost matters — but we also prefer European, sovereign infrastructure where practical, and for this workload we actually get both. English, German, French, and Italian titles are translated primarily through franconia.ai, a German provider running open models on their own infrastructure (currently Gemma 4 26B), with Groq (GPT-OSS 120B) as a fallback if that's ever unavailable. Ukrainian is translated separately, currently via Qwen through opencode-go.

The model selection isn't arbitrary: we benchmark real product titles — including tricky Albanian ones a generic test wouldn't catch, like "maja" meaning yeast only in a baking context — across the available candidates for speed and accuracy before picking one, and re-check periodically as providers change what they offer. As a rough quality reference, we've also spot-checked the same titles against a frontier model (GPT-5.6); it edges out the small open models on precision, but running something that size for every one of tens of thousands of routine batch translations wouldn't make sense for what this is.

Smaller and open models are not only cheaper; they are faster and use meaningfully less energy per translation than large frontier models. Each product is translated only once, when the nightly crawl discovers it for the first time. The crawl itself uses no AI at all — it simply fetches public product pages and data — and existing products are not continuously re-translated.

Privacy & hosting

All data — the database, images, and everything else — lives on our own server, physically hosted in Germany by netcup GmbH. We do not use a US hyperscaler such as AWS, Google Cloud, or Azure, so German and EU law, including the GDPR, applies directly rather than through a separate international data-transfer arrangement.

The application itself does not write visitor IP addresses. Server-level access-log handling is configured separately; until that configuration has been verified, we make no claim about IP truncation or pseudonymisation. Access logs are used for security and abuse prevention, not for profiling visitors.

There are currently no third-party analytics or tracking scripts. We may later add a privacy-respecting, self-hosted Umami installation solely to see rough daily visitor counts, and nothing beyond that.

Feedback

This is an early, independent project with no visitor accounts and currently no third-party analytics or tracking scripts. If something looks broken, a shop is missing, or a translation is off, email hello@loeser.solutions.