This is the story of FeedClinic — its idea, its architecture and its business value. The platform grew out of a very specific, very ordinary pain in Greek e-commerce: the product feed that breaks just before the weekend and quietly wipes an entire shop off the price comparison engines.
1. The Problem: When a Single File Decides Your Revenue
Every Greek e-shop listed on the price comparison engines (Skroutz, BestPrice) is legally required to publish a machine-readable product catalogue in XML. That file is remarkably fragile: one bad encoding character, a fragment of HTML inside a description, an invalid availability value, or an image served over http instead of https is enough for the entire feed to be rejected. The consequence is immediate and expensive — the shop disappears from comparison results and loses sales in real time.
Traditionally, tracking down the error falls to the developer or the agency, usually under time pressure with an angry client on the phone. Conventional validation tools simply return dense lists of errors without proposing or applying any fix, turning a small technical fault into hours of manual debugging.
2. The Solution: One Input, Four Stages
FeedClinic was designed around a deliberately simple idea: the only input is the feed URL. From there the platform handles everything in four stages. First it diagnoses the feed against the official spec, producing a clean Health Report that separates blocking errors from warnings. Then it automatically repairs whatever is safe to repair — encoding and XML fixes, HTML stripping, availability normalisation, http to https image rewriting, duplicate removal — with every change recorded in a transparent before/after diff.
Next, it hosts the clean feed at a stable CDN URL that the merchant registers once and never has to change again. And finally, optionally, it generates bulk Greek SEO product descriptions with AI, closing one of the biggest quality gaps in Greek catalogues.
3. The Philosophy: Safety and Absolute Trust in the Data
The most important design principle was absolute trust in the data. The tool never invents information: critical identifiers such as MPN or EAN are only extracted from existing text and explicitly confirmed by the user — never guessed. Automatic fixes are strictly limited to changes that provably cannot break the correspondence between the feed and the actual shop, so price and availability always stay accurate.
The AI follows the same rule: descriptions are generated strictly from each product’s real data (title, attributes, category), with a mandatory human approval step before any export and a credit ledger with clear limits. The technology here is not asked to impress — it is asked to be reliable and legally safe.
4. The Architecture: Reliability at Scale
Behind the simple experience sits an architecture built for endurance. Feeds are always parsed in streaming mode, so even catalogues of 50,000 products are processed without exhausting server memory. The validation rules are data-driven — declared in YAML — which means a new check can be added without touching the core code.
The most important reliability feature is the keep-last-good mechanism: if the upstream feed source breaks, the last valid file stays live on the CDN and an email alert goes out, so the shop never drops off. The whole system runs on Python/FastAPI with Celery workers and PostgreSQL on the backend, Next.js 15 with TypeScript on the frontend, containerised with Docker for a zero-ops setup that a single engineer can maintain.
5. The Outcome: From Technical Anxiety to Quiet Nights
FeedClinic turns one of the most stressful parts of Greek e-commerce into an invisible, managed service. For a shop owner it means the feed stays live and validated around the clock. For an agency maintaining dozens of e-shops it means every feed lives on one screen, with problems caught proactively — before the client even calls.
For me as an engineer, the project captures a conviction that runs through all my work: the best technology is the kind that disappears, leaving only the result behind. In FeedClinic’s case that result is simple but decisive — a feed that just works.
