CijeneHR
A price comparison engine whose real product is trustworthy data.
- Type
- Product
- Industry
- E-commerce / comparison
- Platform
- Web application
- Role
- Design, build, deploy
What made this hard.
A price comparison platform for Croatian retailers: it scrapes and normalizes products across stores, tracks price history, and puts a fast search in front of it all.
Scraping many retailers with different page structures is the obvious challenge. The unobvious one is knowing which failures are worth fixing: a wrong selector and a blocked request look identical from the outside, and fixing the wrong one burns weeks.
On top of that, a comparison site has to decide what to do with a suspicious row. Quarantine it on write and you lose real offers; show it anyway and you lose the user's trust the first time a price is wrong.
What we built.
- Built a price-extraction diagnostics classifier (ok, price-present-unparsed, JSON-LD-no-price, no-price, antibot, transport-failed, empty) as a pure function behind a swappable page-source seam, so it runs through the real fetch path yet stays unit-testable.
- Added an artisan report ranking stores by recoverable volume. It inverted the plan: the ceiling was transport, not selectors, so selector work was correctly not scheduled.
- Chose trust-but-verify over quarantine-on-write, and recorded the rejection as an ADR: suspicious rows carry review-only flags that never auto-disable them, with display-time validation as the single enforcement point.
- Diagnosed a whole-pipeline outage to a missing working directory in the process-manager config: workers started with a root cwd, the headless-browser smoke check failed to resolve its module, and every source false-negatived into an open circuit breaker. Fixed by pinning the directory in both the process manager and the node launchers.
- Wrote an artisan command that builds a browser-test manifest from the route table, reflection-resolving route-model-bound parameters to real seeded rows. It caught genuine production bugs, such as a controller querying renamed columns and a view missing data, that unit tests never would have.
- Pinned the behaviour of the largest legacy classes with golden-master tests before refactoring them, and gated dead-code removal on evidence rather than intuition: static analysis for safe deletes, access-log evidence for routes, a deprecation-log shim and a watch window for anything ambiguous.
What it runs on.
Chosen to be maintainable for years, not to win an argument on the internet.
- PHP 8.2+
- Laravel 12
- Vue 3
- TypeScript
- Tailwind CSS
- MySQL
- Redis
- Playwright
- GitHub Actions
Where it landed.
CI runs a backend job plus a fast browser subset on every push, and the route-smoke manifest regenerates itself as routes change, so coverage tracks the application instead of drifting behind it.
The diagnostics work is the part we'd reuse anywhere: measuring which failures are recoverable before choosing what to fix is cheaper than any amount of selector maintenance.
Need something in this territory? This is Custom web app development work. Start from all services or see pricing.
Related case studies.
Your project
Tell us what you're building. We reply within two business days.