What it does
Every 30 minutes it crawls the GPU, CPU, motherboard, memory, PSU, storage and case categories of HardverApró, and Jófogás and Vatera every two hours. It records every price change, title edit, reservation and disappearance, so it knows what a part usually sells for and how fast it goes.
The data problem
Listing titles are written by hand, for example Gigabyte RTX 3070 Gaming OC 8GB garis. A rule parser turns most of them into structured products (brand, model, memory, maker), and a small local language model running on my own server handles the cases the rules are unsure about.
Engineering
- Typed Python (mypy strict), tests and CI, database migrations with Alembic.
- 26 written architecture decision records.
- A server-rendered web UI with deals, price history and watchlists, published behind an invite-only Cloudflare Access gate.
- Runs in Docker on my homelab server, with backups and one-click updates.
