Skip to main content
marcolauxDesigner & Developer
DEEN

TIQ Solutions2026Web Apps, AI

Takeover, modernization and operation of a property meta search that gathers listings from dozens of sources, structures them via LLM, rates them by lifestyle.

Starting point

The portal brings together rental and sale listings for flats and houses in a large German city. The listings come from the websites of estate agents, housing cooperatives, housing companies and portals. Every night a chain of four steps runs: crawlers save the listing pages, a language model reads around 40 structured attributes from them, an import step reconciles them with the database, and the search shows the result as a list and a map. Each listing also receives ratings for five lifestyles, such as family, nature or public transport, based on figures about the surroundings of the address.

The platform had been built by another team between 2022 and 2024 and then ran largely without further development. During that time many sources had rebuilt their sites; crawlers stopped delivering without anyone noticing. The frontend ran on an outdated framework version, and a paid subscription model was to be removed again.

My role

Since 2026 the platform has been in my sole technical care, across all its parts: frontend, API, database, crawlers, LLM extraction and operations. The original application and its design come from the earlier team.

Approach

  • A nightly report instead of silent failures: After every run, an email lists all sources: links found, pages stored, listings after extraction and errors. A source only triggers a warning if it delivered in at least three of the last seven nights, so sources that rarely post listings cause no false alarms.
  • Robust crawlers: The Playwright crawler has a time budget per source, a page limit and a watchdog for the whole run; if one source fails, the others keep going. Listings that only appear in the browser are read by a generic reader controlled entirely by a configuration file. I rewrote the configuration for ten sources whose listings had moved.
  • Repair tool for rebuilt sources: A command-line tool follows the links of a broken source, groups them by URL shape and asks a language model, in a single call, which group holds the listings and which the detail pages. It then checks the proposal without a language model by running the real crawler; a person decides whether to adopt it.
  • LLM extraction and import: The extraction gained district detection, an evaluation per source and a mechanism that keeps the chain running even after the step crashes. Known listings only go back to the language model if their last extraction is more than 14 days old. Listings without an exact address are placed in their district so that their markers don't overlap on the map.
  • Frontend: The application moved to Nuxt 4 and Pinia 4. The paid tier is gone from frontend and API, and visitors browse all results without an account. A revised sign-in dialog and a contact form were added.
  • Operations: Every part has a deploy script that keeps the previous container image for a rollback. The database has a local environment, a schema comparison with the live system and a tool that accesses the live database read-only by default.

Result

The chain runs through every night again, and the report in the morning shows which source delivered how many listings and where something needs repair. New or rebuilt sources can be connected by configuration, in many cases based on a checked proposal. The frontend is on a current framework version.

Contact

Planning a similar project?

A conversation clarifies whether and how your project can be implemented – with an initial assessment of scope and timeline.

hi@marcolaux.com
Contact

Discuss a project

A short conversation clarifies scope, timeline, and the most sensible next step. Email or phone is the fastest way.

Marco Laux · Bahnhofstr. 74 · 54338 Schweich