Asset Watcher
an autonomous agent that watches real-estate and vehicle listings and judges each one against what you actually asked for.
- role
- sole author
- stack
- Python · Playwright · Amazon Bedrock · Claude · PostgreSQL
The problem
Listing sites filter on fields: price, rooms, mileage. What people actually want is harder to express: "a quiet street, not ground floor, close to a train", or "one owner, full service history, no accident damage". Those details live in free-text descriptions, which rigid filters can't read, so you end up checking the same sites by hand every day.
How it works
- Collect. The agent watches listings from public APIs and, where there's no API, from sites it reads with a headless browser.
- Extract. Each listing is turned into structured data, pulling details out of the free-text description as well as the fixed fields.
- Judge. An LLM on Amazon Bedrock checks each listing against the criteria you wrote in plain language, instead of a fixed set of filters.
- Deduplicate. The same property or vehicle is often posted more than once, so duplicates are merged before anything reaches you.
Why an LLM here
Rules break the moment a seller describes the same thing differently. A model reading the whole listing can weigh several soft criteria at once, which is exactly the judgement a person makes when scrolling.