· Lawrence Rowland’s MetaGPT experiment
Choosing a contractor
What would a personalised news feed look like if the things it recommended were UK building contractors?
The original brief borrowed the idea from Toutiao. A homeowner would open a shortlist, look at a contractor’s trade and reviews, and decide who was worth a closer look.
Walk through the screens
The Python names five page templates. The 2023 commit contains none of their HTML and no screenshots. This gallery reconstructs the pages from their routes, fields and messages. The layouts are new and the examples are illustrative.
Open the full screen gallery ↗
Pages named in the code
The signed-in home route asked for recommendations. A reader could open a contractor’s profile and write a review.
A form without a page
The contractor form specified name, specialty, rating and description. The API design described creating a contractor. The Flask routes never gave this form its own page.
Ideas in the product brief
The brief asked how people would narrow a feed and remember a good option. The views below interpret those requests. The data model also recorded a saved-contractor relationship.
The three original drawings
The product brief and system design contained three Mermaid diagrams. Each drawing here renders a 2023 definition. The text links open the original instructions for drawing it.
The reach and engagement chart above records the product’s intended position. Read the chart definition.
What put a contractor on the shortlist?
The proposed recommender took each review’s user, contractor and rating. A nearest-neighbour algorithm would estimate how that user might rate each contractor, sort the estimates and return up to five names. A contractor’s displayed rating and the estimate used for ordering were two different values.
The original test supplied one plumbing contractor called testcontractor with a rating of 4.5. It checked whether a recommendation list contained anything. The training code passed a Python list to a method expecting a data frame, and the project lacked the templates and database setup needed for an end-to-end run. The historical record gives us the intended ordering, but no measured results or observed shortlist.
Lawrence’s MetaGPT commit, run a contract recommender, contains six Python files and three design documents. The file paths and date can be checked against the original revision. Read the 12 August 2023 commit. Original MIT licence.