Pierce Freeman and Richard Diehl Martinez start with a practical problem: remembering when street sweeping requires a parked car to move. They sketch an eventual location-to-calendar workflow, then deliberately reduce the first version to manual location capture and street-schedule lookup rather than trying to solve automatic parking detection immediately.
The build starts from a repository and an existing Python and TypeScript stack, with a database, typed interfaces and a local development environment. The hosts inspect San Francisco's street-sweeping records, consider obtaining additional OpenStreetMap geometry, then simplify the pipeline after finding line coordinates already present in the source data.
AI-generated code is not accepted without review. They examine parsing, model types, migrations and database queries, move conversion logic into a separate utility, and discuss recording architectural preferences in project instructions. Their disagreement over familiar versus opinionated frameworks becomes a broader lesson about giving agents explicit conventions rather than assuming popularity guarantees correct code.
They then develop an in-memory spatial-search component and a test interface for matching locations to street segments and the relevant side of the street. They favor a deterministic geometric rule over making a model call for every ambiguous location, while still using the coding agent to explore and implement the algorithm.
A convincing early result fails on a later test. A map overlay exposes missing coverage in the data available to their prototype, prompting a maximum-distance check so an unsupported location does not silently receive a distant street's schedule. The episode ends with a working lookup prototype, not a deployed, complete parking service: authentication, data refresh, automatic detection and calendar integration remain future work.
Watch on YouTube
