Review-summary pipeline

Each stadium collects many reviews. Summarizing when nothing has changed wastes money, and a summary that doesn't reflect new reviews feels stale.

A GCP Workflows run starts every midnight and checks for new reviews. If there are none, it doesn't run, so the AI isn't paid to repeat itself. When there are new reviews, it regenerates the summary. Two ideas keep it accurate:

• Latest 50. The summary uses the latest 50 reviews, because older ones often describe problems that have since been fixed.
• Tags for unresolved issues. New reviews are tagged by topic, such as cleanliness, food quality and the general experience. A few flagged older reviews are included, so a serious issue that was never resolved isn't dropped. For example, a long-ago "the toilet is not clean" review stays visible until it's resolved.

Summarizing on every request, which wastes money when nothing is new. Waiting for a call before updating, which saves cost but shows stale data. I chose fresh data over saving that cost.

Designed and implemented an automated AI-processing workflow for consolidating user reviews associated with stadium experiences.

When users submit new stadium reviews, the system incorporates the newly available review information into a processing pipeline used to generate an updated consolidated representation of the stadium's review data.

Designed the workflow using GCP Workflows to coordinate the different processing stages required to retrieve relevant reviews, process the information, and generate an updated AI-produced summary.

The workflow was designed to process accumulated review information on a scheduled basis, including a midnight processing cycle, rather than repeatedly performing the complete summarization process for every individual user request.

Also developed AI evaluation tests for the review-summary functionality to validate generated results and support improvements to the summarization process.

Even a single new review means going through the whole summary again.

Each stadium has an up-to-date summary, and unresolved issues stay visible.