This service audits an Artificial Intelligence (AI) API by regularly querying it for predictions with the test set entries of an open source dataset.
The initial version is set up to analyse the Google Cloud Sentiment service using the Rotten Tomatoes dataset. The service could be extended to other APIs (for example, Amazon's Comprehend and datasets).
This cloud run container was created and deployed with the kettle-cli. It streams the Rotten Tomatoes test set, which is hosted by Hugging Face, and queries Google's sentiment API for each entry. Overall performance metrics are tracked with the Hugging Face evaluate library, and a sample of misclassifications are retained.
The results are POST'ed to a Heroku app, where they can be displayed. The repository for that website is open source in the ai-auditor-web repository.
### Data
This cron job was run approximately weekly from 03 October 2022 to 23 July 2023. The data that was collected has been added as a CSV file to this repo.