This collaboration develops a locally deployable interface for analyzing open-ended library survey responses without sending sensitive institutional data to external AI services. A staged map-reduce workflow helps analysts define an interpretable label set, maps bounded batches of comments into validated structured records, and aggregates the results while preserving links to the source comments.
The prototype supports local Ollama models, schema validation, visible fallbacks, versioned consent, pseudonymous participant records, and structured user feedback. Following substantial work on data-use and human-subjects approvals, the project has reached pilot testing and study preparation. It provides a research platform for examining privacy, trust, verification, and effective use of small open models in resource-constrained library settings.