Survey-to-AI Pipeline for Patient Insights

We built an end-to-end pipeline that moved survey data from collection to actionable insight without manual intervention. The architecture used a multi-cloud approach driven by practical constraints: SurveyMonkey webhooks triggered AWS Lambda functions on each new response, which handled initial data validation and routing. GCP Cloud Functions ran the AI analysis layer — we used the ChatGPT API to classify open-ended responses into clinical categories (wait times, staff communication, facility concerns, treatment satisfaction) and extract sentiment scores.

The classified and scored data landed in Snowflake, where we built transformation models that aggregated responses by department, provider, time period, and concern category. The key design decision was near-real-time processing: instead of batch-processing surveys weekly, each response was classified within minutes of submission.

We built two reporting layers. The operational dashboard showed department heads their rolling 7-day patient sentiment with drill-down into specific concern categories — if wait time complaints spiked on Tuesdays, that showed up immediately. The strategic dashboard gave clinical leadership monthly trend analysis with statistical significance testing, so they could distinguish genuine shifts in patient experience from normal variation.

The AI classification required careful prompt engineering and validation. We tested the classifier against 200 manually categorized responses and iterated until accuracy exceeded 90% across all categories. Responses the AI wasn't confident about got flagged for human review rather than silently miscategorized. We also built a feedback loop: when human reviewers corrected a classification, those corrections fed back into the prompt examples, improving accuracy over time without requiring model retraining.