Industrial ETL Modernization and Survey Sentiment Triage

Challenge

The business had important operational data moving through fragile ETL procedures. Some work depended on manual cleanup, one-off files, and institutional knowledge. That made reporting slower than it needed to be and made it harder to trust downstream analysis.

The company also had survey data with useful signal buried in open-ended responses. Leaders did not just need a positive or negative score. They needed to know what people were upset about, which issues were addressable, and what actions might reduce the problem.

Approach

I separated the repeatable ETL work from the judgment work. The data cleanup and transformation logic moved toward clearer, testable procedures. For the survey work, I treated early AI as a triage layer rather than an oracle: classify themes, surface representative comments, identify sentiment, and suggest possible mitigation steps for a human to review.

Solution

The workflow created a cleaner path from raw operational inputs to usable reporting tables. On the survey side, the analysis grouped comments by issue, summarized the underlying concern, and paired each theme with practical next steps. For example, if responses clustered around scheduling, communication, or service quality, the output focused on what a manager could actually change.

Results

The project showed how a company can get value from AI without handing the whole process to AI. Deterministic ETL handled the parts that needed consistency. AI helped read and summarize language at scale. The useful output was not a model score. It was a shorter path from messy feedback to operational action.