Reading handwritten survey responses in the same window the event was happening.

The situation #

A large events team needed to collect feedback from attendees across hundreds of events a year. Surveys went out on paper. Attendees filled them in by hand. Coordinators manually keyed responses into spreadsheets, then pushed those into the analytics platform. By the time the data was clean enough to read, the event was over and the next one was already in flight.

What we found #

The bottleneck wasn't analysis, it was digitization. Off-the-shelf OCR tools handled clean printed text fine, but the team's surveys mixed checkboxes, printed scales, and handwritten comments on the same page. Generic models couldn't read them in one pass, and splitting the page into multiple captures broke the workflow on the floor.

What we built #

A custom-trained object-detection model that ran from a smartphone camera with no special hardware. Field staff could photograph a stack of completed surveys, and the model extracted printed checkboxes, scale answers, and handwritten comments in one pass. The results streamed straight into the team's existing analytics pipeline through API integration, and an admin dashboard let event leads watch sentiment shift through the day.

This work pre-dated the current GenAI wave. The engineering was in training data, model selection, and edge-case handling, not prompt design. We shipped ML when ML still meant building your own model.

What changed #

The digitization step that used to span days collapsed into the same window the data was being collected in. Event leads saw sentiment as it formed, instead of weeks after the room had emptied. The capture system was reused across multiple event series without retraining.

Most AI conversations today are about LLMs. The harder question is which problems still belong to other tools. We start there. Start an audit →