Blogs by Colt

Colt Cruz
September 27, 2026

This month, my main goal was to turn my customer-service handoff project into a working research study that could compare decisions made by general participants, a professional customer-service reviewer, and a GoEmotions-based handoff system. The issue I wanted to address was that automated customer-service systems do not always know when a conversation should be transferred to a human. My strategy was to use the same 25 customer-service conversations across all three groups so that their handoff decisions could be compared consistently.

The main features I worked on included the participant study flow, professional reviewer mode, GoEmotions emotion analysis, data collection, and the research results dashboard. I used VS Code, GitHub, Supabase, JavaScript, Python, GoEmotions, and the customer-service conversation dataset to build and manage the project. Supabase was used to store study responses, while GitHub supported deployment and version control. GoEmotions and Python were important for analyzing emotional signals within the conversations and connecting those signals to the handoff logic.

One of the biggest challenges this month was recruiting enough participants. My goal is to collect responses from at least 30 general participants, but recruitment has been slower than expected even after sharing the study in several Discord communities. At the same time, collecting preliminary data has already been useful. The early results show noticeable differences between general participant decisions, the professional reviewer, and GoEmotions. These differences have helped me better understand where the current handoff system may need improvement.

Several things went well this month. The study became functional enough to begin collecting real responses, the professional reviewer completed the same set of conversations, and I was able to start comparing agreement rates. Working with my advisor also helped refine the study. Feedback included clearly explaining the role of GoEmotions, collecting IP addresses to reduce duplicate submissions, continuing toward the 30-participant target, and preparing the project for presentation. This feedback helped make the study methodology and final presentation much clearer.

The biggest lesson I learned is that emotion detection alone may not be enough to make the same decisions that an experienced customer-service professional would make. My original hypothesis expected the GoEmotions-based system to align more closely with the professional reviewer, but the preliminary results currently show stronger agreement between GoEmotions and the general participants. Instead of treating this as a failure, it has given me a clearer direction for the project: a future system could combine GoEmotions signals with professional customer-service decision patterns to create a hybrid handoff logic.

Moving forward, I need to continue recruitment, complete data collection, and analyze the final agreement rates before drawing conclusions. I also want to explore the hybrid approach and determine whether combining automated emotion analysis with professional judgment can produce more useful handoff decisions. I feel prepared to move into the next phase because the study is functioning, I have preliminary data to learn from, and I now have a clearer idea of what the project could become after the defense. For the blog, useful visuals would include a screenshot of the participant conversation interface, the research results dashboard, and a diagram or screenshot showing the GoEmotions-based handoff process.