Conversational AI
IRI
Model serving and response evaluation for a children’s voice service.
Overview
A voice conversation service for children. I worked on Kanana model serving on RunPod, input/output checks, and development evaluation of the response pipeline.
My work
- Model serving. Worked on the Kanana serving pipeline on RunPod.
- Input and output checks. Evaluated responses and used stage-level decisions to distinguish overblocking from missed risky requests.
- Regression checks. Used timing logs and stage-level decisions to compare changes in the response pipeline.
Development evaluation
The work focused on making pipeline behavior observable during development: where a response was blocked, which risky requests were missed, and how changes affected those decisions.