01Challenge
General-purpose models answer children the way they answer adults, and sending classroom conversations to a third-party API was unacceptable to the schools buying the product. Reading level, tone and safety all had to be controllable.
02Solution
A compact open-weight model adapted with parameter-efficient fine-tuning on curated, age-appropriate material, with educators reviewing outputs in the loop and the whole thing deployed inside the client's own infrastructure.
03Business impact
Data residency objections from school buyers were removed, and per-conversation inference cost dropped far enough that the feature could be included in the standard plan.