Framework-driven conversation analysis — comfortable designing LLM-graded prompts against a research framework (e.g., KAB—Knowledge, Attitudes, Behaviors) to evaluate conversations for things like whether a user received relevant information, showed shifts in confidence, or surfaced intent signals such as asking for help drafting a message to an employer. Build and maintain a rigorous evaluation harness: regression tests on prompts, eval sets grounded in real worker needs, working with product team on red-teaming for safety, and bilingual (English/Spanish at minimum) quality checks learning implementation.