Clean OCR and manually entered interview experiences before storage so the RAG corpus has less noise, fewer duplicates, and more consistent company, department, and question fields.

Compare the current interview against historical sessions at the dimension level, highlight progress and regressions, and turn weak areas into concrete reinforcement advice.

Control live interview pacing so the interviewer behaves like one realistic person, avoids repeating the question list, stays on the current question, and only switches mode when the candidate explicitly asks for explanation.

Ground question generation and interview follow-up in the candidate's actual resume by extracting internships, projects, tech stack, achievements, and likely deep-dive topics.