Human-in-the-Loop

AI classifications aren't always right. When the model is confident, auto-accept. When it's not, ask the human.

This pattern:

1. Check the confidence score against a threshold
2. High confidence → accept automatically
3. Low confidence → show the suggestion, let the user confirm or override

Why care? HITL is the difference between a trusted AI assistant and one that silently makes mistakes. The cohort's Telegram bot uses this exact flow — high confidence auto-files, low confidence sends an inline keyboard for human review.

Task

Complete process_with_hitl(result, threshold):

• If confidence >= threshold, return the category directly
• Otherwise print the low-confidence result and prompt the user
  • Prompt: f"Accept '{result.category}'? (Enter to confirm, or type a category): "
  • Empty input or "y" → return the suggested category
  • Anything else → return the user's input as the category