Business Context

At AcmeAI, account managers use a generative assistant to draft client-facing answers about product capabilities, pricing, and compliance. Leadership wants a lightweight NLP system that flags responses with high hallucination risk before they are shared with non-technical stakeholders.

Data

You have 180,000 historical prompt-response pairs labeled by reviewers as Low Risk, Needs Review, or High Risk for hallucination. Text is English-only. Prompts range from 10-120 words, and model responses range from 30-600 words (median 145). Labels are moderately imbalanced: 62% Low Risk, 28% Needs Review, 10% High Risk. Metadata includes model version, retrieval-used flag, and product domain (sales, legal, support), but the core task should rely on text.

Success Criteria

A good solution achieves High Risk recall >= 0.90, macro-F1 >= 0.82, and supports analyst review with interpretable signals. Inference should stay under 150 ms per response in batch scoring.

Constraints

• Must run inside a secure Python service with no external API calls
• Model size should be small enough for CPU deployment
• Review teams need explanations they can use in stakeholder training
• False negatives on High Risk are more costly than false positives

Requirements

1. Build a 3-class NLP classifier to score hallucination risk from prompt-response text.
2. Design a preprocessing pipeline that captures unsupported claims, hedging, and overconfident language.
3. Implement a modern Python solution with a strong baseline and a transformer-based approach.
4. Explain how you would evaluate class imbalance, thresholding, and calibration.
5. Describe how outputs would be translated into plain-language guidance for non-technical stakeholders.
6. Identify common failure modes, including domain shift across model versions and product areas.