All case studies
Hallucination Control
Hallucination Optimization Training
Developed a rigorous hallucination taxonomy and annotation pipeline processing 100K+ model outputs, directly reducing factual error rates by 40% in production models.
Hallucination taxonomy
Fabrication— Inventing nonexistent people or events
Date/Time— Incorrect temporal information
Numerical— Incorrect stats or measurements
Logical— Incorrect inference chains
Factual— Wrong person-event associations