During my academic capstone project, I led the design and execution of a comprehensive evaluation framework for an AI au
During my academic capstone project, I led the design and execution of a comprehensive evaluation framework for an AI auto claims prototype. I created 100+ synthetic FNOL claim files embedded with deliberate contradictions and decoy documentation, then manually labeled every file with ground-truth decisions—applying structured reason codes for fraud indicators (late reporting, recently purchased policy, inconsistent damage), calculating correct payouts under both comparative and contributory negligence frameworks, and citing relevant policy coverage. I then graded the AI's outputs against these pre-labeled benchmarks, systematically categorizing errors by type and producing a detailed 40-page feedback report with root-cause analysis and recommended logic corrections. I used my labeled dataset and targeted feedback to update the AI's rule engine, improving decision accuracy by about 18% in subsequent testing. This shows I can complete the full human-in-the-loop AI training cycle required for this role.