Adam Meskouri

I train AI models to take on research questions that once required substantial manpower. My system prompts, some as long as 50 pages, have enabled commercial AI tools to perform complex content analysis and research tasks consistently, with about 99% accuracy.

I specialize in six kinds of model training:

• Adversarial testingProbing models with deliberately misleading inputs to expose classification failures.
• Perturbation testingMeasuring output stability when inputs are reworded, reordered or reformatted.
• Ablation experimentsRemoving prompt components one at a time to isolate their effect on output.
• Subgroup fairness testingComparing accuracy across subgroups to detect disparities that aggregate metrics conceal.
• Calibration testingChecking whether a model's stated confidence matches its observed accuracy.
• Repeat-run consistencyMeasuring agreement across independent runs of identical queries.