Tobiasz Chatys

Designed adversarial evaluation tasks to expose failure modes in frontier models. Authored a rubric of 20+ graded checks, including negative non-hallucination criteria generalized to root cause, and tuned task materials so a strong model can still fail for legitimate reasons while an expert can reach a defensible answer. Wrote distractors that induce specific wrong answers to make grading diagnoses traceable to stated reasoning errors. Evaluated speech-to-speech models side by side against written guidelines, scoring turn-by-turn and attributing hallucinations and audio defects to the responsible model.

Produced validated lift and drag coefficients from first-hand measurements by testing a NACA 0015 aerofoil at angles of attack from −5° to 15° in an open-return wind tunnel at steady speed. Logged roughly 40 pressure taps using a Scanivalve scanning system and a Furness FCO510 transducer, tuning averaging time against drift versus scan-time trade-offs and deriving freestream velocity from the static-pressure drop across the settling chamber. Integrated pressure data numerically in MATLAB and compared results against published NACA reference data, matching lift curve trends up to stall where flow separation reduced measured lift.

Separated two-phase flow structure from instrumentation noise using Proper Orthogonal Decomposition (POD) implemented in MATLAB, improving Reynolds shear stress estimation accuracy by removing noise and retaining 30 components (about 1.2% of the basis). Scaled processing by coding an incremental SVD that streamed snapshots instead of loading full batches, reducing runtime by ~10x across thousands of snapshots and multiple operating conditions and flow phases. After noise removal, performed Welch PSD on temporal coefficients to reveal a frequency reversal and increased excess kurtosis, recovering buried physics in the raw signal.