Notes
Short writing around validation choices, metrics, optimization, visual odometry, and speech-processing workflows.
Whisper becomes more useful when transcription is treated as one stage in a larger NLP system.
A confidence selector can turn deep and classical visual odometry branches into a more robust research architecture.
The fastest pipeline improvement is often moving repeated Python-level work into vectorized array operations.
Separating sale occurrence from sale magnitude can make sparse forecasting systems easier to reason about.
WAPE is useful, but sparse demand and high zero ratios can hide important failure modes.
A practical note on preserving temporal integrity when lagged features depend on previous predictions.