Blog
Notes on ML, model evaluation, and what I learn building things.
A rigorous benchmark of 9 Chain-of-X prompting techniques across 2 LLMs and 5 task types reveals that plain baseline prompting outperformed every technique on quality - while costing 2.27× less latency.
A controlled study comparing model performance on raw versus engineered feature sets across clinical diagnosis and financial market prediction, proving that thoughtful feature engineering consistently improves model performance.
A systematic experiment across 20+ regression and classification models with varying feature-target correlation (ρ = 0.0 to 1.0) reveals that signal quality dominates model complexity - and exposes an unexpected MAPE artefact along the way.
Accuracy is the first metric you learn - and the one most likely to lie to your face. A structured experiment across 6 classifiers and 7 regressors exposes exactly how far accuracy can mislead you.
A hands-on experiment: what happens to a Linear Regression model's predictions when the test distribution shifts away from training? The results are more nuanced than you'd expect.
In ML, models get all the glory - but data does all the work. A controlled experiment across 5 classifiers proves that balanced, clean data outperforms complex architectures on messy data, every time.