I build small, verifiable data tools.

My work is measurement-first: I take a messy input, define what "correct" means, and report the number that proves it. On a 561-feature sensor dataset I ran a 96-configuration grid search with 5-fold cross-validation (480 fits) and reported both the cross-validation score (99.3%) and the held-out score (95.4%) — the gap is the part that tells you whether the model actually generalises. On a 10-class image task I fine-tuned VGG19 and logged accuracy every epoch, from 90.0% to 94.8%.

I hold the Stanford / DeepLearning.AI Machine Learning Specialization (supervised learning, advanced algorithms, unsupervised learning & recommenders), verifiable at coursera.org/verify/specialization/BOPE9L7H8GCS.

What I sell here is the same discipline applied to your file: send it, get back a normalized table plus a row-by-row report of every change. No credentials, no personal records — public sources and files you hand me directly.