Engineering AI systems at scale.

→

Case study  The same cat, before and after a single tap.

Kids love colorful art. I do too. A child makes a simple drawing, taps once, and a few seconds later it comes back in full crayon color, still theirs. The crayon adapter is a rank-16 LoRA trained on 53 images.

The app first has to recognize what a kid drew. That part is SketchNet, the project right below.

A convolutional neural network for finger-drawn sketches. Four experiments closed the training-inference gap. The key change was rendering original stroke vectors at 128×128 instead of enlarging 28×28 bitmaps. An X-Ray view shows what each layer learns.

It is also the eyes of DoodleCraft above, reading a child's doodle before the app paints it in crayon.

0.98 ms is the time to classify one prepared 128×128 black-and-white drawing. It excludes canvas, camera, and browser time.

Five algorithms scoring the same molecule for drug-induced autoimmunity risk. The spread between their probabilities is the headline; the chemistry interpretation runs once and stays the same regardless of which model the user picks.

A predictive-ML learning project, not a clinical decision tool.

Two PCA pipelines on audio: lossy compression on raw waveform blocks, and denoising on STFT spectrograms. Without spectral subtraction enabled, stationary noise hides at the top of the variance ranking instead of the bottom, and the denoiser does almost nothing.

The data platform processed 6 billion records at peak. I improved EMR and partitioned S3/Parquet ETL pipelines so downstream analytics and ML had fresher data to work from.