Building a Fast CPU correlation kernel

Sun, 20 Sep 2026 00:00:00 +0000

The task: Pearson correlation Link to heading

This is the story of a task I solved during a performance-programming course from Aalto University’s Programming Parallel Computers:

given an ny × nx matrix, compute the lower triangle of its ny × ny Pearson correlation matrix as fast as possible on a CPU. Pearson correlation measures how similarly two rows vary after removing their means. A value near 1 means they move together, -1 means they move in opposite directions, and 0 means no linear correlation.

Building a Compact Handwritten OCR Pipeline

Thu, 18 Jun 2026 00:00:00 +0000

Introduction Link to heading

This project is my solution to the Handwritten to Data Kaggle competition for Ukrainian handwriting recognition. Its dataset is publicly available on Hugging Face as UkrainianCatholicUniversity/rukopys.

Part of my day job is on the deployment side of ML — porting and optimizing models to run on mobile phones, where the recurring questions are “does this fit in memory?”, “does it hit the latency budget?”, and “does it run on the target hardware?”. Within that, I work mostly on document AI: page detection, document classification, image enhancement. Handwritten-text recognition, though, was a piece of the stack I had never built end-to-end myself. I treated this dataset as a chance to fill that gap while staying inside the constraints I optimize for at work.

Engineering on Andrii's Notes

Building a Fast CPU correlation kernel

The task: Pearson correlation Link to heading

Building a Compact Handwritten OCR Pipeline

Introduction Link to heading