Most posts here are about data science and machine learning; some are about engineering. The split is deliberate.

In real production systems the actual model — the math, the loss, the algorithm — is roughly 5% of the codebase. The other ~95% is data pipelines, configuration, serving, monitoring, and the glue between them. The notebooks and simulations on this site cover that 5%: their job is narrow and falsifiable, to show a trained baseline beats random — that the algorithm is not a coin flip.

On Sculley et al., Hidden Technical Debt in Machine Learning Systems (NeurIPS 2015).

Are the essays written with AI?

The content is edited with AI, but it is not unsupervised slop. My process, simplified in three steps, is: research, experiment and validate, then build. Most of the content I produce from it is technical/research notes, diagrams, and case studies—all this stuff is the real experience, but it is not very consumable. So I have to trade time between writing prose and earning experience. I choose the latter because AI can fill the gaps by editing my notes, generating demos, and doing other routines that save me a bunch of time. And believe me, if I had to edit it myself, I wouldn't bother to start this blog in the first place.

If you want to see what I did before AI era, see vski.science and my github.