Serhii Zubkov

Node.js / Python & AI Engineer

3 articles

I studied physics, but I've spent my whole career writing software. I started freelancing in 2012 and worked on many different projects for many different clients. In 2020 I moved to Node.js, and it has been my main backend platform since then.

Today I maintain real production projects built with NestJS and NextJS, running on PostgreSQL, Redis, and RabbitMQ. On ProdStories I write about what that work actually looks like: the architecture decisions, the queues and caches that misbehave, and the fixes that held up in production.

In 2024 I started working seriously with AI and machine learning. My background helps here: I studied at the Faculty of Physics at Zaporizhzhia National University from 2005 to 2011, and the math from those years is finally paying off. I'm now learning Python, LangChain, and LangGraph, and exploring machine learning, data engineering, and computer vision.

I write about this part as a learner, not an expert. You'll see what I tried, what I misunderstood, and what finally worked, so you can skip the dead ends I already hit.

Machine Learning Math

How noisy is an LLM reviewer? We scored one post 30 times

We ran csn-review v0.0.3 thirty times on the same article. Proof barely moved (σ = 0.084), Originality spread over 2.5 points (σ = 0.68). Here are the raw scores, the standard error, the Gaussian curves, and why our old "about one point of noise" rule was wrong for both scales.

How we use AI to check posts before publishing

Every post on ProdStories passes two AI skills before it goes live. One decides whether the post is worth reading, the other helps the right reader find it. We don't ask who or what wrote the draft.

Building an AI reviewer skill that scores developer blog posts

csn-review skill scores a developer blog post on three scales before it goes live: Proof, Originality and Author Authority. Here is how it works, what broke when we ran it on our own posts, and what is still weak.