Notes and case studies on machine learning, AI systems and engineering practice.
In previous articles, I wrote about collecting data from PLCs to build time series. But a couple of months ago, a new consumer appeared — the MES production management system — and things started getting messy.
Recently I got a very practical engineering task: build an API for several industrial glue units.
In manufacturing, we have a typical factory problem: there is a lot of documentation, it comes in different formats, some of it is in Russian, some in English, some in Chinese, and you need to search through it quickly and preferably without mistakes.
In my previous article, I wrote about how I cleaned PLC signals, removed leakage, separated real precursors from the usual machine modes, and generally tried to understand what exactly the model found.
So, we have a few thousand signals from the PLC, and the first thought is: great. A lot of data means the models will have plenty to choose from. But it turned out to be almost the opposite.
The ML model starts with the question: do we even understand what our equipment is giving us?
As counterintuitive as it may seem, an ML model doesn’t start with Python. It starts with the question: what do we even consider a defect?
If you have a lot of data (whether you're collecting it or working with big data), you need not only to store it but also to display it properly.
Continuing the series of articles "ML at the Factory." Before building models, forecasts, and pretty graphs, you need to learn to understand what actually comes from the equipment.
I recently started working as an ML engineer at a factory, and I quickly realized one simple thing: a model is not just a notebook, features, and metrics.
When I started getting into embodied AI, I thought I would quickly bump into neural networks, datasets, and model architectures. But reality turned out to be different: first, I bumped into the factory.
There's a feeling that development has gone a bit crazy over the last couple of years.
A couple of months ago, I announced a new project.
Recently, I was building a small OCR service for recognizing Russian passports.
About a year ago, I started working on a small commercial product with a telling name, Zada4kin...
How it works and why prompts are critically important here (unlike LLMs)
The other day, I decided to Google my credit history. Just for general knowledge. Here's what came out of it.
Nowadays, almost everyone has heard of the trendy term "vibecoding." But can we call everyone who generates code using neural networks a vibecoder?
Here, I will detail what I encountered during the creation of my N-GPT RAG LLM Service.
You encounter this every day, but you probably don't fully understand what it is.
I got tired of manually sifting through irrelevant job postings and put together a simple tool that parses job listings, filters them through an LLM, and leaves only those that really match my stack.
Why ML engineers often underestimate it. How chunk size affects the quality of the entire LLM+RAG system.
A simple explanation of why taxi prices are constantly changing: demand, supply, coefficient, adjustments, and the role of machine learning...
How machine learning algorithms decide which videos and posts you see in your feed.
A simple explanation of what machine learning is and how artificial intelligence really works.