About
What this site covers, how the posts are written and how to find your way around.
What this site covers, how the posts are written and how to find your way around.
About this site
kozodoi.me is a notebook on machine learning and data science. It collects hands-on tutorials, package overviews, competition write-ups and findings from applied projects, with a strong focus on code that can be read, run and adapted. Most posts started life as Jupyter notebooks, so explanations sit right next to the Python or R code that produces every table and figure.
The tone is practical. Instead of surveying a whole field, a typical post takes one concrete problem – computing channel statistics for an image dataset, accumulating gradients when a GPU runs out of memory, tracking hundreds of experiments without a spreadsheet – and walks through a working solution step by step.
What you will find here
- 📝 the blog with tutorials on PyTorch, tabular data, experiment tracking and algorithmic fairness
- 📁 projects that combine several techniques into complete pipelines, from text readability to molecule captioning
- 🥇 Kaggle notes on competitions across computer vision, NLP, tabular data and time series
- 📚 research notes on credit scoring, sampling bias, fairness and demand forecasting
- 🗣 talk notes – short, talk-sized explainers on recurring ML topics
- 🖥 a guide to certifications and courses for people building ML and cloud skills
How the posts are written
Every tutorial follows the same structure: a short statement of the problem, the minimal amount of theory needed to understand it, and then the implementation, broken into small cells with comments. Figures are produced by the code shown in the post, and data sets are either public or synthetic, so the results can be reproduced on a laptop or a free cloud notebook.
Competition write-ups separate what worked from what did not. Ideas such as pre-training on surrogate labels, test-time augmentation or stacking several CNNs are explained in general terms first, so that they remain useful outside the specific leaderboard they were tested on.
The posts were written between 2020 and 2022, and library versions from that period are used throughout. PyTorch, PyTorch/XLA and the Hugging Face stack have evolved since then, so some function signatures may have changed, but the underlying techniques carry over unchanged.
Site notes
The site is static: there are no accounts, comment forms, newsletters or tracking scripts, and no personal data is collected from visitors. Opinions expressed in the posts are editorial and do not represent any company, university or competition organiser mentioned in the text.