Notes on ML, computer vision and network automation.
Together with Kirill Pishchalnikov I spoke to young engineers about building a personal brand inside a company. My part covered two levels of visibility, six common mistakes and a minimal routine that even an introvert can keep up.
A link incident can last hours or days, while a plain script does not even survive a restart. Three design decisions that make incident automation trustworthy enough for production.
In a conflict-monitoring agent the model takes part in only one step out of ten. Here is why the other nine are better left to ordinary code.
A single number in a forecast is not enough for planning. This post covers P10/P50/P90, per-sector normalization and how a forecast helps decide whether to buy more capacity.
Master's thesis on a PPO agent for MiniWoB++ with 16 GB of VRAM, an adaptive curriculum and graded reward. What worked, what did not and what the results were.
How an LLM agent uses an MCP server to find a failed Juniper port and restore it without an operator.
Why segmentation masks help a transformer match people while ignoring the background.