Notes from
the build.
Short write-ups on LLMs, computer vision and evaluation — the things that took me longest to figure out. If one of them saves you an afternoon, it did its job.
How to cut your Claude Code token usage by 20x↗
Claude bills a cached token at 0.1x and a cache write at 2x – a 20x spread on identical content. How context and prompt caching work in Claude Code, the hygiene that cuts always-on context, and a one-plugin guard that blocks cold-cache turns before you pay.
How to debug Python like a pro↗
Debugging in PyCharm is not print-spam. Breakpoints freeze the process, Step Over/Into/Out move with intent, diffs show what changed, and one mute-all trick lets you jump without getting snagged.
What quantization really is↗
Quantization is not magic compression. It is a controlled change in numerical representation that trades precision for memory, bandwidth, and often speed – illustrated with Gemma 4 12B.
Rather talk than read?
I am always up for a conversation about LLMs, applied ML or getting a model into production. LinkedIn is the fastest way to reach me.