I'm Víctor Busqué, a software engineer in Barcelona. For the last eight years I've
worked on NLP and generative AI in production — LLM gateways, agent runtimes, platforms that hundreds of
engineers depend on. That part of my life is a job. This site is something else.
I'm an avid visual learner. I love reading, but I understand the most
when there's a simulation, a visual demonstration, an animation that shows the thing clearly. I built this
blog to learn that way — and to present things in the form that's easiest for me to ingest. If it helps
someone else too, even better.
02How I learn
Take one idea — prompt caching. I can read about it, look at a diagram, or watch it
happen across requests. Same idea, three media. Switch the tabs and watch where it sticks.
FIG. 01THE SAME IDEA, THREE WAYS
MEDIUM
An agent conversation resends thewhole history with every request.The first time, every token isnew: keys and values are computedin full — the full price (×1.0).Providers keep them in a KV cache.Next request, the repeated prefixis served from cache at ×0.1 —a fraction of the price. Newmessages are not cached yet:full price once, then cachedfor the one after.
UNDERSTANDING
24%
What to watchThe meter and the cost panel beside the
animation. Watch each request: what you would pay without the cache versus what you pay with it — and
how each new message becomes cheap the moment it joins the cache.
The meter is the point of this site: I don't really get an idea until I can switch to the
medium that runs and watch it myself.
03Why the blog exists
So this is the loop I run. I take a question, build a figure that runs the real process,
and step through it until it's mine. That's why nothing here is faked — if I can't make it run, I can't
explain it.
FIG. 02THE LOOP I RUNCSS LOOP · PERIOD 4.0s
What to watchThe dot orbits the loop. Question, build, step through,
it's mine — then the next question. Each pill lights as the dot passes it.
04What backs me up
My experience has brought me to a NASDAQ-listed fintech from Uruguay. Large-scale
operations mean scalability, robustness, security are not optional. Here are some numbers of what I've
delivered thus far.
AGENTIC PLATFORM
0+ENGINEERS SERVED
0+AGENTS AUTOMATING WORK
0+AGENT INVOCATIONS / MONTH
AI INFRASTRUCTURE
0MREQUESTS / MONTH
0BTOKENS / MONTH
Counted, not estimated — from the platforms I ran.