TENS LAB

A lot of what happens in AI today is built on conventions that simply became normal over time. A model works a certain way, a system is built around it, everyone uses the same approach, and after a while it becomes difficult to tell whether something is actually necessary or whether we just got used to doing it that way.

That is the part I find interesting.

I like understanding how these systems actually behave underneath — what is happening before the tokenizer runs, where latency and memory are wasted, and what changes when you treat context as a compiled system primitive rather than an arbitrary string.

Sometimes that leads somewhere useful: a deterministic compiler that eliminates 70% of prompt token overhead, an inference-native storage pattern, or an architectural idea that holds up under benchmark scrutiny.

Sometimes it doesn't. And that's fine too — figuring out why an assumption breaks is half the work.

TENS LAB is where I document the systems, compilers, and architecture work that comes out of those rabbit holes.