Seven years of systems that have to stay up. Now learning the ones allowed to be wrong.

Where I am. Seven years of distributed systems: streams, databases, change capture, circuit breakers, and the pager. Highly scalable, fault-tolerant, resilient architectures — that is the record, and it is the part you can check against employers and dates.

Where I am heading. AI platform, infrastructure and evaluation: eval harnesses and the statistics that make their numbers mean something, judge validation, how a vector index behaves once you start writing to it, what inference actually costs per request, and what an SLO looks like for a system that is allowed to be wrong some of the time. I am learning it and building it in public. None of it is production work yet, and I am not going to write it as though it were.

That is a different ladder from modelling or research, and naming it is the point. Left generic, a research-leaning reader sees a backend engineer who has drifted, and a platform reader never notices I am talking to them. Named, seven years of streams, databases and on-call becomes the asset rather than the thing to explain away.

Anything here with a number in it is measured on one 4-vCPU, 7.6 GB VPS with no GPU, the same box that runs a six-agent system round the clock. Nothing on the ledger has earned that word yet — both rows say preliminary, and they will keep saying it until there is a repro command beside them. The constraint is deliberate. Most of the interesting numbers in this field come from people with a GPU budget, most readers do not have one, and almost nobody is writing for them.