José Neves

LLM-written code tends to default to database locks to deal with concurrency. A small Go and PostgreSQL benchmark shows what that costs: last write wins served 1.5 to 3 times more users than the locking patterns.

An internal AI assistant — named after the Japanese 時 / とき (time, moment, era, opportunity) — autonomously resolved 32.5% of intervened engineering support requests over four weeks. A look at how it's plugged in, its quirks, and where it's actually helping versus where humans still own the work.

After repeated attempts at LLM workflow automation, a first real success comes not from smarter models but from scaffolding — small, scoped tools around the LLM that iterate, validate, and fit into existing workflows without a chatbox in sight.