Notes
Occasional build notes — decisions, trade-offs and lessons from shipping. Mostly agent-assisted development, system design, evaluation and product scoping.
Environments are the unit of reproducibility, not code
Most "works on my machine" failures with coding agents are environment drift, not model error. A note on why I made the environment the artefact.
When real-time data lies, prefer stability over freshness
A predicted arrival that jumps backwards destroys trust faster than one that is a little stale. Notes on reconciling noisy multi-source feeds.
Scope an AI feature by what happens when it is wrong
The useful question in product scoping is rarely "can the model do this". It is "what is the cost of a confident mistake".