Field notes

Short observations and implementation lessons captured while building.

2026-08-06KPI Math Belongs on the ServerA KPI that lives in a spreadsheet formula is only as correct as whoever last edited that cell. Move the calculation into the system that owns the underlying data, and correctness becomes a property of the architecture instead of a training problem.Engineering
2026-07-28Meet the Data Where It LivesAsking a customer to re-enter data they already keep somewhere else is a design failure, not a convenience gap. Integrating with the system they already run is a product decision before it is a technical one.Product
2026-07-20Write Decisions for Whoever Inherits the CodeCode shows what a system does. It rarely shows what you deliberately chose against, or why. A few lines of decision record saves the next person from re-solving a problem you already solved.Systems
2026-07-08Local Inference Is About EconomicsRunning models on your own hardware is usually sold as a privacy move. The more honest reason, most of the time, is cost and latency control during evaluation.AI
2026-06-24Provider Adapters Are Cheap InsuranceWiring your app directly to one model provider feels fast. A thin adapter behind a stable interface costs a little now and saves a lot the day pricing, quality, or availability changes.Engineering
2026-06-03The Kill Switch Is a FeatureIf you cannot turn an AI off in one move, you do not really have it under control. Treat the off switch as a product requirement, not an ops afterthought.AI
2026-05-20One Confidence Score, Two LanguagesA model's confidence is not a fixed currency. The same number can mean different things across languages, so a single global threshold quietly under-protects your second language.AI