AI that invents missing evidence is worse than no AI.
When a system cannot verify what a person said, uploaded, or approved, it should ask—not fill the gap with something statistically plausible.
Notes
What I believe about building useful software.
Short arguments from current product work. These are working positions, not permanent doctrine; better evidence is allowed to change them.
“Production-ready” includes the boring failure paths.
Payment verification, idempotency, retries, recovery tools, ownership, and clear error states matter as much as the happy-path interface.
African products should not treat local infrastructure as an integration detail.
M-Pesa, constrained devices, mobile-first behavior, and regional buying patterns shape the product itself. They should influence architecture from the beginning.
A polished demo is not proof of a useful system.
The real test is whether people can trust it with their work, recover when something fails, and understand what happened without calling the builder.