Responsible AI is easy to put on a slide and hard to put in a pipeline. We keep ours honest by making it mechanical: gates in code that a deadline can't argue with.
Why gates beat good intentions
Intentions drift under pressure. A gate doesn't. When the gate lives in CI and blocks the merge, "we'll fix it later" stops being an option, and that's the point.
The gates we never skip
- Provenance, we know which code an AI drafted and a human reviewed. Nothing is anonymous.
- Security review on every change touching auth, payments, or personal data.
- Accessibility verified on real screen readers, not just linters.
- No silent failures, errors surface loudly in dev and degrade gracefully in production.
A worked example
When AI writes a database query, a human checks three things before it merges:
- Is the access scoped to the right tenant or user?
- Is every input validated and parameterised?
- Does it fail closed if a permission check is missing?
-- Row-level security: the database itself enforces the rule,
-- so a bug in app code can't leak another user's data.
create policy "read own posts"
on posts for select
using (auth.uid() = author_id);The safest line of code is the one the system won't let you get wrong.
Responsibility is a feature
Clients don't ask for "responsible AI" by name. They ask for software they can trust with their users and their reputation. The gates are how we earn that, quietly, on every commit.