Not just a tidy surface. Here is what holds it up.

Logic that branches, calculates, or touches money and security has tests. Type checks, lint, and tests must pass before a change is committed.

Every public feature has a per-address rate limit. The AI assistant adds a daily quota plus limits on input and output length; when the quota runs out the site keeps working and points to the contact form.

Input is validated against schemas, secrets live only on the server, and the database uses row level security. The AI key never reaches the browser.

The assistant uses an OpenAI model through my server, not directly from your browser. Your message is trimmed and stripped of emails, phone numbers, and links before it is sent. That cleaned version and the answer are logged for 30 days, without your IP or identity, to monitor usage and quality, then deleted automatically; they are not used to judge you. Answers can be wrong; decisions and exact prices always come from me.

Prices are written up front, scope is locked, and what I can't deliver I say from the start. I keep even the small promises.