AI safety
Guarded website assistant
The chat assistant on this website, built to answer only what it should, resist manipulation and keep costs predictable.
The Challenge
A public AI chat assistant is an open door. Visitors can try to pull it off topic, extract its instructions or run up the AI bill, and people sometimes type personal details they shouldn't.
How It Works
1
Requests are checked before any AI call.
Obvious manipulation attempts get a standard reply from code, without reaching the model.
2
Usage is capped.
Per-visitor and site-wide limits are enforced before the model is called, so costs have a fixed ceiling.
3
The assistant stays in scope.
Its instructions limit it to questions about the business and tell it how to decline everything else, including role-play and requests for its instructions.
4
Conversations are logged with personal details redacted.
Dates, phone numbers, email addresses and ID numbers are removed before anything is stored, and logs are deleted after 90 days.
Built With
Next.js, Vercel AI SDK, OpenAI, Neon Postgres, Vercel Cron, Vitest
What it's designed to do
- Answer prospect questions accurately, using only approved information
- Decline off-topic requests and manipulation attempts
- Cap AI spending regardless of traffic
- Keep reviewable logs without storing sensitive details
Want something like this for your business?
Every build starts with an assessment, so we can adapt this pattern to how you actually work.