Context
A consulting site needs to demonstrate the craft it sells. So the site is the demo: don’t take my word for it, you’re using my work right now.
This site and its agent
The contact pipeline you’re using right now: an AI agent that answers questions about the work, qualifies fit, reports to me, and books the call. Built in place of a contact form.
One operator, spec to production. Headline number in progress.
The problem
A public endpoint that ingests untrusted visitor text and can send email and write to real systems is the sharpest risk profile a small site can carry.
What I built
- A streamed AI agent over a public-safe knowledge base, with visitor input treated as untrusted data, never instructions.
- A verified-email entry gate before the agent will engage.
- Cross-instance rate limiting and a spend-cap circuit breaker.
- An HMAC-verified booking webhook that mirrors a confirmed call into Notion, idempotently.
- All of it delivered through the same spec, evaluate, and gate cycle I’d bring to your team.
Results
Metric-led results block in progress.
What this shows about working with me
Maps to: Delivery rescue
This is the rescue skillset in miniature: hostile-input thinking, hard gates, and honest calls about what should be AI (the agent) and what should be plain software (everything else).