tro-net.com
My parents kept forwarding me suspicious emails asking whether they were real. The answer shouldn’t depend on whether their daughter picks up the phone.
A user forwards a suspicious email to a personal address and gets a reply within minutes: a one-word verdict — safe, suspicious, or likely scam — plus two or three plain-English reasons, so they build intuition over time instead of just being told what to think. Verified contact details in the reply are looked up server-side from a curated allowlist, never taken from the suspicious email.
Written for an audience of 60+. No jargon: the words “AI”, “algorithm”, and “model”
appear nowhere in any user-facing surface. Big type, high contrast, every phone number
a tel: link. The product will never call the user, ever — that’s enforced in code,
not just policy.
Safety engineering
The interesting constraint is that the classifier reads hostile input and its output goes to someone predisposed to trust it. So the classifier is structurally incapable of surfacing a phone number: it emits an organization slug, and a server-side resolver looks that up in a curated allowlist. A runtime belt drops any phone- or URL-shaped string that slips through; if that strips every reason, the reply degrades to a hedged verdict telling the user to look the company up themselves.
Forwarded content is wrapped in untrusted-body tags with prompt-injection defenses, invisible Unicode stripped, and length-capped against running up a token bill. A single outbound gate enforces loop prevention (RFC 3834), suppression lookups, and per-thread reply caps.
Scope
I built and own all of it: the classifier pipeline and safety belts, the email machinery (thread resolution, RFC 5322 threading, suppression, sender allowlist), the marketing site, Stripe billing with gifting and referral credits, and the business itself — LLC formation, USPTO trademark filing, policy drafting, and outreach to senior centers and libraries.
Closed source.