What I've shipped.

A missed call at a local trade business is usually a lost job — the caller dials the next plumber within minutes. Voicemail doesn't fix it, and for immigrant-owned businesses in Vancouver the caller may switch to Persian mid-sentence, which most voice agents can't follow.

One agent definition per business, kept as a config file rather than a hand-tuned dashboard — greeting, services, approved prices, service area, emergency rules, transfer number. A ten-section system prompt is built from that config, with hard rules the agent can't talk its way past: never quote an unapproved price, never confirm a booking the calendar tool didn't return, admit to being an AI when asked. English and Persian switch mid-call, and every call is summarized by Claude into the same Supabase the site's chat leads already land in.

The web channel is live — the button in the corner of this page is the product, not a mock-up. Phone is wired but waiting on carrier setup, so the honest status is half-shipped. Guardrails are tested against a written adversarial matrix (price fishing, discount pressure, prompt injection, angry callers, mid-call language switches) plus 26 automated tests; a run costs roughly twelve cents a minute, which is the number that decides whether this is a business or a demo.

I needed a working demonstration of the way I claim to build — bilingual, honest about method vs. history, and with an AI diagnostic tool that produces something specific enough to defend.

A Next.js site with an AI Readiness Audit powered by Claude, a document-grounded assistant (Nova) using Voyage embeddings, dual payment gateways, and a full admin panel — deployed on Vercel and Supabase.

The audit engine generates personalized reports in ~30 seconds with a schema-validated fallback path if the model is unreachable. Monitoring, cost limits, and error handling are baked in from day one — because the site itself is my day-300 test.

Local businesses juggle reviews and DMs across Google, Instagram, WhatsApp, and Telegram with no single place to keep replies on-brand and timely — and most AI reply tools either sound generic or send something risky with no human checking first.

A unified inbox behind one channel-adapter interface, so live integrations and still-pending ones (Google/Meta approvals take weeks — Telegram needs none) share the same code path. Claude drafts replies in the business's own voice with structured-output validation, a do_not_say filter, and a prompt-injection wall around customer content — nothing sends without approval.

Same discipline as the audit engine above: row-level security from the first migration, a central logger with auto-redaction, and the public interactive demo below, protected by three layers of rate limiting so the cost stays predictable before there's a paying customer.