Build · AI front desk, 48-hour build

Man in the Mirror front desk

A three-chair barbershop loses calls every time the owners are mid-cut, which is most of the day. I pitched a missed-call agent, then built the whole thing in 48 hours for an owner demo: a real phone line, live bookings, a cancellation waitlist, and a front-desk dashboard.

Open the live dashboard

What it does

Call the shop line and hang up: June texts you back in seconds.
Text her instead: she recognizes returning clients, quotes real prices, offers real open slots, and writes the booking.
Shop fully booked? She offers the cancellation list. A slot frees up? She texts the first person whose window fits and holds it for a reply.
The dashboard's first number is the one an owner cares about: dollars booked by June today.

Man in the Mirror front desk dashboard with live conversations, schedule, and waitlist — The front desk, live: revenue booked by June, every conversation threaded per caller, the three barbers' schedule, the cancellation waitlist, and upcoming rebooks.

Runs on the business phone line through Zoom Phone webhooks: missed and rejected calls trigger the text-back, inbound SMS goes straight to the agent.
Mock Setmore and Clover adapters return real API shapes, so going live swaps adapters instead of rebuilding.
Booking writes go through a guarded sub-workflow that refuses double-bookings and past times, and recovers with fresh alternatives.
Compliance is enforced in code: STOP, START, and HELP keywords, an opt-out list checked before the agent ever sees a message, and a rule that "cancel" always means an appointment, never an unsubscribe.
Protection limits on every number: reply rate caps, a loop guard against other bots, message dedup, a kill switch in the dashboard, and an error catcher that texts me when anything fails.

How I built it

My execution loop at full speed, in four beats.

Day one: working core. Data spine, seeded clients, the agent adapted from my June build, and an end-to-end booking written through the mock calendar.
Day two: real phone line. Bridged through Zoom's API and live-tested from my own phone. Then the dashboard, lighter than my Man Cave app on purpose, reusing its hard-won lessons about polite polling.
Fix what only real phones find. A session-memory flaw in the workflow engine quietly merged different callers' conversations. Traced, fixed, and later replaced with durable transcript-backed memory after watching context drop once in a mock demo.
Attack my own work. A four-lens audit before the owner saw anything: backend, mobile UX, live customer personas texting the real line, and a business lens, with every finding challenged by two skeptics instructed to refute it.

The audit confirmed real revenue-costing bugs: a race that let two customers book the same chair, and the word "cancel" being read as an unsubscribe. Every confirmed finding was fixed and live-tested the same day. That is why the guarded booking workflow and the opt-out enforcement exist.

What the build taught me: speed and rigour are not opposites, they are a sequence. Build fast on mocks with real API shapes, test on the real channel early, then audit adversarially before the buyer ever sees it. And demo on the owner's actual pain: the demo's opening beat is the exact scenario the owner complained about in the pitch meeting, a call ringing out while he is mid-cut.

Why this matters for selling AI

Diagnose-then-close, compressed: the pitch meeting surfaced the owner's biggest annoyance, and two days later he could text the fix from his own phone.
Commercial structure that protects everyone: free demo on dummy data, live systems connected only after a yes.
Selling AI at any price point takes exactly this: hear a business problem, then describe concretely and honestly what an agent will do about it. I can, because I have shipped it.