Production proof

An AI concierge for real estate — what running one in production taught us

A California coastal platform let an AI advisor sit on its own listings. Real buyers, real inventory, real after-hours traffic — and the lessons that only show up in production.

A California coastal listings platform — Anonymized by agreement — real buyers, real inventory, live in production.

96 / 96 / 100 / 100

Lighthouse, live production (median-of-3)

Performance · Accessibility · Best Practices · SEO.

76 bytes

First-paint media payload

A 40.5MB master sat in the critical path before it.

SSIM ≥ 0.97

Quality gate, enforced in CI

Every re-encode must re-pass the production filter stack.

The 9pm problem

On a coastal listings platform, the most valuable inquiries arrive when nobody is awake — evenings, weekends, the first hours of a new listing. The brand's promise was a private advisor; its surface behaved like a database: filters, grids, a form. High-intent buyers arrived, asked questions, and waited until business hours. Sometimes they didn't wait.

The most valuable demand arrives outside business hours
The advisory promise stopped at the copy — the surface was a database
That gap became a build spec

What went live

A conversational advisor on the platform's own listings — not a chatbot bolted on, not a demo on toy data.

Streaming dialogue with live property cards — matched homes appear in the conversation; a detail drawer and a voice guide follow the same thread
Two lanes, by design — a fast hosted model streams the conversation, a separate lane parses intent; they fail independently
Honest degradation — if a provider key is missing, the surface says so; no silent wrongness
Media that paces itself — first paint is a 76-byte inline placeholder that owes the network nothing; image rungs climb 960w → 1920w, video loads when the browser is idle, data-saver devices stop at the poster
A measured floor — live production Lighthouse, median-of-3: 96 performance · 96 accessibility · 100 best practices · 100 SEO

What production taught us

Five things that only show up when real buyers and real inventory meet the system.

1 · Honest degradation beats hidden failureWhen a provider key was absent, the surface stated it. Buyers forgive a stated gap; they never forgive a confident wrong answer — and neither do the platforms staking a brand on it.

2 · Streaming and intent are different jobsOne lane keeps the conversation human; another reads the actual need. Fusing them makes both worse — the reply gets slower while the understanding gets shallower.

3 · Media weight is a product decisionThe luxury feel is the pacing: weight climbs only after attention arrives. A 40.5MB hero never loads for someone still deciding whether to scroll.

4 · Quality is a tested invariant, or it is a vibeEvery video re-encode must re-pass SSIM ≥ 0.97 through the production filter stack, enforced in CI. If the gate cannot fail the build, it is not a gate.

5 · A client re-theme should be two knobsTwo-tier color tokens, guarded by tests — because the brand will change, and the build should not have to.

The patterns behind this, for builders

The numbers

Measured on the live production surface, not in a lab.

Measure	Before	In production
First-paint media payload	40.5MB master in the critical path	76 bytes inline; the ladder climbs after
Hero video weight	40.5MB	6.8–18.6MB, codec-negotiated by device, SSIM-verified
Lighthouse performance	unmeasured (local 66)	96, live production, median-of-3
Quality gate	none	SSIM law in CI — re-encodes must re-pass
Test battery	none	64 unit/structural + 6 Playwright e2e + Lighthouse floors
Client re-theme	CSS sprawl	two token knobs, guarded by tests

If you run a real-estate platform

The concierge runs on the client's own listings, in production — we do not demo on toy data. The same motion applies to any platform whose search box should be a conversation: answer in the buyer's words, show real inventory, degrade honestly, and measure the floor instead of promising it.

Fixed scope, fixed price — a named acceptance test closes the engagement
One workflow live in 30 days, on your data
You own the code

How an after-hours desk is built — and what it costs

Running a listings platform? Write us — we reply in writing, with a fixed scope and a fixed price.

Write us ← All work