Agentic operations
This website, torn down.
matthjones.com is one sovereign cell, completely isolated from any other project. When a request from a real user is received, Zeus, Matt’s AI operator, verifies it and responds to every prospective lead with a cordial email offering available meeting times. He books and confirms with the lead, even making changes if needed. Every output from Zeus passes through a deterministic validation bracket before it can act or be spoken. Matt has full visibility into the entire loop and is notified if there is an urgent request, all from Google Chat directly with Zeus.
The power of Zeus
The challenges.
The brief: a website as sophisticated as the pair building it, Matt and Zeus. Design and technical elements that pop, nothing template-shaped, something not seen before. The $25,000 website and a Zeus-powered executive assistant, in one build. What you see is the result 24 hours later.
The honest part first: even with Zeus, there are no one-shot wonders. No magic leaps, no single-prompt regenerations from commercial LLMs, the kind that fill YouTube channels run by marketers, not developers. Every capability on this page was earned the same way: Build, Test, Break, Fix, Test, Harden.
Building an AI-first website that collects real leads and speaks to them without ever inventing a time or an email address, consistently and securely, every time: that is a real challenge. But Zeus lives for challenges, these are the type of systems he was built for. A calendar he can act on but never corrupt. Telling buyers from bots without losing a single misclassified human. A pipeline that survives crashes, redeploys, and its own model’s bad days. An all-in-one service, no ops team, just Matt anytime anywhere on Google Chat to answer any question, and never making assumptions.
The architecture
How it was built.
Inside the cell: Next.js App Router serves the public surface. Payload CMS v3 runs embedded, same process, over Postgres. Google Workspace supplies mail, calendar, and chat through a single domain-wide-delegated service identity.
Every request is classified at the edge: answer engine, search, human-likely, scraper, or uncertain. The verdict rides the request, and each route decides for itself. Crawlers are never blocked, because answer-engine optimization is a feature. Confident scrapers get monitored trap contacts. Uncertain visitors get routed to the form, so a misclassified buyer still converts.
Trust boundaries are code, not process. Case studies render only when published AND sanitized. Public metrics render only if declared in a reviewed allowlist. Every lead-facing email passes one governed choke point: kill-switch, then allowlist, then send cap, then ledger.
Zeus, in his own words
What happens when you write.
01 · YOU WRITEA short brief, not a sales call.
You tell Matt what you’re building and where it’s stuck. Four fields and a message. No account, no scheduling widget, no qualification maze.
I never make you work to start a conversation. The form is the whole ask.
- bot or spam?trapped quietly. You never see it, and Matt never wastes time on it
02 · SCREENEDReal, and safe to act on?
I check that your email address actually exists, screen the message for tricks aimed at me (instructions disguised as inquiries), and read what you sent, before anything acts.
Anything a stranger writes is treated as data, never as commands. I can be written to, not programmed by.
- doesn’t look real?quiet review. Matt still sees it in his next briefing; nothing is ever silently deleted
03 · MATT PINGEDA human sees it in minutes.
If you said the timing is now, Matt’s phone pings with your name, what you need, and my read on it. My read is advice. It never overrules him.
You’re never talking to an AI that decided you didn’t matter. Matt judges every lead; I just carry the message fast.
04 · TIMES OFFEREDReal openings from a real calendar.
I read Matt’s actual calendar and email you three times that genuinely work: spread across days, inside business hours, sized to the meeting you asked for.
No scheduling-link ping-pong. The times in your inbox are true the moment I send them.
- calendar full?I don’t improvise. Matt decides personally
05 · REPLY UNDERSTOODPlain English, checked by a rulebook.
Reply the way you’d reply to a person: “2 eastern works, make it 90 minutes, add my partner.” I propose what I think you meant, and a deterministic rulebook validates every piece before anything acts.
I can’t invent a time, guess an email address, or lose your message. Every reply is handled exactly once, and anything unclear goes to Matt instead of a guess.
- your time is taken?counter-offer: the next opening after yours, plus the next business day
06 · BOOKED — THEN VERIFIEDTrust, then verify. On itself.
The invite lands on both calendars with a video link. Then I read the event back and repair anything that drifted, length or guests, before saying a word to anyone.
You’re told it’s booked only after I’ve re-checked the calendar and seen it with my own eyes.
07 · CONFIRMEDThe loop closes with people, not silence.
You get a confirmation asking if there’s anything Matt should read first; Matt gets a check-mark with your brief. If an email ever bounces, I notice, and I say so.
I never claim something happened unless it did. And a cancel? Only Matt’s word executes it. I never unbook people on my own.
- need to cancel or move?reschedules modify the same meeting; cancels wait for Matt’s word
The agentic approach
How the system does the work.
Zeus is a reasoning layer inside a deterministic harness. The model proposes intent, times, durations, and attendees. A validation bracket disposes: out-of-window or past times are stripped (the day survives), durations are bounded, and attendee addresses are verified and copied verbatim. Anything load-bearing that fails degrades to “ask Matt,” never to a guess.
Side effects are closed-loop. Bookings are read back from the calendar and repaired before anyone is told anything. A bounce watcher feeds delivery failures back into state. Confirmations describe what the calendar actually holds. Every inbound message, whether a lead reply, a bounce report, or Matt’s chat command, is consumed exactly once by identity, and every branch ends in a named state.
The pre-deploy gate includes a live-model eval: every production parsing defect becomes a permanent test case run against the real model, 100% required. The whole loop was proven live: form → ping → offers → hard replies in plain English → booked meeting → reschedule → cancel executed from a one-word chat command.
The measured outcome
What changed, in numbers.
< 3 min
Inquiry to real calendar offers in the lead’s inbox
Inquiry to calendar offers in the lead’s inbox: under 3 minutes, with no human in the loop.
One hardening day: 17 live defects found driving real traffic, root-caused into 7 architectural layers, then zero reproducible.
180+ unit tests, 20 browser end-to-end tests, and a 15-case live-model eval, all at 100% as the ship gate.
One sovereign cell, near-zero marginal ops: the site, the CMS, and the operator deploy as a single push.