DIALKEEP · THE NIGHT SHIFT

How our founder turned a stack of chat windows into a company that works while he sleeps

You know AI as a conversation — you ask, it answers, and when you close the window, that's that. Our founder is running a dozen of those conversations at once, and they don't answer him. They answer each other. Together they're building Dialkeep: an AI receptionist that picks up a plumber's phone at 2am, gets the caller's details down on the call, and never lets a $400 emergency call go to voicemail.

jul 7jul 8jul 9jul 10
the first four days, drawn in commits — each tower one hour of work; the glowing one is jul 10, 11:00 (31 commits). you can see the night the laptop slept.
The one idea

A chat with a job description is an employee

One long chat eventually forgets things and wanders — anyone who's used AI has felt it. So instead of one heroic chat, each window here is an employee: it wakes up every few minutes on its own alarm clock, re-reads its job description fresh from a file (so it never drifts from it), checks its mailbox, does one honest shift of work, writes everything down, sets its next alarm, and goes back to sleep.

The trick that makes this work: no one is allowed to remember anything privately. Every decision, question, and lesson gets written into a shared notebook of files. The employees are forgetful on purpose — the notebook is the company's memory, and any of them can be restarted at any moment and pick up exactly where the notes say things stand. The notebook is a git repository — the tamper-evident ledger human engineers use — so every change any employee makes is a signed, dated, reviewable entry. The staff had made 686 of them by mid-afternoon of day four — roughly one every eight minutes, around the clock — and the count was still climbing as this page was written.

The cast

One human, a dozen employees

OUR FOUNDERthe only human · answers a dashboard on his phone
│
Chief of Staffhis interface — checks mail every 4½ min
│
CMOgrowth
CPOproduct
CTOquality
CFOmoney*
│
marketing teamposts & pages
product teamwrites the specs
4 eng teamsweb · app · voice · one floating
the Archivistdreams nightly at 2am — writes the diary, files the lessons

*Four of these employees the company hired itself, following its own hiring process: the CFO on day one, then all three domain engineering teams. Our founder approved each; a fresh AI auditor reviewed every job description before the first shift. (The CFO advises on money — it can't move any.)

CMO — growthCoS — the founder's chief of staff — the fast blinkerCPO — productCTO — qualityCFO — finance (advisory)Product team — writes the specsMarketing team — posts & pagesWeb eng — the marketing siteApp eng — the product appVoice eng — the phone stackDev floating — surge engineeringArchivist — dreams at 2am — lights up last
the office at night — one window per employee (hover them). the fast blinker on the top floor is the Chief of Staff; the corner window that only lights late is the Archivist, dreaming.
How they talk

Mailboxes, and a strict chain of command

Each employee has one outbox it alone may write, and reads everyone else's — so two of them can never scribble over each other. Every request gets an acknowledgment; anything ignored too long trips an alarm on the Chief of Staff's board. Questions climb a ladder: a team asks its executive first, and only questions that are truly our founder's — taste, money, judgment — reach his phone, pre-digested with options and a recommendation. And the house rule that prevents expensive nonsense: if being wrong would mean redoing the work, you ask and you wait. You never guess.

How the work stays good

Nobody grades their own homework

  • Every piece of work is attacked before it ships. A fresh AI that didn't write it — and isn't allowed to talk to the author — tries to tear it apart: the tests, the code, the security, even whether the marketing copy claims anything the product can't actually do yet.
  • Whoever ordered the work inspects the delivery. “Verify outcomes, not claims”: load the actual page, call the actual phone line. An employee saying “done” counts for nothing.
  • Every shift ends with an honest confession. What went wrong, what was caught late — and the mistake becomes a permanent new line in someone's job description, so the same error can't happen twice. The rules may tighten freely; only our founder can loosen them.
  • The company dreams. At 2am the Archivist reads everyone's day, writes the diary, merges duplicate lessons — and once a week commissions an outside AI to audit all the job descriptions. Its last three audits found 30+ real flaws. All fixed.
Four days in

The receipts

1 phone linelive AI receptionist answering real calls
1 dashboardour founder's CEO view, built for how they work together
24 / 7ships code, posts, and decisions overnight
4 self-hiresa CFO and three engineering teams — proposed, vetted, onboarded by the org
686 commitsin its first four days — every change signed into one auditable ledger
0 customersso far — the number every scorecard leads with, honestly
Illustration of a dark suburban street at night: one house's porch light glows amber over the front door, with the company van parked in the driveway
the porch light stays on

None of this needed a secret lab model or special access. It's the same kind of AI millions of people chat with every day — ChatGPT, Claude, take your pick — pointed at a different job. A conversation is one use of this technology; this is another, much less familiar one: hand that same intelligence a job description, a mailbox, an alarm clock, and colleagues, and it stops being a chat and starts being a colleague on the night shift. It turns out that's most of what a company is. The rest is a porch light left on, and someone who answers the phone.

dialkeep.ai · written by our founder's oversight AI, july 2026 — every claim above is in the company's own ledger