What a deployment looks like when it's built to be handed over
This is the reference engagement behind the 0tomait method: a freight brokerage drowning in billing paperwork, two shipped automations, and the documentation discipline that keeps them alive after the engineer moves on.
note: Meridian Freight is fictional, built as the worked example for the published method. The discipline shown here is the real product, and every manifest artifact on this page sits in the public repository where you can read it in full. The tool views are stylized recreations, marked as such.
The setup
Meridian brokers truckload freight between shippers and roughly 900 carriers. Nothing gets invoiced without a signed proof of delivery on file, and matching that paperwork to loads by hand kept billing three to nine days behind. Two staff spent most of their week rekeying documents out of a shared inbox while days-sales-outstanding drifted past 45.
Learn what the code can't tell you
The first weeks produce no code. They produce the context that makes every later decision safe: what words mean in this building, who decides what, and which data quirks will silently corrupt a naive integration.
"POD" means proof of delivery. The signed document that gates all revenue. Newcomers confuse it with the BOL, issued at pickup. An AI agent would confidently read "pod" as Kubernetes. The glossary exists so neither mistake happens.
Load numbers reset every fiscal year. Joining on load number alone silently mixes a 2025 load with a 2026 one. The safe key is fiscal year plus load number. This one line in the manifest prevents a data corruption bug nobody would trace for months.
Dana decides, Priya gates access. Anything a customer could notice needs Dana, the VP of Operations. All system access flows through Priya, an IT contractor reachable two days a week. Miss that fact and every schedule slips a week.
Automation on the client's rhythm
Two flows went live. One watches the billing mailbox and files carrier paperwork against its load. Here is what that looks like in the field, minutes after the morning trucks report in:
unread mail is the review queue: nothing is ever lost, it stays visible until a human clears it
The other flow runs on Meridian's twice-weekly billing rhythm and raises the invoices. Both fail closed: an item that errors is skipped whole and reported, never half-executed. The operator's interface is the run summary:
every skip carries its reason, and skips self-heal: fix the cause and the next run picks them up
Invoices over $25,000 stay human-approved, because the whole project was approved on that condition. And the manifest records how the intake flow fails, honestly:
"on_failure": "Retries three times with backoff, then leaves the email unread and emits a CloudWatch alarm. Unread mail is the safety net: no POD is ever lost, it just stays in the inbox for a human.", "who_notices": "CloudWatch alarm to #meridian-alerts, which Marcus watches. Nobody watches it on weekends."
That last line is deliberate. Weekend coverage was priced, declined by the client, and recorded. Honest gaps beat comfortable silence.
Decisions keep their reasoning
Every consequential choice is recorded with the alternatives that lost and the condition that should reopen it. Here is one: the mailbox is polled every three minutes instead of using push notifications, which sounds backwards until you read why.
"decision": "Poll the billing mailbox on a 3-minute delta query rather than subscribing to change notifications." "context": "Subscriptions expire every three days and must be renewed by a running process. Meridian has no weekend coverage, so a renewal failure on a Friday would silently stop POD intake until Monday. Polling costs trivially more and degrades visibly." "rejected": ["Graph webhooks: invisible single point of failure", "IMAP: basic auth disabled by tenant policy"] "revisit_when": "Meridian establishes real on-call coverage."
A successor who prefers webhooks now knows this was a considered trade, what would make it wrong, and when to revisit. Without this record they would "modernize" the trigger and reintroduce the exact failure it avoids.
Staffed roles and a talking deployment
Every human step has a named owner and a family tree. Meridian's queue has already survived one staffing change, and the record shows both the failure and what it taught:
Tina Okafor · queue-operator
Left Meridian. No walkthrough existed yet, so the queue reverted cold to her supervisor and a two-week backlog followed. The pairing script below was written as a direct result.
Marcus Bell · queue-operator
The next transition follows a procedure: close the tenure, open the new one, deliver the walkthrough.
The tooling then reads the document back and says what a polite consultant might soften. This is the actual validator output for this deployment:
! warnings (2) systems[3] (accounting): System last verified 93 days ago. flows[0] (pod-intake): nobody is recorded as noticing this flow failing. Record it as a risk if genuinely unmonitored. · notes (4) flows[0] (pod-intake): every human step runs through one person (Marcus Bell). The family tree shows no backup — a planned absence becomes an outage. systems[0] (tms).lineage[1]: Priya Chandran inherited 'administrator' with no walkthrough on record.
The deployment also talks back. Operators' complaints become field reports with a tracked lifecycle, so the system surfaces its own gaps:
"The handwritten ones take three or four times longer than anything else in the queue. Most mornings they are the whole queue."
"I cannot see which delivered loads are still waiting on paperwork without asking Marcus. Before the automation I could at least see the pile on his desk."
A duplicate POD slipped past dedup when a carrier re-sent it with a cover sheet. The resolution is written down where the next engineer will find it.
Finally, the whole manifest compiles into a briefing an AI coding agent can maintain the system from. Your constraints become its hard rules:
## Hard rules — never violate these - Do not write to the TMS from any deployed flow. Read only, and only against the replica. (technical) - Invoices over $25,000 must remain human-approved. Dana agreed to the whole project on this condition. Removing it without asking her would damage the relationship well beyond the value of the automation. (political)
This is what you'd be holding
Working automation, the reasoning behind it, staffed roles with walkthroughs, and a system that reports its own friction. Maintainable by the next engineer, your own staff, or an AI agent. The method is public; the engagement is yours.