Duncan AndersonIndependent systems
Back to selected work

Governed AI operations · Production operations system

The Lineup AI Operations

Four specialized AI workers monitor the business—but they cannot quietly change production.

The result

Recurring AI work became a visible operating process with clear owners, blocked states, approval boundaries, and durable proof.

Operational proofFounder, system designer, and operator

Instead of treating an agent run as proof of useful work, the control plane follows an inspectable sequence: request, finding, one selected action, approval when required, and evidence of what actually happened.

  1. 01Four specialized operating lanes in one control plane
  2. 02Scheduled and manual requests with durable run and action records
  3. 03Approval and reliability checks expose incomplete or blocked work
Operational proofEvidence view

The Lineup

AI operations control plane

Four operating lanes
Data and pricing
Models and forecasts
User growth
Executive oversight

Selected action

Freshness exception requires review

01Finding recorded
02One action selected
03Production change held
Human approval required

Faithful reconstruction of a private operations dashboard · no production data shown

What changed

Before

  1. 01Recurring monitoring lived across scripts, dashboards, and operator memory.
  2. 02A completed agent run did not prove that an action was delivered or resolved.
  3. 03Risky writes needed a reliable stop before production systems changed.

After

  1. 01Four domain-specific operating lanes can run on schedules or explicit requests.
  2. 02Material findings and actions move through a durable, visible lifecycle.
  3. 03Protected changes stop for approval, while delivery claims require verifiable evidence.

What I owned

I built an internal control plane that schedules recurring work across data, predictive models, user growth, and executive oversight—then records findings, proposed actions, approvals, and outcome evidence.

  • 01Control-plane architecture and durable action model
  • 02Data, model, growth, and executive operating loops
  • 03Approval gates, blocked states, and fail-closed behavior
  • 04Operator dashboard, reliability checks, and automated tests

Where trust was designed in

Runs are not presented as outcomes. The system separates monitoring, proposed action, completed work, and verified evidence. Consequential changes remain approval-gated.

Technical detail

PythonNext.jsPostgreSQLSupabaseAutomation orchestrationApproval workflowsReliability gates

Have a similar operating problem?

Send the current workflow and the point where it keeps breaking. I will tell you whether there is a practical paid first phase.

Discuss your workflow