The models are built. The last mile is not. This series breaks down the systems I have running in production, what they replace, where humans stay in control, and the receipts behind the claims.
Every system in the series uses explicit rules, a human gate for consequential actions, and a record that can be audited after the fact.
Monitoring, extraction, comparison, drafting, and follow-up at a volume people cannot sustain.
Money, facts, promises, strategy, and actions that are difficult to reverse.
Inputs, proposals, tool calls, approvals, final actions, outcomes, and operating cost.
A production AI system needs source data, validation, approval gates, and an audit trail built around how a product or service is delivered.
Marco is an outbound AI agent that begins with sender health, deliverability gates, and signals before it writes a message.
Hunter runs persistent PPC operations across Google and Microsoft Ads, but meaningful changes stay governed by written rules and human approval.
Oscar is a self-hosted YouTube operations agent that turns analytics, SEO, captions, and channel tools into a plain-language interface.
InspectPilot pairs inspection scheduling with a compliance engine built on roughly 11 million Los Angeles inspection records over 13 years.
A field note on replacing selected SaaS tools with self-hosted systems, where ownership saves money, and where it creates new operating costs.
Behind the Watt is an open evidence database that uses AI to extract claims from public records while preserving source-level provenance and human approval.
A production pattern for AI systems: let software handle persistence and volume, keep consequential decisions with people, and log every action.
A one-week audit maps the inputs, rework, approval points, and cost. The result is a narrow sidecar with a baseline your team can verify.
Book a workflow audit