Method
How I work
Seven rules that show up in every system on this site. None of them started as principles. Each one was written down after something went wrong or after a measurement disagreed with an assumption.
Rules
- Dry run is the default. Every mutation against an external system (Google Ads, a helpdesk, a mail API) runs in dry-run mode unless explicitly confirmed. Campaigns are created paused. Sends are drafts until a human says otherwise.
- Never assume success. Every API mutation is verified against the response, and every creation is re-read from live state afterward. This rule came from an audit of my own system that found months of silent failures behind a green log.
- A human sits at every irreversible boundary. Approval gates on quotes, outreach, medium- and high-risk ad changes, and helpdesk replies. Low-risk, reversible actions are the only ones that run unattended.
- Facts come from retrieval, not from the model. Any customer-facing claim is checked against retrieved data by a second pass before it can send. If the retrieved data does not answer the question, the system escalates to a person instead of guessing.
- Incidents become controls. Each production incident gets a structured record: root cause, impact, what would have prevented it, mitigations. The mitigation lands as an enforced hook, gate, or test, and the record is carried into future task handoffs as a warning.
- Autonomy is earned with evidence. Agents and automations start at observe, move to suggest, then act-and-report, then autonomous, and each step requires logged evidence: accepted suggestions, zero corrections over a period, confirmed reversibility.
- Measure before claiming. Controlled A/B experiments, holdout groups, a head-to-head benchmark against an outside agency, anomaly detection against 30-day baselines, and a measured noise floor before reading a visual diff. When the data said an idea did not work, the idea was dropped and logged so it would not be re-tested.
Background
Seven years building systems for marketing, sales, and finance teams, first as a marketer who automated his own work, now as an engineer who ships the automation as product. The last three years have been on LLM APIs. I work daily in Claude Code and build the orchestration layer that other agents run inside. I am focused on the gap between the AI tooling companies buy and the output they actually get, and on the operating models that close it.
Roles: Head of Growth Engineering at a private-equity operator, directing a five-person marketing team and serving as technical counterpart to an enterprise software vendor through a year of monthly product reviews. Director of Business Development at a B2B wholesale ingredients business, where I built the operational stack solo. Earlier, Head of Digital Marketing for a multi-brand consumer portfolio, managing $150,000+ monthly ad spend and a four-person team.
What I can show in an interview
- Code walkthroughs of any system on this site, including the risk-tier taxonomy, the mutation verification layer, the RLS policies and cross-tenant tests, the guardian hook, and the two-layer send validation.
- Nightly engine reports, weekly executive reports, and the experiment log.
- Incident records and the controls each one produced.
- The cutover runbook and visual-diff output for the website re-platform.
What is deliberately not on this site
Client, franchise, and vendor names, revenue by client, source code, and screenshots. All of it is available on request, under NDA where the client requires one. The numbers here are the numbers on my resume, and the resume is the superset: nothing on this site goes beyond it.
Contact
andreasengelbertson@gmail.com. New York, NY. Open to engineer-titled roles at AI-native companies: applied AI, forward deployed, growth engineering, and agentic systems.