I don't manage,
I ship.

The problem. Inbound came from three places (web forms, event lists, Apollo activity) and reps spent their day cleaning contact data instead of calling people. No system turned a raw inbound signal into a rep-ready, enriched contact automatically.

What I built. End-to-end inbound automation anchored on an AI agent (Claude Sonnet) that resolves company identity from an email domain and chains directly into ZoomInfo enrichment, then routes clean, enriched contacts to the right rep. Web inbound, event contacts (400+ event leads, a $38M+ TAM), and Apollo activity all flow through it. I designed the CRM data model from scratch across 4 core objects with bidirectional relationships and shipped 600+ documented changes in 90 days with zero onboarding.

The failure that taught me the most. Before the attribution chain existed, $10.91M in marketing-sourced pipeline sat in the CRM misattributed, invisible to the team that earned it, because the early system assumed source data would arrive clean. It never did. The fix was not cleaning records by hand, it was building inheritance: a Contact to Lead to Account to Opp attribution chain across 4 production automations with a standardized 8-value taxonomy, so the truth propagates instead of being re-entered.

Also in this system: a 4-automation bidirectional Apollo integration, 6 stage gates with a Slack bot that DMs reps their exact missing fields, and the attribution chain above.

The problem. A raw construction TAM is not a market. You have to separate real equipment buyers from dealers, media, and design-only firms, find a real commercial moment, and then actually send something a real company's reputation can stand behind.

What I built. A gate-first qualified TAM of 631 US heavy-civil contractors (the qualification layer rejects dealers, design-only firms, and media before anything downstream treats them as prospects), a Fit Score built on disqualification logic and public award evidence, a reusable ICP Fit Gate function that takes a plain-English ICP as input, an n8n daily permit watcher with stable dedupe, and a full outbound loop: Clay list to n8n send engine to Gmail to recipient, then the reply back through a watcher into Clay.

The failure that taught me the most. The first real reply almost slipped through. My watcher filtered case-sensitively for "Re:" and Gmail delivered "RE:", so the loop dropped it. I made the filter case-insensitive, republished, and the next poll caught it automatically. A bug found in production, fixed live. That is the difference between a loop that could theoretically work and one that actually closed. The reply came from an equipment leader at a top ENR heavy-civil contractor, who walked through exactly how his team decides rent versus own.

The ethics moment. Every draft had a value line offering an "anonymized 600-contractor study" that did not exist. I cut it before sending and replaced it with something true. "Offer value" is not a line you add to an email; it has to be true before you hit send.