THE LEAD ENGINE
A shadow-first outreach machine that won't guess
The Problem
Most lead-generation automation is built around one metric: volume. Scrape more, message more, book more. Sending to bad addresses, anonymous handles and probable noise gets written off as the cost of doing business. Spam filters, domain blacklists and burned sender reputations are costs the tools don't track and the operators don't feel, until they do.
I needed something different. The targets were independent trade businesses: real companies, real owners, real inboxes. If a cold message from an automated pipeline lands badly with one of them, it costs a relationship and a reputation that a retry can't buy back.
So the engine had to err on the side of silence. I wanted it to report “uncontactable” rather than send into the void.

Chapter 1 — The Pipeline That Refused
The first version pulled from social-signal sources: community posts where small business owners describe operational pain in their own words. It produced 42 leads and wrote 126 personalized outreach drafts for them. The drafts were specific and relevant, and each one used the exact phrasing the prospect had posted.
Then the pipeline reached the contact step and stopped.
Every one of those 42 leads was an anonymous or pseudonymous poster. There was no verifiable business name, no domain to query and no address to validate. The system could draft well, but it was pointed at a list it couldn't honestly call contactable. So it didn't. It marked the list UNCONTACTABLE and queued the drafts. They aren't waiting to send. They're waiting for the day the discovery layer can find real contacts behind them.
126 drafts written. Zero sent. The pipeline called the list what it was, and that was the design working, not a failure.
That moment shaped the rest of the architecture. In a v1 there's always a temptation to ship the spray. This system resists it structurally: unvalidated contacts have no send path at all, so no policy has to stop them.
Chapter 2 — Named Businesses, Earned Contacts
v2 rebuilt the discovery layer around named businesses instead of anonymous signals. Public map data, queried by craft and trade tags, turns up real companies. A business listed in a public directory has chosen to be found.
A search-API pass extends coverage past what any single map source holds. A sweep of a third metro found 45 more named businesses, with email harvest still pending at the time of writing.
The crawler is polite on purpose. It identifies itself with a custom User-Agent, holds to one request per second per host, and reads robots.txt before touching any page. Several sites declined, so the crawler honored their exclusion directives and skipped those businesses. I don't treat that as a formality. A scraper that ignores robots.txt has started the relationship on bad terms.
Email addresses come only from a business's own website, from its contact pages and footer, and every address is stored with its source URL. The system knows exactly where it found each one. Every harvested address also gets a DNS MX record check. If the domain has no mail infrastructure, the contact is flagged before it reaches the drafting queue.
Every email address in the CRM carries its source URL, so the system can prove where it found each contact.
Ethics as brand protection. The rate limit, robots.txt compliance and provenance records aren't there for show. They're what let me explain the system to a prospect. If a business owner ever asks “how did you find me?” the honest answer is: their own contact page, fetched once, politely, with the URL on record. I can defend that answer. I couldn't defend a bulk-scraped dump with no provenance.
Craft Details
Provenance chain. Each CRM contact record stores the source map or search result that named the business, the page URL where the email was found, and the MX validation result. The chain can be audited end to end, and every contact traces back to its public source without any third-party data broker.
Draft quality bar. Drafts reference the prospect's business category and observed operational context. The personalization comes from public signals, not inferred private data. The drafts are specific enough not to read as templates, and everything in them is information the business made public itself.
Numbers on disk (2026-07-14). 18 named businesses discovered across two US metros, 5 MX-verified email addresses confirmed deliverable, and 45 more businesses found in a third-metro search-API pass with harvest pending. CRM total: 155 contacts, 126 queued drafts.
Stack
Chapter 3 — The Shadow Floor
The engine has no send path, and that's deliberate. Every draft that clears MX validation goes into a queue the operator can see. Nothing leaves that queue without an explicit human grant scoped to a specific batch. That grant doesn't exist yet.
I apply the same rule to every autonomous system I build: built does not mean armed. A capability that can act on its own before it has earned a track record is a liability. The shadow floor is the receipts period. During it, the system has to prove it can produce contactable leads, well-sourced addresses and relevant drafts before it gets the authority to send any of them.
Reply rate: UNMEASURED. Conversion: UNMEASURED. Open rate: UNMEASURED. Nothing has been sent, and these numbers won't exist until the operator decides to open the gate.
155 contacts. 126 drafts. Reply rate: UNMEASURED, because nothing has been sent. That's the point.
The machine is ready and the receipts are piling up. I left it on a cliffhanger on purpose: an outreach system that asks permission before it acts is one you can trust when it finally does.