Lead engine
It finds the right business owners, checks each one against public sources and writes a personal first line for the email.
- 352
- Stranded leads recovered as sendable
- 7 of 10
- Sending domains it caught landing in spam
- 40 of 50
- Inboxes it found with warmup set wrong
The problem
Finding the real decision-maker at a small business by hand took forever. Generic cold emails get ignored.
What it does
- Searches by market and keyword for owner-run businesses that fit.
- Confirms the decision-maker from the business's own sources and drops anything it can't prove.
- Writes a personal opener from public info, held to my voice rules.
- Hands campaign-ready leads to the email tool on a schedule and tests where the emails land.
Biggest challenge
The research agents kept stalling mid-run. Hundreds of half-researched leads piled up and never made it into a campaign.
How we solved it. I rebuilt it as one scheduled job from discovery to campaign-ready lead, with a repair cycle when a step fails. Then it re-checked the stranded pile and recovered 352 of 430.
How it got built
- 92
- PRs merged to production
- 20 hours
- Average time between PRs
- 38
- PRs in the busiest week
- 77 days
- Jul 15 to Sep 30, 2026
v1: drafts only
Wrote partner emails for me to review. Nothing sent on its own.
Research agents
An autoresearch system to find and verify decision-makers. 5,300 lines in one PR.
One specialist lead hunter
Rebuilt as a single lead-hunting agent and deleted 2,300 lines of the old one. Then 24 PRs of fixes.
Personal first lines
A second pass writes a personal opener from public sources, held to my voice rules.
Running at scale
Rolling lead supply on a schedule, split across a pool of AI accounts. Then a long run of reliability fixes.
v2: one job
One scheduled job from discovery to campaign-ready lead, with self-update and a repair cycle.
Inbox check
Replies fell from 8.5% to 1.6%. This skill found 7 of 10 sending domains landing in spam and 40 of 50 inboxes with warmup set wrong.