District is my name for the outreach software I build. The first version has three agents that work in a loop. One finds local shops. One reads their reviews. One writes each shop a real email about their own work.
It never sends the same shop two emails, and each email costs about a nickel. A second version with a new design is on the drawing board.
The engine
District is the standing codename for the cold-outbound systems I build; the first generation shipped in late 2025 and a second-generation architecture is on the drawing board. V1 is three durable agents in a continuous loop, built on crash-safe resumable workflows with per-step retries. A router governs cadence: it keeps at least 75 ready leads, enforces a daily send cap, sleeps, and resumes. A scout walks the town-by-business-type space deterministically, pulling listing metadata and up to 100 full-text reviews per business, deduplicated at the place level. An outreach agent finds and validates contacts, writes per-lead copy grounded in the business's actual reviews via structured output, and delivers. Row-level locking keeps two parallel runs from claiming the same lead.
Why build instead of rent
The first outbound stack was rented: Apollo, Clay, Instantly, Smartlead. It worked, and every layer of it was somebody else's subscription and somebody else's ceiling. Rebuilding it as owned software brought the unit cost to about five cents per personalized email and put the whole loop under one roof.
Owning it meant owning deliverability too, the part the rented tools hide. That layer was run by hand: DMARC policy, domain warmups, scoring every template before sending, and a full recovery after a volume bump tripped a spam spiral. The scar tissue is the point; the system that came out of it treats sender reputation as an asset with a budget, not a setting.
The voice sibling
The voice sibling re-derives the architecture for telephony, where batch acceptance means nothing and the post-call webhook is the only source of truth. It carries a compliance layer up front: consumer-protection rules encoded as a queryable eligibility policy, with carrier line-type evidence per number, hard blocks for mobile and personal lines and excluded states, and recording-consent friction tracked as its own category.
Questions about this system, or the problem it could solve for you?
Discuss your project