AI SYSTEM · CASE STUDY

Autonomous Outreach Machine

A full acquisition engine built for a top leading insurance company. A six-account LinkedIn fleet finds decision-makers inside an 18,000-company federal-filings corpus, writes to each one like a human, generates a personalized report per lead — and polices itself with killswitches, budget governors, and a fail-closed do-not-contact ledger.

18,000

companies in the filings corpus

6

sender fleet, warmed & capped

100+

engineering PRs, all reviewed

Autonomous Outreach Machine
Fleet admin — six sender accounts with daily caps, 14-day warmup, pause/resume, and one-click onboarding links
Fleet admin — six sender accounts with daily caps, 14-day warmup, pause/resume, and one-click onboarding links
Reply inbox — every prospect reply streams in live over webhooks and queues for human review
Reply inbox — every prospect reply streams in live over webhooks and queues for human review
Resolution lanes — each company is matched against an 18,000-row federal-filings corpus, or routed to a human review queue
Resolution lanes — each company is matched against an 18,000-row federal-filings corpus, or routed to a human review queue
Fail-closed suppression ledger — 1,899 profiles the system is forbidden to contact, with a reconciler watching for stuck records
Fail-closed suppression ledger — 1,899 profiles the system is forbidden to contact, with a reconciler watching for stuck records
Engine internals — killswitches, per-sender budget counters, intent state distribution, and the last 20 reconciler cycles
Engine internals — killswitches, per-sender budget counters, intent state distribution, and the last 20 reconciler cycles
Six layers deep · scroll

Layer 1 · Intake

Thousands of prospects, found and scored automatically.

The lead engine pulls, enriches, and scores prospects — company data, roles, and contact info assembled before a human ever looks at a name.

Layer 2 · AI

Messages written by AI, in the client's own voice.

Every message is generated per-lead, then locked with snapshot tests so the approved voice can never silently drift. Personal at a scale no human team could match.

Layer 3 · Safety

Hard rules, enforced in code — not by hoping.

Budget gates reserve spend before any message fires. Invariants like never double-messaging the same person are compiled into the system itself.

Layer 4 · Scale

A coordinated fleet of sender accounts.

Daily caps, health checks, and automatic recovery when an account goes down. The machine keeps running even when a piece of it doesn't.

Layer 5 · Real time

Every accept, reply, and open lands the moment it happens.

Production webhook receivers stream live events into the ledger — the system reacts in seconds, not at tomorrow's export.

Layer 6 · Trust

The system audits itself every single day.

Daily crons diff vendor billing against internal ledgers and alarm on any drift. One watchdog caught a silent billing leak within 24 hours of shipping.

Lead engine

Every prospect is resolved against an 18,000-row corpus of real federal filings. Matched companies get reports built from their actual government data; edge cases route to a human review queue instead of guessing.

Personalized reports & AI messaging

The system generates a personalized report for each decision-maker and writes every message per-lead in the client's own voice — locked with snapshot tests so approved copy can never silently drift.

Send-safety governors

Budget gates reserve spend before any message fires. Hard invariants — like never double-messaging the same person — are enforced in code, not by hoping.

Multi-sender fleet

Six coordinated sender accounts with per-sender daily caps, 14-day warmup for new accounts, health checks, and one-click reconnect flows when an account goes down. Follow-ups cancel themselves the moment a prospect replies.

Real-time webhooks

Accepts, replies, and opens stream in live through production webhook endpoints — every event lands in the ledger the moment it happens.

Reconciliation & watchdogs

Daily crons diff vendor billing against internal ledgers and alarm on any drift. One watchdog caught a silent leak within 24 hours of shipping.

The build, step by step

How this one actually got made.

01

Systems audit

Mapped how the client actually acquired customers, where leads leaked, and what 'safe' had to mean for a regulated industry before designing anything.

02

Architecture for safety first

Six layers designed around hard invariants: never message the same person twice, never spend unreserved budget, never let approved copy drift. The rules live in code, not in a checklist.

03

Build under review discipline

100+ pull requests, every one reviewed before merge. Snapshot tests lock the client's approved voice so no deploy can silently change what prospects receive.

04

Hardening

Kill switches on every external vendor. Budget governors that reserve spend before a single API call. Daily reconciliation crons that diff vendor billing against the internal ledger and alarm on drift.

05

Live operations

Real-time webhooks stream accepts, replies, and opens into the ledger. Dashboards show fleet health at a glance. A watchdog caught a silent vendor billing leak within 24 hours of shipping.

under-the-hood · outreach-system

  • Never-double-message invariant enforced at the code level
  • 1,899-row fail-closed suppression ledger — forbidden contacts stay forbidden
  • Follow-up sends cancel themselves the instant a prospect replies
  • Budget reserved before any provider call — spend can't run away
  • Daily vendor-billing vs. internal-ledger reconciliation with alarms
  • Lead Trace joins ~12 database tables into one timeline per prospect
  • PIN-protected run schedule with business-hours clamping
  • AI copy locked by snapshot tests — approved voice can't drift
AI messagingWebhooksCron watchdogsBudget governorsDashboards

Want something like this for your business?

One call. I'll tell you exactly what I'd build — and whether your project is a fit.