Alien Campaign Management: Orchestrating Multi-Channel Outbound at 1,000+ Touches Per Day

Running a 1,000+ touch per day outbound operation as a solo operator without a campaign management layer means spending 4 to 6 hours daily on sequencing, reply triage, variant selection, and channel coordination. The design goal was not automation for its own sake. It was automation of the rote and human judgment on the consequential.

Campaign management for a custom development firm has a structural difference from campaign management for SaaS lead generation. The average deal size of $15,000 to $25,000 meant that each meeting was worth fighting for individually. A mass outreach tool that treated prospects as fungible was the wrong model.

The multi-channel requirement added complexity. Cold email was the primary channel. LinkedIn DM was the secondary channel for prospects who did not reply to email. Twitter DM was the tertiary channel for founders visible on that platform. Reddit and Discord opportunity monitoring ran in parallel as inbound-to-outbound triggers.

The orchestration layer used alien/cadence (Rust), which consumed lead state from the PostgreSQL database and emitted scheduled send jobs to channel-specific Redis queues. Cadence logic: 4-touch email sequence (day 1, 3, 7, 14) plus 3-touch LinkedIn DM sequence triggered after email touch 3 with no reply, plus 2-touch retarget post sequence for prospects who opened email but did not reply.

Each channel had independent rate limits enforced at the Redis queue consumer level. Email: 30 to 50 sends per inbox per day, 90 inboxes, total 2,700 to 3,600 sends per day. LinkedIn DM: 30 total per day across all active LinkedIn accounts. Twitter DM: 20 total per day per Twitter account.

Reply classification used Claude API (claude-sonnet-4-6). Every inbound reply was classified into: positive, neutral, objection, bounce, or stop. Classification accuracy was validated at 94% against a manually labeled 200-reply test set. Classification cost approximately $0.001 per reply at Haiku pricing. At 72 replies per day, classification cost under $2.60 per month.

Campaign state in PostgreSQL extended the lead state machine with a campaign_enrollments table linking leads to campaigns with per-touch status tracking. The join between leads and campaign_enrollments made it possible to answer operational questions like how many leads are on touch 3 of the November agency campaign with no reply to touch 2 in a single SQL query. Campaign pausing was atomic: an UPDATE on campaign_enrollments setting status=paused for all future touches.

The alien/ab-tester maintained a variants table in ClickHouse. Auto-promotion logic ran every 6 hours. Criteria: a challenger variant with at least 50 sends and a reply rate exceeding the current promoted variant by more than 2 percentage points at 95% confidence. Chi-squared test implemented in Rust using the statrs crate. The operator could reject the promotion within 2 hours by clicking a link, otherwise it was applied automatically.

The LinkedIn dm-engine simulated human typing cadence with 100 to 300ms delay between characters to avoid LinkedIn's automated bot detection. LinkedIn session cookies stored encrypted in PostgreSQL and rotated on 401 responses.

Campaign management system designed for 1,000+ touches per day across 4 channels. Reply classification accuracy: 94% against 200-reply labeled test set. A/B variant auto-promotion: 6-hour cadence, 95% confidence threshold. Break-up email reply rate: highest in sequence, consistent with practitioner benchmarks. Operator daily review time target: 20 minutes versus 4 to 6 hours without the system.