Self-Running B2B Outreach Engine
Designed & architected. Build in progress.
A specialist engineering consultancy needed a steady stream of qualified sales calls, but the principal could spare only a few hours a week and had been burned by tools that either did nothing or risked spamming a reputation-sensitive professional network. They needed outreach that ran itself, without ever putting the firm’s name at risk.
We designed an end-to-end outbound engine orchestrated in n8n with a PostgreSQL state machine as the single source of truth. AI is used with discipline: a tiered Claude setup (Haiku for high-volume reply classification, Sonnet for copy, Opus to anchor brand voice) classifies and drafts, but every send decision is deterministic code gated by the database. No LLM ever sits in the irreversible send path. Three layers of idempotency make a double-send structurally impossible, and a CASL/PIPEDA compliance spine enforces consent and propagates unsubscribes within minutes.
The system was designed end-to-end across eight subsystems (enrichment, consent, sequencing, dispatch, reply routing, booking, control, and audit), with a fully specified data model and operational runbooks. Its defining property is safety by construction: AI accelerates the work while deterministic guarantees protect the client’s reputation. Engagement in progress; outcomes will be published as the system goes live.