Outbound Lead Generation & Personalization Engine

Outbound has a scale-versus-relevance problem. Generic mass sends get ignored, and genuine one-to-one personalisation does not scale manually. Researching each company, finding the right person to contact, writing something specific enough to earn a reply, tailoring the collateral you attach, and sending it at a sensible time is hours of work per prospect. Do it properly and you can't do volume. Do it at volume and it stops being worth reading.

I treated this as a product problem rather than a volume problem. The insight was that quality per send matters more than send count, so the automation should absorb the mechanical work - discovery, contact resolution, scheduling, tracking - while personalisation stays genuinely specific to each prospect. That meant putting a quality gate in front of the send rather than trusting raw model output. I spec'd the full pipeline before building, cut features that added complexity without improving reply quality, and iterated on what actually moved replies. Building it against my own outreach first meant I felt every failure mode directly, which is why the architecture held up when I refactored it for a client.

An eight-stage agentic outbound pipeline: Playwright and Apify handle multi-source prospect discovery, Hunter.io resolves contacts with confidence scoring, OpenAI generates per-company messaging with playbook enforcement, A/B template testing, and a humanizer plus LLM quality gate, a LaTeX compiler tailors per-prospect collateral, the Gmail API manages OOO detection and scheduled sends with a separate executive track, and Google Sheets carries conversion tracking and analytics. A Claude Code skill orchestrates the full daily run. The same architecture is now being refactored into an AI SDR agent for Sharpenn Technologies, a B2B process-equipment manufacturer, to open their first outbound channel.

A fully automated daily pipeline running in production. Every send is personalised per company, reviewed by an AI quality gate before dispatch, and timed and tracked end to end. The stronger signal is portability: the engine was built for one use case and then refactored into a governed B2B AI SDR for a manufacturing client that had run entirely on referrals, which is the real test of whether the abstractions were right.

Playwright and Apify for multi-source prospect discovery, Hunter.io for contact resolution with confidence scoring, OpenAI for message generation with template A/B testing and an LLM quality gate, LaTeX with tectonic for per-prospect collateral compilation, Gmail API for draft creation and scheduled sends, and Google Sheets as the tracking and analytics layer. A Claude Code skill orchestrates the daily pipeline. OOO detection parses email headers and defers sends to the contact's return date. The client deployment adds a governance layer on top: confidence tags on material facts, claims restricted to a locked credentials file, and a two-stage review before anything leaves the system.