CRM AI Sales Automation

The problem

A U.S. logistics company had 50,000+ contacts sitting in Pipedrive with no way to tell which ones had gone cold, or to follow up with them at scale without every email reading like a copy-pasted template. An earlier attempt to solve this with Make.com didn't hold up.

What I built

A full-stack platform that connects to Pipedrive, scans the pipeline for contacts inactive 60–90 days, and uses GPT-4 to draft a personalized follow-up for each one — pulling in that contact's actual conversation history so the email doesn't read like a form letter. Drafts can be reviewed and sent manually, or the system runs on a daily cron and sends on its own once the client trusts it. Every send is logged for a full audit trail, and a system-prompt editor lets the client tune the AI's tone without touching code.

The engineering had to hold up at the client's actual scale: three-layer pagination so no single request times out against 50,000+ contacts, batched OpenAI calls with retry logic that reads the Retry-After header instead of guessing, and a four-layer JSON-parsing fallback so an unpredictable GPT response never crashes the run.

Result

Fully automated daily outreach running without a human in the loop, with a manual review path still available whenever the client wants one. Two rounds of client feedback were incorporated directly into the product — including a full stale-contact detection engine and a brand-voice system-prompt editor that didn't exist in the original scope.