Agentic Sales Infrastructure
Autonomous lead discovery, qualification, personalization, and outreach for founders and operators. Built because the alternative is 6 hours a day in spreadsheets.
The Problem
Most early-stage teams have no dedicated sales function. Founders do it manually. The process is: find leads, qualify them, write personalized messages, follow up. Each of those steps is 30 minutes per prospect and entirely automatable.
The goal was not to build a spam machine. It was to eliminate the mechanical parts so that the human judgment goes where it matters: closing and qualifying inbound.
The Approach
Four agents in a pipeline: discovery, qualification, composition, and dispatch. Each is isolated, testable, and replaceable. The pipeline can be triggered manually or on a schedule.
Qualification is the hardest part. A rule-based system produces too many false positives. An LLM with a structured schema and a few-shot prompt does noticeably better, and the misses are cheap to catch with a quick manual review before anything sends.
Architecture
LangChain for orchestration. Bedrock Claude for qualification and composition. Each stage writes structured output to a queue. The dispatch agent reads from the queue and sends only when a human has reviewed (for first-run campaigns).
Status
Active build. Running internally. Evaluating for deployment to clients as a managed service.