AI B2B Lead Engine — LangGraph Multi-Agent Sales System
A 6-agent hub-and-spoke StateGraph that qualifies, scores, and works B2B leads across LinkedIn, email, and voice
A 6-agent hub-and-spoke system built on a LangGraph StateGraph that runs B2B sales end-to-end: it qualifies and scores leads with a MEDDIC model (0-60), drafts outreach grounded in 7 embedded sales frameworks, and executes across three channels — LinkedIn, email, and AI voice calls — while a real-time Streamlit dashboard keeps a human in the loop for review. Shipped with 124 tests and typed throughout with Pydantic v2.
The pain this had to solve
B2B sales teams lose most of their time before a single real conversation happens: researching accounts, qualifying whether a lead is worth pursuing, deciding which channel to use, and writing outreach that does not read like a template. Done by hand, this work is slow, inconsistent, and impossible to audit when a deal stalls.
The goal was a system that runs the qualification-to-outreach loop end-to-end with a defensible scoring method, draws on proven sales frameworks instead of generic copy, reaches prospects on whatever channel fits, and still lets a human approve or override before anything goes out.
What I built — the architecture
A LangGraph StateGraph coordinates six specialized agents through a central hub with shared state, so research, scoring, and outreach all read and write the same lead record instead of drifting out of sync.
Each lead is scored 0-60 on a MEDDIC model (Metrics, Economic buyer, Decision criteria, Decision process, Identify pain, Champion), making "is this worth pursuing?" an explicit, auditable number rather than a gut call.
Seven sales frameworks are embedded in the agents so outreach is grounded in real selling methodology — discovery, value framing, objection handling — not generic AI-written filler.
The system reaches prospects across three channels — LinkedIn, email, and AI voice calls via Twilio and Retell AI — choosing the channel that fits the lead and the stage.
A real-time Streamlit dashboard surfaces every lead, score, and drafted message for human review and override before outreach is sent.
The whole pipeline is typed with Pydantic v2 and backed by 124 tests, so state transitions and agent outputs stay validated and the system is safe to extend.
What it delivered
The system replaced hours of manual research and qualification with a defensible, auditable pipeline: every lead gets a MEDDIC score (0-60), outreach is grounded in seven sales frameworks across three channels, and a human approves before anything sends. Shipped with 124 tests, it is safe to extend as the sales motion evolves.
Want results like this for your team?
Tell me what you want to automate. On a free 30-minute call I’ll tell you straight whether it’s worth building, roughly what it costs, and how I’d approach it — no pitch, no obligation.
Book my free 30-min AI scoping call →