I built an automation pipeline that discovers buying signals across the web, qualifies them against an ICP using LLM scoring, enriches the prospects, and pushes them into Amplemarket's Duo via their Custom Signal API. Duo then auto-generates personalized multichannel sequences.
https://artemisgtm.ai/blog/amplemarket-duo-signal-playbook
Here's the technical breakdown.
The Architecture
Signal Discovery (Exa MCP)
|
v
ICP Qualification (LLM Scoring)
|
v
Enrichment (Amplemarket MCP)
|
v
Webhook Push (Custom Signal API)
|
v
Duo Sequence Generation (Automatic)
Phase 1: Signal Discovery
I use Exa for semantic search. Unlike traditional keyword search, Exa understands intent. Querying "founders who discussed lead response challenges on podcasts in the last 30 days" actually returns relevant results.
Signal types: podcast appearances, conference talks, funding announcements, hiring patterns, competitor reviews, regulatory changes. Each represents a verifiable moment where someone in your target market demonstrated a specific pain point.
Phase 2: LLM-Based Qualification
Each signal gets scored against ICP criteria. I pass the signal context plus ICP definition to Claude and ask for a structured score across four dimensions: company fit, role fit, pain alignment, and timing urgency.
Anything above 70% aggregate confidence triggers the next phase.
Phase 3: Enrichment via Amplemarket API
Qualified prospects get created in Amplemarket via their lead management API. Enrichment fills in contact details, company data, tech stack info. Typical enrichment latency is 30-60 seconds per batch.
Phase 4: Custom Signal Webhook
This is the interesting part. Amplemarket's Custom Signal API (POST /custom_signals/{token}/entries) accepts structured data about a prospect plus context, and routes it to their Duo AI copilot.
Key insight: create separate Custom Signals for each signal type. A "Podcast Pain Signal" webhook has different messaging instructions than a "Funding Round Signal" webhook. This means Duo adjusts its tone and angle automatically based on how the prospect was discovered.
Payload structure:
json{
"first_name": "Sarah",
"last_name": "Chen",
"company_name": "Acme Corp",
"email": "sarah@acme.com",
"linkedin_url": "linkedin.com/in/sarahchen",
"extra_context": "Appeared on SaaS Scaling podcast March 2026. Discussed challenge of converting website visitors to pipeline. Company recently expanded SDR team from 2 to 5."
}
The extra_context field carries the signal intelligence. Duo reads it and generates sequences that reference specific details from the prospect's public activity.
Phase 5: Human Review
Duo generates the sequences. A human reviews before sending. I don't recommend going fully autonomous on outreach copy. But the difference between reviewing context-rich AI drafts versus writing from scratch is 20 minutes versus 4 hours per batch.
Implementation Options
This can run as a Claude skill (what I built first), an n8n workflow (for scheduled execution), or a Python script. The architecture is tool-agnostic.
The important decisions are: what signals to scan for, how strict to make your ICP qualification, and how to write the Duo messaging instructions for each signal type.
Results
Podcast signals convert at 2-3x any other signal type. Funding round signals need careful framing (skip "congrats on the raise"). Hiring signals are consistently underrated.
Full non-technical breakdown: https://artemisgtm.ai/blog/amplemarket-duo-signal-playbook
Top comments (0)