I built a full B2B lead gen pipeline in n8n — one Telegram message triggers everything. Here's the full breakdown

Apr 11 • General

Hey everyone,

I just finished a production-ready AI lead generation workflow in n8n and wanted to share the full architecture with the community. Happy to answer questions and hear how you'd improve it.

🔧 What it does (end-to-end):

You send a search query to a Telegram bot (e.g. "IT consulting firms Amsterdam")
Google Maps Places API returns up to 10 businesses
Leads are split, deduplicated across runs (persistent memory), and paginated
Prepare Keys node enriches each item with company_name, website, phone, address and generates a company_key for merging
Two parallel Serper searches run simultaneously:
Two sequential Merge nodes combine everything by company_key
Edit Fields normalizes all data into a clean unified object
Research Agent (GPT-4o-mini) generates 3 automation opportunities + personalized icebreaker per company
JSON Cleaner (Code node) parses the AI response and merges it with contact data
Contact Validator scores email/LinkedIn quality (0–100)
Lead Scoring calculates total score: corporate email (+40), LinkedIn (+25), website (+15), phone (+10), icebreaker quality (+10)
IF Score gates the outreach — only leads ≥ 55 proceed
Google Sheets Append logs everything to CRM
Email Generator (GPT-4o-mini) writes a personalized cold email
Gmail creates a draft → waits 3 days → follow-up draft created → Telegram report sent

🐛 Interesting bugs I hit and fixed:

$json context loss after Serper nodes — fixed by using $('Prepare Keys').item.json.company_key explicitly
websiteUri returning undefined for businesses without a website — fixed with || $json.nationalPhoneNumber || 'unknown' fallback in the key
JSON Cleaner failing with raw.slice is not a function — GPT-4o-mini returns a parsed object, not a string, when json_object mode is enabled
Merge nodes not joining — caused by null company_key on one input branch
Remove Duplicates discarding all items — cleared deduplication history after test runs
💡 Key design decisions:

Used company_key = displayName.text + '_' + (websiteUri || phone || 'unknown') as a composite dedup key
All Serper queries include city/country extracted from formattedAddress for geo-accuracy
JSON Cleaner checks typeof content before parsing — handles both string and pre-parsed object responses from OpenAI node
Error handler workflow connected globally — sends SOS to Telegram on any node failure

🔗 Full source code + README on GitHub: https://github.com/Evgeniy1970-ai/AI-Lead-Gen-Production-Full-Workflow_production

Would love to hear your thoughts — especially on the scoring logic and whether you'd handle the Merge differently. Drop your questions below 👇

#n8n #automation #leadgeneration #AI #GPT4 #workflow #nocode #lowcode

1 comment