Project Chimera: Autonomous B2B Lead Qualification Engine

Overview

In B2B SaaS, speed to lead is everything. Yet, our marketing team was drowning in unqualified leads from our static contact forms. Sales Development Reps (SDRs) spent 30+ minutes manually researching each prospect on LinkedIn and company websites before deciding if they were worth a call.

I designed and built "Project Chimera," an automated workflow that acts as a virtual SDR. It instantly enriches lead data, uses an LLM to score relevance against our Ideal Customer Profile (ICP), and drafts a personalized email for the human sales team to review.

The Problem

The "Contact Us" form was a black box. We faced three core issues:

• Low Signal-to-Noise: 60% of submissions were unqualified (students, hobbyists, competitors), clogging up the CRM.
• Slow Response Time: It took an average of 4 hours for an SDR to pick up a lead, by which time the prospect had moved on.
• Generic Outreach: Because volume was high, SDRs used generic templates that had low open rates (< 20%).

My Role & Contributions

I wore multiple hats, bridging the gap between UX design and technical implementation:

• Workflow Architecture: Mapped the entire data journey from Typeform submission to HubSpot deal creation.
• Prompt Engineering: Designed and tested the system prompts for the OpenAI GPT-4 API to ensure accurate lead scoring and tone-appropriate email drafting.
• UX Design: Redesigned the Typeform to be more conversational and gather "signal" data without adding too much friction.
• Implementation: Built the automation logic in n8n and set up the database schema in Supabase.

Process

1. Discovery & Definition

I shadowed three SDRs to understand their manual research process. I found they looked for specific keywords on LinkedIn (e.g., "Head of Operations," "Scaling") and revenue data on Crunchbase. We defined these as the "scoring signals" the AI needed to replicate.

2. Designing the Logic

Instead of a visual UI, the "design" here was a logic flow. I treated the LLM as a user, designing prompts that gave it clear context and constraints.

3. Prototyping in n8n

I used n8n for its visual workflow builder. The hardest part was handling edge cases—what if the website URL is broken? What if the LinkedIn profile is private? I added conditional logic to route these "error" cases to a human for manual review.

The Solution

The final product is an invisible but powerful engine running in the background.

The Workflow

• Intake: Prospect fills out a Typeform.
• Enrichment: n8n grabs the domain and pings Clearbit/Apollo for company size, funding, and tech stack.
• AI Analysis: GPT-4 reads the prospect's input and the enriched data. It assigns a "Fit Score" (1-10) and writes a 1-paragraph summary of why they are a good fit.
• Action:
  • If Score > 7: Create Deal in HubSpot, draft personalized email, Slack alert to Senior AE.
  • If Score < 7: Add to "Nurture" sequence in HubSpot, send polite automated rejection/resource email.

Impact

TEST