Clay + Claude scoring pipeline

Problem

Raw lead lists pulled from scrapers, public datasets, and broker sites contained thousands of records with wildly varying quality. Manually qualifying them against acquisition criteria — industry fit, structural quality, buy-box alignment — wasn’t tractable at the volume we needed to operate. Off-the-shelf scoring (“good lead / bad lead”) breaks the moment you’re targeting more than one thesis at once.

What I built

A data enrichment and lead scoring system in Clay that aggregates raw leads from multiple sources, enriches them with firmographic and web-scraped data, and runs them through an LLM-based scoring framework. The 1–6 scoring scale evaluates industry fit, structural quality, and overall attractiveness using AI prompts grounded in website-derived inputs. Iterated on scoring logic to balance strict industry alignment against broader structural fit, so the pipeline surfaces both bullseye targets and high-quality adjacencies.

1. Data aggregation (Clay) — website flatten, NAICS, reviews, firmographics
2. Industry-specific evaluation (Claude) — thesis-comparison prompt, on-site evidence only
3. Structured scoring output — 1–6 scale with banded meanings
4. Override logic — caps for edge cases, softer penalties on missing data

Architecture principle

Separation of concerns. Clay handles data + orchestration. Claude handles reasoning + judgment. Neither tool is stretched to do the other’s job, which is what makes this cheap to operate and easy to change.

The scoring band

• 1–2 — not a fit — hard pass
• 3–4 — edge case — manual review before send
• 5–6 — strong / near-ideal — send with primary campaign

Outcomes

time to live

First working version within a week

filter rate

Raw lists → high-confidence subset before any send

cost

Operational cost in the low hundreds per month

coverage

Underpins both outbound programs — same engine, different rubrics

iteration

Continuously modified as buy-box criteria shift