ENRICHMENTSCORINGSIGNAL DETECTION
Account Intelligence Engine
10,925 accounts. Four layers. 1,631 qualified against a forecast of 1,000 — and a validation harness that can't be accused of cherry-picking.
Clay · PredictLeads · Python · Instantly + HeyReach
The Problem
A data consultancy client was launching an enterprise campaign and needed a target book, not a list. The brief said "companies using a specific AI stack," which is a technographic question most vendors answer badly and expensively. I built a four-layer pipeline — firmographic, suppression, technographic, paired contacts — over the whole US universe of companies above $500M, ran a vendor bake-off before spending a credit, and tiered the output into a priority book the copywriters and the campaign could run on. Then I built the harness that checks whether the vendor data is true.
Stack
🧱
Clay
Native company search — 19,994 US companies ≥$500M at 0 credits
🔍
PredictLeads
Technographic layer at 1 credit/company after the bake-off
🐍
Python
Suppression, scoring, tiering, and the md5-ordered validation harness
📤
Instantly + HeyReach
Email and LinkedIn campaigns the book fed
How It Works
The execution path
Execution flow
Firmographic: 19,994 US companies ≥$500M— Clay native search, 0 credits — whole universe, not a sample
Suppression: 4 sources, 413 companies removed before scoring
Vendor bake-off → PredictLeads at 1 credit/company— 10× cheaper than incumbent, richer output
Technographic sweep: 8,827 companies → 1,631 hits (18.5%)— 80.8% coverage of a 10,925-account universe
Tier: Diamond 262 · Platinum 268 · Gold 551 · Silver 167 · Greenfield 117
Paired contacts → copywriter handoff → Instantly + HeyReach
Validation harness: md5(domain)-ordered sample, verdicts sourced per account— 2 confirmed · 2 partial · 1 contrary · 5 uncorroborated of 10
Two decisions carried the project. First, the technographic vendor bake-off: I ruled out two providers on capability and found one at 1 credit per company — 10× cheaper than the incumbent with strictly richer output. The recommendation held in production. Second, the client's written ICP disagreed with their own closed pipeline: the stated 15,000-employee gate would have excluded 8 accounts their team had already flagged green. I documented the disagreement in a comparison table and built to the pipeline, carrying the excluded verticals at neutral weight for review instead of silently dropping them.
Key Design Decisions
🎲
Deterministic Sampling
The validation sample is ordered by md5(domain). It re-draws identically every time, so nobody can call it cherry-picked.
📣
Ship the Contrary Finding
One of ten sampled accounts contradicted the vendor. It went in the report next to the confirmations.
⚖️
Pipeline Beats the PDF
When the written ICP disagreed with closed deals, I built to the deals and documented the disagreement instead of inheriting it.
💸
Bake Off Before You Buy
Whole universe pulled free. Technographic vendor chosen on a head-to-head, then validated in production.
By The Numbers
1,631
Qualified vs. 1,000 forecast