0→1 Launch · Data Platform · Visa
35% of transactions were
invisible to ML models
I found a critical data gap — declined authorizations covering 35% of all transaction volume were invisible to downstream ML. Then I built the product that closed it.
The Data Gap — Before & After
35%
Declined authorizations — invisible to ML. No alerts. No tickets. No one was looking for what wasn't there.
→
94%
Full signal coverage. Declined auths now a first-class payment signal feeding all downstream systems.
What I
Built
🗺️
Message Flow Mapping
end-to-end txn lifecycle
📐
Data Model Design
field-level mappings
🔧
Ingestion Pipeline
dedicated for declined auth
Feeds
Into
🤖
ML Models
fraud detection
📊
Analytics
risk dashboards
🔍
Agent Tool-Use
RAG retrieval source
signal → ML models underperforming on specific merchant categories
clue → Data science flagging incomplete training sets
action → Pulled raw transaction message flows end-to-end
found → Declined authorizations (35% of volume) never reach ML pipelines
why → Historically treated as terminal events — 0 alerts, 0 tickets, 0 owners
impact → To the models, 35% of transactions didn't exist.
Discovery Decision
Find what's NOT there
Nobody filed a ticket. I found the gap by reading raw data flows and noticing the absence.
Business Case Decision
Translate tech → executive language
"Signal coverage" didn't move the CRO. "Revenue lost to false declines" did.
Architecture Decision
New pipeline, not a patch
Rejected quick fixes that would create tech debt. Built a first-class global signal.
Scale Decision
30-person team across 6 regions
Engineering, data science, risk, legal, compliance. Briefed the CRO directly.
✓ Chose
New data model + dedicated ingestion pipeline
First-class global payment signal — scales across all regions and clients
Scales globally ✓
✗ Rejected
Patch existing pipeline for partial declined data
Fast but creates tech debt and inconsistent field coverage
Creates debt ✗
✗ Rejected
Bilateral data sharing with individual issuers
Wouldn't scale and creates governance complexity
Doesn't scale ✗
Rollout — From Pilot to Global
Phase 1 — Pilot
3
enterprise clients
validated data model
→
Phase 2 — Scale
30+
clients onboarded
integration guides + FAQs
→
Phase 3 — Global
92%
adoption in 6 months
6 regions live
The most impactful platform improvements are never on the roadmap. Nobody filed a ticket saying "we're missing 35% of our data." The skill is finding what your platform is invisibly ignoring. And infrastructure investment requires a business case in the language of the people approving it. Translate the technical problem into the business cost. Every time.