ProfessionalRole
500+Connections
Substack
--Articles
--Subscribers
About
Product leader who turns messy document workflows into one-click automations. Power of 3.
Proof of Work
2 plays
AI Document Extraction Pipeline for Tax Forms
Built AI doc extraction that cut manual corrections by 72%
Before71% OCR accuracy, 3.2M manual corrections/season
After94.3% accuracy, 72% fewer corrections, $1.9M saved
ProductSystem
Situation
TurboTax users uploaded 12M+ W-2s and 1099s annually, but OCR accuracy was 71% causing 3.2M manual corrections per season. Each correction cost $0.85 in support overhead and added 4 minutes to filing time.
Action
Designed a multi-model extraction pipeline using Claude for field validation and GPT-4V for image preprocessing. Built a confidence-scoring system that auto-routes low-confidence extractions to human review. Ran A/B test across 500K users over 8 weeks.
Result
Extraction accuracy improved from 71% to 94.3%. Manual corrections dropped by 72%, saving $1.9M/season in support costs. Average filing time decreased by 6 minutes. Feature became default for all TurboTax tiers.
312 copies · 67 forks
View ArtifactClaudeFigmaSQLAmplitude
Self-Service Document Upload Redesign
ProductWorkflow
Situation
Document upload flow had a 34% drop-off rate at the review step. Users were confused by the multi-step verification process.
Action
Redesigned the upload flow with inline previews, progressive disclosure, and real-time validation. Conducted 20 user interviews and 3 rounds of prototype testing.
Result
Drop-off rate decreased to 12%. NPS for the upload experience went from 23 to 61.
0 copies · 0 forks
View ArtifactFigmaAmplitudeUserTesting