Overview
Top 3 Finalist in the Google x UC Berkeley Analytics Hackathon. Diagnosed a 3x customer churn spike for a SaaS company by analyzing messy, real-world datasets and delivered an executive-ready recommendation projected to retain $23M in annual contract value.
Problem Statement
BizGrow, a SaaS inventory management platform, saw churn spike from 4% to 12% in one quarter. Leadership was divided:
- Sales blamed product performance
- Product blamed low-quality customer acquisition
- Support felt overwhelmed by tickets
The COO needed one actionable recommendation with limited resources—not a dashboard or 50 charts.
Solution
Reframing the Problem
Defined churn by revenue impact rather than customer count. Enterprise contracts were 46x more valuable than SMBs, so logo-based churn metrics obscured the real risk.
Root Cause Analysis
Analyzed three datasets with real-world flaws (missing values, corrupted EU server logs, unstructured text):
- CRM Data – Contract dates, company size, industry (~30% missing)
- Usage Logs – Daily login activity (EU September data corrupted)
- Support Tickets – Unstructured customer complaints
Key Findings
Customer Success > Technical Failures
- 58% of churned customer tickets were customer success issues (onboarding friction, sales expectation gaps)—not product bugs
- Customers with only CS-related tickets churned at 21.5% vs 14.9% for technical-only tickets
Small Companies Drive Volume, Not Value
- Small companies (1-10 employees) drove 75% of churn but represented only 34% of customers
- Revenue impact was concentrated in enterprise accounts
Usage Drops Signal Frustration
- Session time dropped significantly on ticket-filing days (p=0.017)
- Validated that complaints reflected real user friction, not noise
Recommendation
The fastest lever: close the expectation gap between what sales promised and what customers experienced.
- Retrain sales on accurate product positioning
- Implement structured onboarding check-ins for first 60 days
- Proactively monitor low-usage customers as churn indicators
Projected Impact: Retain ~$23M ACV and reduce daily support load by 193 hours.
Technical Stack
- Analysis: Python, pandas, scipy
- Methods: Statistical significance testing, ticket text classification, cross-dataset validation
- Deliverable: Executive slide deck with data-driven narrative
Impact
Demonstrated ability to navigate ambiguous business problems, handle messy data, and translate analysis into actionable recommendations under time pressure—core skills for data science in business contexts.