Data Scientist - PNC Bank, Cybersecurity Organization - Columbus, OH
(2021-08)
Fraud detection team that catches internal-actor fraud across PNC's banking products. Work splits between production rule-based alerts and ML triage models. Stakeholders: cybersecurity, compliance, operations.
- Built a Bagging PU (Positive-Unlabeled) classifier to triage alerts where confirmed fraud is a small fraction of total alert volume for Same Day Withdrawal fraud model.
- Lifted F1 from 2 percent to 8 percent at 80 percent recall. Analysts now triage off model output rather than raw rule output, which has cut the time spent on false alerts.
- Consolidated six fraud rules (Same Day Withdrawals, Teller Reversals, Self Dealing, DDA Non-Employee to Employee, SVG Non-Employee to Employee, Teller Override) into one unified training set spanning 2021 to 2024 to prevent overfit.
- Engineered transaction-pair behavioral features: how often a specific customer and teller interact, same-branch visit patterns, employee tenure flags to cut false positives.
- Used frequency encoding for high-cardinality fields (Bank No, Branch No) since one-hot would have blown up the feature space.
- Applied time-based split for validation: trained on 2021 to 2024, validated on the next 7 months, tested on the 7 months after to avoid information leakage.
- Wrote the production pipeline for data ingestion, feature engineering, and alert scoring.
- Built Private Bank Address Change rule combining Jaro-Winkler and Levenshtein fuzzy matching with geospatial proximity scoring against 1,000-plus PNC branch locations.
- Built Self-Dealing Fee Refund rule catching employees processing fee refunds on their own accounts.
- Built Employee Phone Number Linking rule firing when an employee links the same phone number to multiple customer accounts.
- Built DDA Non-Employee to Employee rule (SparkSQL) flagging transfers moving from a regular customer checking account into an employee-owned account.
- Built EDGE Executive Snooping rule (SparkSQL) detecting employees pulling up executive accounts without a documented business reason.
- Built Demo Check Abuse rule (SparkSQL) triggering when an account receives more than five demo checks inside a 90-day rolling window.
- Built Quick Balance Anomaly rule catching employees skipping a full cash drawer count in favor of posting a lump-sum quick balance.
- Built Instant Card Order self-issuance rule flagging cases where the issuing employee is also listed as Cardholder Name or Authorized Signer.
- Built Tableau dashboards and emulation reports for production fraud rule monitoring.
- Set up the framework for recurring monthly fraud reporting.
- Pushed updates to demand-deposit account fraud rules into production along with matching reports.
- Designed and shipped the phone-number-change alert system from scratch including requirements, build, testing, and deployment.
- Rewrote fragile legacy query logic against Criss Index databases and Hadoop extract tables for self-dealing fee-refund investigations.
- Fixed broken production code on the debit card disputes pipeline after business-rule drift and reworked the logic.
- Served as stakeholder lead across cybersecurity, compliance, and operations with regular updates and on-time delivery.
Career Break
(2015-10 - 2021-08)
Career break to care for my first child during a period when my husband's role required frequent travel.
Data Analyst - BNP Paribas Bank
(2012-06 - 2015-10)
Data analysis and validation across the lending portfolio: Consumer Credit Card, Home Lending Servicing, Auto, and Personal Loan Originations. Worked closely with developers, BAs, and PMs.
- Led the Home Lending lead-creation project, which fed into the customer acquisition pipeline by translating business requirements into data analysis across multiple lending systems.
- Ran data quality checks and validation, catching consistency issues before they hit production reports.
- Validated mappings on imports and exports.
- Analyzed JSON and XML payloads and ran API analysis on REST/HTTP and SOAP to trace data flows.
- Used POSTMAN heavily for API validation.
- Worked with integration partners on process improvements and reliability fixes.
