Cindy – Python, Pandas, Data Science

It's a rebuild of a degraded scoring model for a major social platform's ad inventory, which had collapsed into producing near-identical predictions across all input combinations.

• Reframed the problem from classification to regression for meaningful score differentiation across inputs.
• Designed split-first architecture to prevent training/inference data leakage.
• Encoded domain constraints into preprocessing, blocking structurally impossible input combinations.
• Validated non-linear feature relationships with mutual information, justifying tree-based models.
• Delivered first differentiated scores across device/format combinations in 4 years.

It's a rebuild of a manual, analyst-dependent scoring process into a fully automated event-driven ML pipeline on AWS.

• Designed S3 → Lambda → Airflow → Docker architecture eliminating all manual steps.
• Built a multi-response ensemble scoring with row-level keys for safe prediction join-back.
• Implemented global score distribution tracker with automated drift detection alerting.
• Created a graceful market size fallback chain across three data sources for global coverage.
• Enforced destination schema validation pre-write, replacing silent failures with clear errors.
• Reduced turnaround time from 1.5 weeks to under 2 minutes, enabling 5 repeat global clients.

It’s a production ML issue I handled: one is designing an interpolation approach for a new, unseen product variant without retraining, and the other is ongoing model governance and drift monitoring tied to data sources.

• Traced data creep to single client rollout and analyzed score distributions across variants.
• Verified new variant blend formula maintained monotonicity between existing bounds.
• Corroborated internal distributions against third-party benchmarks for relative impact.
• Rejected unsupervised clustering due to insufficient data and traceability needs.
• Secured senior data scientist, customer team, and VP alignment before shipping.

It's a reusable statistical testing framework for measuring outcome lift across binary and ordinal survey responses in a controlled exposure study comparing two treatment conditions among thousands of respondents.

• Implemented Chi-square, Fisher's Exact, and Z-tests for binary outcomes, validating convergence across methods.
• Added Mann-Whitney U for ordinal Likert-scale data, selecting appropriate non-parametric tests.
• Built null imputation sensitivity analysis as a conservative lower bound for robust significance.
• Documented test selection rationale for non-technical stakeholders.
• Designed a reusable statistical framework for future studies across response columns and grouping variables.

It's a research project at a major university combining NLP-based signal extraction from SEC 8-K filings with event study methodology to identify language patterns associated with accounting fraud disclosures.

• Built a keyword extraction pipeline on SEC 8-K filings to detect fraud signals like blame-shifting and regulatory mentions.
• Combined SEC enforcement actions and restatements as additional fraud indicator categories.
• Applied event study methodology with CRSP-Compustat returns to measure abnormal returns around flagged disclosures.
• Identified market reaction patterns consistent with material adverse disclosure events.
• Sustained year-long independent research from data acquisition through analysis and documentation.