Data platforms and analytics#
The data foundations that analytics and models depend on, and the models that pay for them: pipelines, attribution, document NLP and prediction APIs for retail, fintech, education and pharma. Six projects, 2019 to 2025.
Data platform engineering for a UK retailer#
Problem · A UK retailer's data platform runs raw, conformed and consumer data products on a shared pipeline framework. The products, from point-of-sale to orders and payments, had to stay working while the framework moved through several major versions.
What I did · Upgraded the data products across framework versions, fixed stale change streams and typing in the raw layer and duplicate rows in consumer fact tables, added a new raw data product for payments, and wrote the developer and release-process documentation.
Result · Products upgraded and documented, and a new payments source live on the platform.
Stack · dbt, Airflow, Snowflake, Terraform, GitLab CI, GCP.
BI and NLP for tax filings#
Problem · Tax refunds go unclaimed because the relevant details are buried in millions of heterogeneous documents from the tax authorities.
What I did · Parsed assessment notices into the data warehouse, classified raw assessments with NLP, added PII detection, built an OCR-quality monitoring pipeline, and benchmarked a new document-OCR vendor against the old one offline.
Result · Faster analysis of tax documents and better identification of overpaid income taxes, such as missed childcare deductions. The OCR benchmark showed document accuracy up by half.
Stack · Airflow, spaCy, Snowflake, Python.
Marketing data foundation for a medical education platform#
Problem · Marketing and sales data sat scattered across the warehouse. The company needed one governed foundation for sessions, traffic categorization and attribution of sign-ups, purchases and licenses, and documentation of what already existed.
What I did · Built the marketing data foundation as a Dataform model on BigQuery, covering attributions, journeys, users, events, costs and monitoring, and produced lineage analyses of the existing view stack across billing, CRM and sales.
Result · A documented, governed foundation for marketing analytics.
Stack · Dataform, BigQuery, Airflow, Python.
Marketing data team
"Thanks for bringing some order into chaos in the marketing data world."
Multi-touch marketing attribution modeling#
Problem · Marketing needed customer-acquisition cost by channel, campaign and country, which means attributing conversions across web visits, influencers, affiliates, search ads, newsletters and video, and tracking it daily against a lifetime-value target.
What I did · Built an attribution MVP with first-click, last-click and extended models, journeys and channel co-occurrence, then a production PySpark pipeline on AWS EMR: visits landed in S3 as partitioned Parquet, attribution extracted with a three-month conversion window into the PostgreSQL warehouse, with config-driven cleansing rules and visit-user matching with data provenance. Answered the marketing team's questions in a series of analysis notebooks and fed a propensity model and incrementality tests.
Result · Daily attribution over millions of visits and hundreds of thousands of conversions, in production.
Stack · Piwik/Matomo data model, PySpark, AWS EMR, S3, Parquet, MySQL, PostgreSQL, SQLite.
Senior Performance Marketing Manager
"Thanks, Michele. What are we going to do when you finish?"
Discovering and mitigating inaccuracies in tax filings#
Problem · A tax-filing app estimates refunds from questionnaire answers, but the tax office's assessment often differs. That gap hurts trust and creates risk for a product that pays refunds out early.
What I did · Helped define the business problem, acquired and integrated the training data from hundreds of thousands of filings, and built regression, binary, multi-class and quantile models. Validated them with domain experts, found error-correcting signals in the tax engine's intermediate calculations with SHAP analysis, and implemented the prediction API in production, wired into the tax-operations tool so manual review goes to the riskiest cases.
Result · Refund-estimate error down by roughly a fifth, with calibrated quantile predictions for the payout decision.
Stack · Google AI Platform, App Engine, Kubernetes, Flask, Snowflake, JupyterLab, scikit-learn, XGBoost.
Head of Data
"Would you like to join us full-time as Lead Data Scientist?"
Visual analysis of pharma business processes#
Problem · Pharmaceutical processes around confidential information, product development and trials are complex and interconnected. Their recorded logs can reveal where automation and optimization pay off, if people can see them.
What I did · Developed an application component that visualizes business processes and highlights areas of interest, integrated into a business-intelligence dashboard.
Result · In use as part of the dashboard.
Stack · Node.js, Vue, TypeScript, yFiles.
Working on something similar?
The quickest way to find out whether I can help is a short call. Prefer email? Write to [email protected].