Experience
My work combines modeling, software engineering, product intuition, and scientific rigor. I have built AI systems for messy real-world data, launched production fraud models, improved identity verification accuracy, investigated unfamiliar datasets, traced data anomalies, and created infrastructure that makes data science teams faster and more reliable.
Case Study
Abodemine clients upload loan tapes in inconsistent spreadsheet formats when evaluating groups of mortgages to buy or sell. I architected an LLM-driven column matching system that maps ambiguous client workbook fields to a canonical property and loan dataset of more than 1,000 attributes.
Problem
Client spreadsheets often use different field names, workbook layouts, enum values, and data conventions for the same underlying loan and property concepts.
Approach
I built the prompting framework, deterministic rule engine, data normalization, enum mapping, and multi-stage validation flow using Python, LLM APIs, Anthropic, and Gemini.
Reliability
I built the evaluation data pipeline and a human-in-the-loop feedback system that captures client corrections so matching quality can improve as the product scales.
Related ML Systems
I also designed a scalable entity resolution system for real estate transaction records and developed model monitoring infrastructure for reliability, drift detection, and maintainability.
Case Study
At Prove, I worked across ML product impact, identity and fraud risk, and production data science infrastructure. I grew from Data Analyst to Senior Data Scientist while leading initiatives that improved core verification products and created new model deployment patterns.
Risk Score
I built and championed Risk Score, a product that collected onboarding signals into a trained XGBoost model. It caught 98% of fraud at a 3% friction rate, reducing friction by more than 80% versus previous solutions.
Executive Buy-In
Risk Score was a self-directed project that I pitched up the management chain to the CTO, who greenlit it and two additional ML products after hearing the proposal.
TrustScore
I architected and deployed the International TrustScore ML model from concept to production, establishing a model microservice framework that accelerated deployment cycles by 200%.
Assurance Level
I overhauled the legacy Assurance Level verification system with ML enhancements that increased AL=3 accuracy to 99%, directly strengthening Verified Users.
Data Investigation
I was often the go-to person for new client or vendor datasets: probing data structure, identifying usable signal, clarifying product possibilities and limits, and tracking down the source of anomalies when the data behaved unexpectedly.
Engineering Leverage
I developed the team's core Python repository with pytest and GitLab CI/CD, reducing scripting overhead by 70%, and collaborated on LazerDome, a centralized AWS environment for secure access and improved collaboration.