Computer vision, fraud detection, predictive modeling — built on problems I know from the inside.
A computer vision pipeline that analyzes photographs of structures — identifying roofing materials, siding, slope geometry, ancillary components, and fencing — then generates construction estimates and replacement cost valuations automatically.
Built originally in 2020–2021 using TensorFlow's Object Detection API, the system included custom-trained models for multiple construction trades, integrated labor cost data by ZIP code, and surfaced estimates through a REST API. The current version uses YOLO (Ultralytics) for detection across roofing materials, slope classification, siding type, and fencing. No comparable product existed when this was built — and years later, the gap in the market remains.
The system received a significant acquisition offer, which I declined. The domain knowledge embedded in the feature engineering — which materials to detect, which combinations drive cost, how slope affects labor — is what separates it from any generic computer vision demo.
A Streamlit application that predicts appraisal award amounts and case complexity from claim characteristics available at demand receipt — before the appraisal process even begins. Built on top of ML regression and classification models, with Claude (Anthropic) integrated for plain-English analysis and follow-up Q&A on each prediction.
The feature set was engineered from my own appraisal experience: what's known at demand, what signals are predictive, and what a practitioner actually needs to know going in. The LLM layer translates model output into the kind of summary an adjuster or counsel can act on.
An end-to-end fraud detection pipeline that approaches the problem the way an experienced investigator would — engineering features around billing behaviors, patient patterns, and physician relationships that signal fraudulent providers — then applying ML to scale that intuition across thousands of claims.
SHAP explainability is built into the workflow from the start, because flagging a provider without a defensible explanation isn't useful in a real investigation. The system trains three models (Logistic Regression, Random Forest, XGBoost) and selects the best, producing feature importance plots that make predictions auditable.
Houston, TX