Brian K

Lead engineer for the document intelligence system behind Complya’s healthcare compliance platform, taking work from problem framing through production rollout. Built a RAG pipeline over 40,000+ regulatory and accreditation documents, improving answer relevance from 0.61 to 0.84 on a held-out set via chunking fixes and hybrid retrieval with re-ranking. Fine-tuned an open-weight model with LoRA on 12,000+ labeled credentialing and policy examples and delivered frontier-quality performance on the classification task while reducing inference cost. Implemented an evaluation harness for every model change using automated metrics, LLM-as-judge scoring, and compliance-specialist review. Reduced inference cost ~38% using prompt compression, response caching, request batching, and routing low-complexity queries to smaller models. Made training reproducible with versioned datasets, MLflow tracking, and a model registry tied to deployed data and commits for audit readiness. Deployed containeriz

Trained and shipped fraud-scoring models for contractor pay-outs, improving precision at fixed recall from 0.68 to 0.83 against the rules engine replacing models. Built Pandas and Spark feature pipelines feeding both offline training and online inference using the same definitions, covering transaction, pay-out, and KYC signals. Served models through a FastAPI service behind Kubernetes at scale (300,000 requests/day) with ~45 ms median latency. Orchestrated training and data workflows in Airflow and used upstream validation to catch schema breaks before poisoning training runs. Implemented prediction logging, feature drift detection, and performance dashboards to cut detection/response time from weeks to under a day. Ran A/B tests on production pay-out traffic to validate model changes and communicated results to product and risk stakeholders.

Built and validated risk-scoring models for insurance claims and exposure data using scikit-learn and XGBoost, starting from exploratory analysis. Prepared training data from claims and policy systems, handling missing values, class imbalance, and mitigating label leakage that inflated early results. Wrote Python pipelines that replaced analysts’ manual scoring workflows. Built dashboards to make model behavior understandable to underwriters. Improved team standards for experimentation and reproducibility through code review and research discussions.