At Marsh, I designed and own an agent platform using Semantic Kernel to answer natural-language questions over ERP accounting data, automate manual steps, and generate reporting. Lightweight classifiers for request routing cut model costs 30%.
I also designed and own the platform’s AI observability layer and SSRS reporting environment, and established Azure DevOps CI/CD across 10 applications. I own the production Windows Server environment and technically lead a team of five across engineering and analysis.
I built Reader++, an iOS PDF and EPUB reading app with document-grounded AI Q&A, hybrid retrieval, and study tools. Its retrieval evaluation framework improved top-5 evidence retrieval from 65.8% to 79.3% over BM25 across 24 documents and 429 generated cases, with no unsupported-question false positives in the benchmark.
