At Indicium AI, I built ingestion and transformation pipelines that integrate 13 operational data sources into Databricks Bronze, Silver, and Gold layers, with data-quality and governance controls. I also developed reusable Claude Code Skills and evaluated AI workflows for legacy-code analysis and code review.
I built Doris, an independent Python/FastAPI LLM assistant that routes quantitative questions to governed SQL and uses hybrid retrieval for rule and methodology questions. Its reliability features include citation validation, fail-closed behavior, MLflow tracing, and automated tests; retrieval benchmarking reached an MRR of 0.631 across 14 rule-identification queries.
Previously, at Appen China, I led two QA teams across nine ByteDance AI projects, coordinating evaluation guidelines and root-cause investigations. I also trained reviewers and helped raise success and acceptance rates from approximately 63–68% to 94–98%.
