work
selected systems
AI evals
Evaluation infrastructure for Spaces
Brex, 2026
Built evaluation infrastructure for Spaces, Brex's AI financial analyst, including a 200+ scenario suite covering SQL correctness, chart generation, insight quality, and output formatting. The system combines deterministic checks, narrowly scoped LLM judges, and trace review so product iteration does not outrun measurement.
Semantic layers
Semantic layer and AI infrastructure for Spaces
Brex, 2025–2026
Worked on the semantic engine and agent infrastructure behind Spaces: translating warehouse data into business-facing concepts, enforcing safer access patterns, and exposing financial reporting workflows across Brex's product surface, MCP infrastructure, and various external integrations.
Why the semantic layer is execution infrastructure · Brex journal · Cube case study
Metadata and provenance
Metadata catalog evaluation framework
Brex, 2025
Developed an evaluation framework for metadata catalog tooling across discoverability, lineage, and governance. The goal was to ground platform build-vs-buy decisions in real access patterns and technical-user workflows instead of vendor demos.
Streaming systems
Zero-downtime versioning for Flink
Brex, 2025
Designed deployment and versioning workflows for stateful Flink pipelines, with concurrent job promotion, canary outputs, and rollback paths for migrations where failed restores could mean multi-day backfills.
Data platforms
Mortgage data platform behind Credit Karma Home Loans
Credit Karma, 2021–2024
Architected and led a Kafka-based mortgage application monitoring platform for visibility after partner handoff. It became a core system behind Credit Karma Home Loans and later shaped a broader platform pattern for loan lifecycle analytics.
talks
How Brex Built AI-Native Financial Reporting
Agentic Analytics Summit (Cube) · 2026
How we built Spaces, Brex's embedded AI financial analyst, on a semantic layer, and what it takes to make an LLM agent answer real financial questions correctly: the ontology, certified queries, and the evaluation infrastructure that keeps it honest.