I build AI systems, and the evals that show whether they actually work.
Senior software engineer, San Francisco. Seven years at Carta: when the CEO said move the company onto AI, I built its first cross-product agent, brought Carta into Claude via MCP, and taught the org to work with agents. Oxford MSc on evaluating LLM assistants.
nowRunning my life on a compound Claude Code system — skills, loops, and a knowledge base that handles everything from taxes to this job search — and teaching executives to do the same, one day at a time.
actBuilt Carta's first cross-product AI agent as a hackathon prototype on internal Model Context Protocol (MCP) servers, pitched it to the CEO, and grew it into a company-backed, multi-team program. Then shifted the program to bringing Carta into external assistants: the Carta→Claude connector, via MCP.
eval✓Carta inside Claude — personalization and handoff, shipped in three weeks
✓Agent service and runtime — tracing, CI/CD, agent-to-service auth
✓Created and led the org-wide “Working with Agents” trainings; AI on-sites with Carta's largest customers
loop 2023–25
Measure whether an LLM system works before trusting it.
actOxford MSc dissertation — built a retrieval-augmented (RAG) assistant over health records, then tested every combination of chunk size, retrieval depth, and embedding model against factual, multi-document, and contradiction queries: 540 controlled runs, with formal statistical testing.
eval✓Isolated which design choice moved accuracy — on the 2025 models tested, reasoning, not retrieval, was the bottleneck on hard queries
✓Failure-mode taxonomy — hallucination, missed context, contradiction — later applied to production agents
✓Peer reviewer for digital-health journals (JMIR)
loop 2019–25
Ship across the platform without breaking it.
actBuilt the prototype that became Total Compensation, presenting to executives weekly as the team formed around it; then Cap Table, LLC, Stakeholder Management, and the HRIS layer that feeds them all.
eval✓Total Compensation, prototype to product line — founding backend engineer
✓Two migrations off PII-heavy legacy data models, run in shadow mode until the diff was clean
loop 2014–19
Learn where it hurts.
actAnti-spam ML at Kik (tens of millions of users); ad-tech data at Sortable; WordPress sites end-to-end, solo.
eval✓Spam reports down ~70%, false positives held flat
Play
./originSmall-town Canada. Studied development economics before software; I care about what systems do to the people inside them.