Work
Longer write-ups of systems I have designed and shipped. Figures are given as ratios and rates rather than absolute volumes or costs.
An LLM PII classifier, and the flywheel that fixed it
Roku · 2024 – presentBuilt in-house sensitive-data discovery for the ads data platform instead of buying it, then closed the loop between analyst corrections and training data to move precision from 84% to 91% without giving up recall.
An agentic Hive-to-Iceberg migration platform with mandatory human approval gates. Migrated 450 tables across 40 workflows with no production incidents, delivered by three engineers instead of roughly ten.
A governed MCP data-access layer over roughly 5,000 tables, enforcing schema, lineage, ownership, identity, and metric certification before a query runs. Adopted by 120+ platform engineers.
Deleting the tracker table
Roku · 2024 – 2025Re-architected hourly ad impression and conversion processing from mutable tracker state to stateless, idempotent Iceberg pipelines. Runtime down 35%, Trino latency down 20%, incident recovery from a day to two hours.
From 90 days to 5 — rebuilding GDPR deletion
Roku · 2024 – 2025Redesigned deletion across a petabyte-scale data lake using virtual-identifier rotation for immediate disassociation, cutting end-to-end fulfilment 94%. Separately re-architected data download from MapReduce to Spark for a 50× intake increase.
Productionised Slack incident-history and Confluence runbook-retrieval MCP servers for an internal AI incident assistant, with read allowlists, least-privilege scopes, and button-gated writes. Adopted by 20+ teams and 150+ engineers.