The provenance layer beneath frontier models.

Member of Technical Staff · San Francisco, CACurrent

• Work on model provenance and model governance — the systems that make it possible to say, precisely and after the fact, which datasets went into a given model and under what terms they were used.
• Build dataset lineage tracking across the training pipeline, so the composition of a training corpus is a recorded fact rather than institutional memory.
• Support training on data under strict privacy protections, including workloads that run inside secure enclaves where the data itself is never directly accessible.
• Build the data infrastructure that prepares training data: storage layout, bucket organisation, and reliable transfer across account boundaries at the scale frontier training demands.
• Own data deletion and retention as an enforced property of the platform — built on Apache Iceberg and Spark, so a deletion commitment resolves to jobs that actually run.

Senior Software Engineer, Data & ML Platform · Cupertino, CATech Lead

• Led design and implementation of the Data Retention Service across Apache Hive, Iceberg, and S3 datasets, using Spark-based enforcement jobs to hold data compliance at petabyte scale, with first-class integration into the Lakehouse platform.
• Built AI-powered agents that give teams a conversational path through policy creation, onboarding, and troubleshooting — RAG pipelines over internal knowledge bases, Splunk logs, and MCP servers for automated diagnosis.
• Developed a partition pattern detection system using LLMs to auto-suggest retention policies, accelerating governance onboarding and cutting manual configuration errors.
• Provided technical leadership for Apache Iceberg adoption, shaping Lakehouse table format standards and cross-functional practice for large-scale data management.
• Built a semantic search system on Ray, LangGraph, and a vector database for embedding-based dataset discovery, turning metadata retrieval into governance insight.

Senior Software Engineer, Observability · San Francisco, CA

• Worked on the company-wide monitoring and alerting system, backed by an in-house timeseries database with aggressive compression that stores every metric across the fleet — 22 billion per minute — along with the read and write infrastructure around it.
• Drove a key part of the company-wide hybrid cloud investment by extending the Observability stack to support Prometheus-style dimensional metrics with labels, including structured querying of dimensional and histogram metrics in the query service.
• Shipped reliability improvements that took a stateful, distributed temporal key-value indexing service from 97% to 99.99% success, making container and metric indexing dependable across the data center.
• Developed Kafka-based real-time indexing, cutting p99 delay for newly seen metrics from 30 minutes to 10 seconds.
• Led the indexing service's lossless, zero-downtime migration from bare metal to shared cloud — a service holding 1 billion metrics in memory and serving 14 million Thrift RPCs per minute.
• Extended the timeseries query language and scaled the query engine to 100M+ unique timeseries reads per minute, while reducing false alerts from 50/day to under 5/week.

Progression: joined as Software Engineer I and grew to Senior Software Engineer over three and a half years on the same team.

Safety application stack for connected autonomous vehicles (CV2X)

• Built an intelligent transport systems stack for connected autonomous vehicles transmitting Basic Safety Messages over the IEEE 1609 protocol.
• Wrote a C library for parsing and interpreting received safety messages, with detection algorithms driving real-time safety warnings.

Resource-based REST APIs for Unified Computing System Director

• Built a workflow exposing virtual machine and cluster operations as REST APIs, using Java reflection and annotations to reshape the existing architecture into a single-pane-of-glass model.
• Designed role-based access control for the generated APIs to guarantee secured access.