Work

Architected and led Kinesso's next-generation audience data platform on Snowflake, replacing a legacy Elasticsearch system at scale. I drove the technical strategy, team execution, and stakeholder alignment across the full platform lifecycle.

NeXus spans 60+ markets, holds hundreds of terabytes, and resolves hundreds of millions of consumers. It powers B&AA, the product where client and sales teams build and activate audiences for pitches and live campaigns, with hundreds of client workspaces and tens of thousands of audiences built on top. That makes it revenue-linked infrastructure, not back-office plumbing.

• NeXus 1.0 — Directed the core migration from Elasticsearch to Snowflake. The goal was to get off the legacy stack as fast as possible, so I adapted a data model that could reuse the data we had already built for Elasticsearch, which made for a quick migration. That move eliminated a roughly $2M/yr always-on infrastructure cost base and re-platformed onto elastic, pay-per-use compute, driving 40%+ annual savings with query times up to 95% faster.
• NeXus 1.1 — Spearheaded the extension of the platform to support survey datasets, solving a fundamentally different join model at scale. Achieved 85x reduction in data scanned and ~25x faster query execution.
• NeXus 2.0 — Defined the architecture for a wide/column-based data model, eliminating upfront pre-materialization for complex fused datasets and enabling new categories of data providers.

Oversaw the design and delivery of a governed, API-driven ingestion service that handles end-to-end onboarding of first-party and third-party data into NeXus, from identity mapping and transformation through metadata registration and audience publication. Authentication runs on an OAuth layer built with AWS API Gateway and Cognito.

I also led a single-tenant deployment of SANDS that runs in fully isolated environments, built for first-party data with no PII in scope. It extends the platform to client setups the shared multi-tenant model was never meant to serve.

Led the data engineering team delivering the audience-building platform on Elasticsearch, an inherited architecture. Directed the pipeline and infrastructure optimization that kept it reliable at scale, cutting data-processing time by 40%. Recognized where Elasticsearch would break at higher volumes and made the case for the migration to Snowflake, which became the NeXus program.

A prototype I researched and built to test whether a multi-agent system could turn a plain-language brief into a finished, validated audience. It is not in production yet. I took it end to end as a working proof of concept: it reads the brief, extracts the targeting criteria, runs entity-aware semantic search across the taxonomy, assembles a boolean audience tree, and returns geography, composition, and overlap validation on the result.

I designed it to run entirely inside Snowflake: LangGraph orchestration on Container Services, Cortex LLMs for extraction and reranking, vector embeddings for semantic search, Hybrid Tables for agent session state, and a Streamlit-in-Snowflake interface.

Led the creation of a suite of Claude Code skills the data team uses every day, wired to Snowflake and Jira over MCP. They automate the work that used to eat hours: standing up data-ops tickets, auditing query performance across the warehouse, running multi-account security and DBA checks, profiling tables, and producing a morning brief over Jira plus 24 hours of Snowflake health. The idea is to encode senior data-architecture judgment into tools the whole team can run.

Owned the delivery of an automated audience system on Redshift that served enterprise clients across retail, auto, and media. Config-driven pipelines and a lightweight DSL let teams stand up new client audiences without re-engineering each one, with access governance built in from the start. I also built the Scala and Spark encoding layer on AWS that made very large audiences fast to assemble, and integrated a wide range of third-party data sources behind it.