Divvy

Overview

Divvy is Carvana's internal experimentation platform … the system nearly every product team relies on to run A/B tests and make data-driven decisions before shipping. It buckets users into experiment variants at the edge and turns the results into evidence teams can act on.

At scale, that means 10M+ bucketing requests per day, 1,000+ concurrent experiments, and sub-50ms assignments at a 99.9% uptime SLA.

I own the platform end to end … spanning the Divvy Dashboard (a NextJS app for experiment creation and config), a fleet of NodeJS and .NET backend services, a Kafka-driven event pipeline into Snowflake, the monitoring and observability layer, and the SDKs product teams build against.

I’ve been on the Experimentation team since I started at Carvana and I’ve been fortunate to work on a talented team of A-Players that make this product awesome.

Why it matters

Experimentation is how Carvana ships with confidence. Product teams run controlled tests, measure impact against real KPIs, and roll out only what's proven — instead of shipping on instinct. Divvy is the infrastructure that makes that the default way of working.

Four parallel threads of work

1. Platform stewardship across the stack: Keeping a high-throughput, low-latency system healthy — sub-50ms assignments and a 99.9% SLA — through deep observability, disciplined dependency management, and fast incident response.

2. Modernizing the Dashboard: Bringing the frontend up to the NextJS App Router, React 19, retiring legacy patterns, and making the platform's primary interface faster and easier to build on. I pro-actively lead an initiative to make deploying NextJS App Router apps possible within Carvana’s infrastructure.

3. Customer-driven feature work: Delivering new platform capabilities from the product roadmap … expanding what teams can test and how … while holding the line on stability.

4. AI for experimentation — Divvy Oracle + a platform MCP: Two efforts pushing the platform into AI-native territory:

• Divvy Oracle: an internal R&D prototype, born at a company hackathon, that predicts an experiment's winning variant earlier and with less manual number-crunching. Rather than a fixed reporting pipeline, an AI agent retrieves an experiment's results on demand and runs statistical analysis …. significance testing, confidence scoring, and a plain-English recommendation for what to do next. It's model-agnostic across leading LLM providers and can run on a schedule to surface early signal across active experiments.
• A Divvy MCP server (in development): built in direct response to demand from the teams who use the platform. It exposes Divvy to AI assistants through the Model Context Protocol, so engineers and PMs can spin up, inspect, and reason about experiments in natural language — right inside the tools they already work in …. instead of clicking through a dashboard. It turns the platform into something you can talk to.

Together they point at where Divvy is headed: experimentation that's not just measured by AI, but *driven* through it.

Outcome

A platform that quietly gets better and more reliable over time … one teams can trust without thinking about it. The goal has never changed: let teams run their experiments, read the numbers, and make decisions without ever having to think about the platform underneath.