Case Studies
Production systems I worked on at Peloton from 2019 to 2026, and the independent work I’ve done since March 2026. Each row names the problem or the decision; the case studies show what changed.
Production experience
Peloton, Sep 2019 – Feb 2026: commerce, checkout and reliability.
Cart-Checkout Platform
Rebuild the checkout monolith, or buy the commerce backbone and build one checkout layer in front of it: the second let funnels move across one at a time.
Warranty Checkout Migration
Which funnel goes first: small enough to recover from, complete enough to exercise every part of the checkout path.
Catalog Migration to Commercetools
A multi-million-dollar contractor proposal, or a pipeline the team could already debug on call: build or buy for permanent infrastructure.
Checkout Observability & Reliability
Incidents were an ownership problem before they were a tooling problem, so the ownership map came before the monitors.
Legacy Catalog Service Retirement
A redundant service nobody could justify a quarter to remove: argued as risk instead of cleanup, then retired consumer by consumer.
CI/CD and Testing Modernization
A flaky end-to-end suite had trained the team to ignore red runs; auto-waiting removed the cause instead of the symptom.
Independent work
Since Mar 2026: developer tooling for AI coding agents, applied AI, evaluation and data.
AgentGrid: coordination for parallel coding agents
Two agents editing the same file is a coordination problem, not a prompting one, so the rules live in a deterministic policy layer with no model logic.
agent-playbook
How the work is split when agents write code: an Architect session plans vertical slices, each slice gets its own worktree and worker session, a fresh Reviewer checks every PR, and no agent merges.
Cortex
Getting model output that code can use: three passes on local models, JSON into typed records, and explicit fallbacks when a response doesn't parse.
AI coding-model evaluation
Judging generated patches the way a reviewer would: apply them to the repository, run the tests, and score correctness and engineering quality.
Scheduled data pipeline
Python on OCI with PostgreSQL: runs on a schedule, scores what it collects, alerts on it, and feeds a dashboard.