Architecture · Amazon
Interstitial Upsell System
16xsubscriber acquisition, 11K to 181K a month
Grew monthly subscriber acquisition from about 11,000 to 181,000 and conversion from 0.3% to 10%, across 7+ microservices.
- The decision
- I led architecture and delivery. We wrote up three options and I chose bounded orchestration: the upsell flow is coordinated in one clear place instead of being spread across every service it touches.
- The tradeoff
- Each option traded technical debt against delivery timeline. The bounded approach could ship on schedule without adding new coupling to every service in the path.
- What it changed
- The eligibility integration and experimentation setup let every later change ship behind an A/B test, so improvements come with a number attached.
Cross-team · Amazon
Grocery Subscription Landing Page
3teams, one flow
One customer flow, delivered by three teams that each owned different widgets and frameworks.
- The decision
- I designed the end-to-end architecture and turned the flow into explicit integration contracts, so each team could build in parallel.
- The tradeoff
- Keeping each team on its own framework avoided a rewrite. The cost was more upfront work defining exactly what crosses each seam.
- What it changed
- The contracts became the shared reference for the page, and funnel views made experiment results readable widget by widget.
Compliance · Amazon
FTC Double Opt-In
2engineers led, fixed legal deadline
Shipped on a legal deadline, with the rule built into the system so a future change can't quietly undo it.
- The decision
- I owned the system design and led two engineers, working as an away-team contributor in another team's services.
- The tradeoff
- The date couldn't move, so scope stayed narrow. The effort went into making the rule hard to break.
- What it changed
- The requirement lives in service contracts and tests. A change that skips it fails before it reaches a customer.
Redesign · Amazon
Physical Location Finder
Took over an undocumented system and rebuilt it on new AWS infrastructure with integrated auth and monitoring.
- The decision
- Before changing anything I traced the dependencies, auth and deployment. Then I led the redesign on that map.
- The tradeoff
- Mapping first cost time up front, but the redesign was based on what the system actually did.
- What it changed
- A system nobody could explain now has known infrastructure, real authentication, and monitoring that says when it's unhealthy.
AI agents · Amazon
On-Call Investigation Agent
↓drastically reduced investigation time
An agent does the first pass on on-call tickets, so engineers start from the relevant logs and context.
- The decision
- I encoded the team's on-call knowledge into custom agents and skills that gather logs and context before an engineer picks the ticket up.
- The tradeoff
- The agent does the first pass and hands off, rather than closing tickets on its own, so a person still makes the call.
- What it changed
- Investigation time dropped drastically, and it drove tuning of transient alarms that removed recurring tickets.
Project-status agent in Slack
Answers PM and stakeholder status questions from code changes, docs and threads.
Scheduled agents and MCP servers
Email digests, 1:1 summaries, and custom MCP tools in daily workflows.
