One thing has always had a habit of leading to another.

Full-chip layout on flash memory products, working to constraints that are physical rather than logical: area, timing, manufacturability. I led four full-chip projects, which was the first time the work was as much about coordinating a system of dependencies as about the engineering itself.

The other thing I remember has nothing to do with layout. The team needed somewhere to keep documents, so I built one in PHP and MySQL — picked up on my own time, at a job that had never asked for it. That turned out to be a pattern.

Design kits and technology files across a long run of process nodes, and the reference-flow work that decides how a design flow behaves before anyone runs it.

This is where automation became the job rather than a convenience. Scripting a flow, checking it, and making the results repeatable is a different discipline from designing the chip — and it is the one that carried forward into everything after. Three papers came out of these years, on physical verification and lithography inside the design flow.

A start-up, and the first cloud work: an API that let EDA software run on AWS, and the backend automation carrying a customer’s design from their desk to the cloud and back. Fully parallel design-space exploration and timing closure, which ended up as a co-authored paper at DAC 2012.

Short, and disproportionately formative. It established the instinct that compute is elastic, and that the interesting constraint is usually orchestration rather than capacity.

A year outside semiconductor, on market-data systems where an outage is measured in money. Monitoring automation, rollouts, and the investigation work when something broke: backtraces, log tracing, capacity review, server tuning. The task I remember best was unglamorous — cutting false alerts by going through each one’s timing, the binary behind it, and the algorithm that raised it. An alarm nobody trusts is worse than no alarm.

I got this role largely on Linux I had taught myself at home. Owning something in production teaches you to design for the day it fails, and that is a habit rather than a technique.

Execution frameworks for semiconductor manufacturing workflows, enterprise web services, integration APIs, workflow monitoring, centralised credential management — systems architected from a blank page and then owned in production as the person called when they broke. Underneath sat the unglamorous plumbing: queue-driven processing, event handling, asynchronous execution for anything long-running.

One project pulled on the whole career at once. A metrology application I took over was not producing good results, and fixing it needed the IC design knowledge from 2003, the software engineering from every year after, and a machine learning approach to something conventional methods handled badly. Ten years of living with what you build is the part that changes how you design.