Zhekai Li · Capacity & Demand Planning
For three years that has meant capacity planning for semiconductor fabs in the US, Japan, and Singapore: turning demand into equipment and CapEx, allocating capacity with optimization, and running the plan-vs-reality loop that shows exactly where a plan and the floor disagree.
Four habits I bring to any capacity problem.
Keep what demand will need, what capacity to build or buy, and which tools serve which products as separate, reconciled objects. Most planning arguments are two of these being mistaken for each other.
Score every projection against what happened, then triage each variance: bad input data, a known model gap, or a real change on the floor. Each goes to a different owner: a data fix, a model change request, or a process problem the floor has to solve.
A tool that is up is not a tool that is producing. Decompose output into availability, occupancy, and rate before planning against it — the loss is often rate, not downtime, and a tool count on paper hides all three.
If a forecast cannot beat a flat line at last week's average, its daily numbers carry no information. Measure the error, keep the baseline in view, and prefer simple models you can inspect over clever ones you cannot.
How these are masked. Each screenshot is de-identified at the source — tools, sites, lots, suppliers, and internal systems become aliases — and every number is masked after rendering; chart labels are drawn as pixels, so charts are blurred. A screenshot is refused if any digit or protected name is still visible on the page. Results in the case studies are stated as ratios, not operating figures.
I am a senior industrial engineer at Skyworks Solutions, where I own capacity planning and fab simulation for wafer fabs in the United States, Japan, and Singapore — the models, the data behind them, and the software people use to act on them.
What I care about is plans that survive contact with reality: forecasts scored against a naive baseline, simulation whose error is measured instead of assumed, and capacity models that keep what demand needs apart from what we choose to build and what each tool actually runs. I write the SQL and Python myself, and I would rather retract a finding than let a wrong number steer a decision.
Before this I built a fab discrete-event simulation engine from scratch as an intern; it became the foundation of the company's simulation program.
Email, phone, or LinkedIn — whichever is easiest for you. Based in Thousand Oaks, California.