Weeks 1–2: the detective phase
The engagement starts with a question the plant thinks it has already answered: "we want AI on the inspection step." The FDE's first job is to test that assumption on the floor. That means walking the line at different shifts, because factories behave differently at 3 p.m. and at 3 a.m. It means discovering that the official quality metrics come from a system nobody trusts, while the real record is an Excel file the third-shift supervisor maintains by hand.
It also means the first hard conversation with the production chief, whose line stops for no one. The FDE who wins here does it by trading in the plant's own currency: scrap rates, changeover minutes, missed OEE points. Two weeks in, the outcome is a scoped decision — one process step, one measurable number, one deadline.
Weeks 3–4: building on real data
Industrial data is nothing like benchmark data. The vision sensor on the line has been miscalibrated since March, and the historian logs a tag that was renamed two years ago. So the build phase starts with plumbing: pulling from PLCs, reconciling the hand-kept records with the automated ones, and labeling images with the person who actually knows what a defect looks like — usually a veteran operator, not a quality manual.
The modeling itself is often the easy part. What consumes the FDE's judgment is the engineering around it: where inference runs (edge box on the line, not a cloud round-trip), what happens when the model is unsure (route to a human, never silently pass), and how the system degrades when a camera fails mid-shift.
Weeks 5–6: production and handover
A system is not deployed when it works; it is deployed when the plant stops calling. The last stretch is spent running in shadow mode next to the existing process, comparing calls with the line's inspectors, and tuning thresholds with the people who will own the system. The handover artifact that matters is not documentation — it is an operator who can retrain a threshold and a maintenance tech who knows which cable to reseat.
| Phase | Working weeks | Deliverable |
|---|---|---|
| Detective phase | 1–2 | Scoped use case with a measurable target |
| Build on real data | 3–4 | Pipeline + model running in shadow mode |
| Production & handover | 5–6 | System live, plant team operating it |
Timelines assume a scoped single-line use case; multi-site rollouts run in quarters. The pattern, though, repeats: specificity in, working system out.
The skills the job actually uses
- Process fluency — reading a line the way a controls engineer does, workarounds included.
- Data plumbing under constraints — PLCs, historians, hand-kept spreadsheets, and the politics of each.
- Deployment engineering — edge inference, fallbacks, degradation modes an operator can trust.
- Floor credibility — the ability to be believed by a shift crew, which no course teaches.
Roles asking for exactly this mix are in the open searches right now, with base ranges published on every one. If the week-by-week above reads like your last project, one interview puts you in the pool — and every future search finds you by itself.
FAQ
Is a forward-deployed engineer a data scientist?
No. A data scientist optimizes models; a forward-deployed engineer optimizes outcomes on a specific production line. The FDE spends more time with maintenance crews and PLC data than with model architectures, and success is measured by a system running unattended, not by accuracy on a test set.
How long does a typical industrial AI deployment take?
For a scoped single-line use case, working weeks four to eight are typical from arrival to a system running in production under supervision — provided the FDE controls the scope. Multi-site rollouts take quarters, not weeks.
Does an FDE need to write production code?
Yes, but less than a software role: pipelines from PLCs and historians, model serving on edge hardware, and fallback logic operators can trust. The scarce skill is judging what to build, which comes from plant experience.