From the approved drawing to the welded joint

In 2021 a security review, prepared ahead of a naval customer's supplier audit, found that nineteen of our sixty-three CNC machines loaded programs from computers running an operating system its vendor no longer supported. Machinists moved programs on USB sticks. Four of the nineteen controllers also had failing parts that were becoming hard to source. Our production general manager and chief financial officer favoured replacing eleven machines at a quoted A$13.4 million, arguing that new equipment would add capacity and settle the auditors' concern in one move.

I recommended a different first step. We separated the machine network from the office network, installed a program server that released only approved, revision-matched programs, sealed the USB ports, locked the old computers to the applications they needed and retrofitted the four failing controllers. The cost was A$1.9 million over nine months. The trade-off was real: we gave up the capacity new machines would have brought, and machinists lost a quick habit they valued. The production general manager told the board that program loading would slow the night shift, and for two months he was right.

Finance also worried that the customer might still insist on replacement and leave us paying twice. I asked the customer's security team to review the design before we built it; they accepted it with two changes.

By month nine every program reached those machines through the server. Controller-related stoppages fell from 31 hours a month to nine, mainly because of the retrofits. The audit closed with two minor findings, both about documentation. In 2024 we replaced three of the nineteen machines, for A$4.6 million, on capacity grounds the board considered separately.

What I would do differently is involve the machinists before the first design. Our initial version required walking to a terminal at the end of each bay to request a program. A request screen at each machine, added in month three, should have been there from the start.

By 2022 our naval customers issued three-dimensional models with manufacturing information embedded, and revisions arrived as often as weekly while a structural unit was in build. Our shop still worked from printed drawings and paper travellers. In the 2022 financial year we recorded 41 revision-related nonconformances on defence work, and one batch of machined seatings made to a superseded revision cost A$212,000 to scrap and remake.

The general manager of our commercial division, which serves water and transport customers, proposed one cloud-hosted product lifecycle tool for all work, quoted at A$2.1 million and deliverable within a year. I did not support it for defence data. Some of the technical data we receive is export-controlled, and our customers' contracts set conditions on where it may be held and who may view it. Our export-control officer could not confirm that the proposed service met those conditions, and the vendor could not do so without qualifications.

I recommended building the path from model to traveller to inspection result for defence work first, inside a separately administered environment with named access, at A$3.6 million over two years. Commercial work would stay on paper for eighteen months longer. The commercial division lost a faster improvement, and we accepted the cost of running two systems. The chief executive and board approved the plan; the commercial general manager's argument for speed was sound, and I said so in the paper.

Revision-related nonconformances on defence work fell to six in the 2025 financial year. First-article inspection of a typical seating fell from 5.5 hours to 3.2, because inspectors received the correct model characteristics with the part. In 2025 commercial work moved onto a simpler, separate instance costing A$0.8 million.

I would now ask the export-control officer and the customer's security team to write the access requirements with us at the outset. We spent four months redesigning permissions that a joint workshop in the first fortnight would have settled.

In 2024 we prepared a tender for repeat structural units on a naval programme, worth up to A$46 million over five years. A large share of the welding was stiffener-to-plate joints of stable geometry. Our owners and the chief executive favoured three robotic welding cells costing A$7.8 million, which would have supported a sharper price. Adelaide was short of experienced welders, and the case for automation was strong.

My concerns were qualification and people. Every robotic weld procedure for naval work needs its own qualification testing before production, and we had not yet qualified one. Three cells would also need operators we did not have, and our welders' representatives feared the cells were a first step towards fewer welders. I recommended one cell, qualified for the two most repeatable joint types, with eight experienced welders trained as its programmers and operators at their trade rate plus an allowance. Complex nodes and closing welds would remain manual.

The trade-off was price. Our tender was about 3.5 per cent higher than the three-cell model implied. A family director argued that we were handing the contract to a competitor. Our welding supervisor, Kerri-Anne Mulcaster, doubted at first that the cell would ever match her best welders.

We were awarded a reduced share of the programme, about A$31 million, with another supplier taking the remainder. The cell cost A$2.9 million. Procedure qualification took seven months rather than the four I had planned. Once running, the cell reached 58 per cent arc-on time against 22 per cent for manual welding of the same joints, with first-pass acceptance of 97.4 per cent. Nobody lost a position. In 2026 the board approved a second cell on the evidence of the first.

What I would do differently is qualify the robotic procedures before the tender rather than during it. The board would then have priced the work on demonstrated capability instead of my estimate, and the reduced share might have been larger.