Chief Data Officer Canada

Decisions about records, models and the people who use them

Serialised parts in one genealogy record: 2.6m; Six plants, parts made since January 2020
Answer to a 3,800-part traceability drill: 3 h 10 min; March 2026; the 2022 customer question took 23 days
Tessandier Aerospace revenue: C$1.18bn; 2025; six plants in Quebec, Ontario and Mexico

Decisions & consequences

Building a serial genealogy record instead of replacing every plant system

In September 2022 an engine customer asked us to identify every part we had made since 2016 from fourteen heat lots of titanium bar whose mill certificates were under investigation. It took twenty-three days and thirty-one people in four plants, working from three execution systems, scanned files and paper travellers. Our first answer listed 1,162 parts. Eleven days later we corrected it to 1,140, because 22 reworked parts had been counted twice. The material was eventually cleared, but the customer's supplier quality lead warned that a second slow answer would affect our standing on a new engine programme.

The obvious response was a group-wide execution system. Our chief operating officer, Yannick Desormeaux, had a C$48 million proposal ready, and it had real merits: one way of working on every shop floor. I argued instead for a genealogy layer above the existing systems, linking material certificate, heat lot, work order, operations, special-process runs, inspection and shipment for each serial number, at a budget of C$12.8 million. Mirabel's general manager opposed any new data-entry steps for his machinists. Our chief financial officer, Laurent Bégin, wanted the cheaper option but doubted my estimate. The board approved my proposal in February 2023, and the execution-system replacement was deferred indefinitely, with the inconsistent shop-floor practice that implies.

We started at Boucherville, because special-process records carried the most risk. We added validity rules at source, such as a heat-lot format check that blocks an operation from closing, and digitised 410,000 traveller pages from 2016 to 2019 for parts still on active programmes. Quarterly timed drills, reported to the audit and risk committee, test the result.

The record now covers 2.6 million serialised parts made since January 2020. Heat-lot completeness rose from 91.4% to 99.7%. In March 2026 a drill covering 3,800 parts in four plants took three hours and ten minutes, with no corrections afterwards. The programme cost C$14.2 million, above budget, and older parts still take days.

What I would do differently is the scanning. I hired a document contractor for the first batch, and its staff could not read handwritten amendments or inspector stamps. Experienced inspectors, paid overtime, did the work properly later. I should also have told the board about the pre-2016 gap when it approved the budget.

Narrowing a spindle-prediction programme the plant had already announced

In 2023 our operations team proposed a fleet-wide predictive maintenance system for the spindles on our 64 five-axis machining centres at Longueuil, Mirabel and Querétaro. A spindle failure stops a machine for days and costs C$90,000 to C$160,000 to repair, and the vendor's business case claimed annual savings of C$6.1 million. The rollout had already appeared in the operating plan presented to the board, and the vendor offered a C$0.8 million discount for signing before the end of the quarter.

My team ran the model in shadow mode on 22 machines for six months. It produced 47 alerts, of which 11 corresponded to faults that developed within thirty days, and it missed three failures entirely. When we examined the fleet's history, we found 58 recorded spindle failures in six years, and 27 of those work orders had no usable cause code. The model had very little to learn from.

I recommended against the fleet rollout. Instead I proposed rebuilding failure coding with the maintenance planners in all three plants, installing condition monitoring with fixed vibration and temperature limits on all 64 machines, and keeping the prediction model on the 14 Longueuil machines with the cleanest histories. Yannick Desormeaux, our chief operating officer, disagreed strongly. He had told the board the programme was coming, and he thought my standard of evidence would stop every improvement in the plants. Our chief executive supported my recommendation on condition that I explained the change to the board myself. We gave up the discount and the headline saving, and the operations plan lost a year.

By 2025 unplanned spindle downtime across the 64 machines had fallen from 4,900 hours in 2022 to 3,580, a reduction of 27%. Spindle replacements fell from 19 to 14 a year, and the saving was about C$2.2 million a year. After retraining on recoded histories, the Longueuil model's alert precision rose from 23% to 52%. Our investment policy now asks how many coded failures a model will learn from before any predictive programme is approved.

What I would do differently is protect the planners. At the peak, condition alerts reached forty a week at Longueuil, and planners began closing them unread. We should have capped alerts per planner from the start and tuned limits machine by machine before switching them on.

Leaving the data with the plants and taking the definitions to head office

In 2023 our board asked for a single dashboard comparing the six plants on scrap, first-pass yield and equipment effectiveness. When my team assembled the figures, we found six definitions of scrap. Granby counted rework as scrap; Querétaro excluded scrap caused by customer drawing changes; Mirabel recorded scrap when a part was quarantined rather than when it was condemned. Scrap reported by the plants differed from scrap cost in the general ledger by 11%. Querétaro, which head office regarded as our weakest plant, was being judged on figures that were not comparable.

Laurent Bégin, our chief financial officer, wanted a central team to take raw data from every plant and calculate one set of figures by the end of the quarter. Sylvain Arcand-Moreau, the general manager at Mirabel, wanted no comparison published until the definitions were settled, and he feared central analysts would misread his plant's data. I proposed a third course: the plants would keep their data and its local detail, and head office would own a dictionary of shared definitions, each with an owner, a formula, its exceptions and a monthly reconciliation to the ledger.

The decision cost time and some control. The board dashboard arrived three quarters late, in the third quarter of 2024, and our chief financial officer had to explain the delay to the audit and risk committee. Head office gave up direct access to raw plant data for most purposes. Each plant general manager and I signed a short data agreement in October 2023, and my office funded half the cost of eleven plant data stewards.

The definitions committee agreed 46 terms. Plant-reported scrap now reconciles to the ledger within 0.6%. Querétaro ranked second of six on first-pass yield once rework was counted consistently, and the plant's investment request was approved the following spring. Group scrap cost fell from 3.9% of revenue in 2023 to 3.2% in 2025, mostly through operations work that the comparable figures made possible. Plant managers have since stopped contesting the numbers in reviews.

What I would do differently is the downtime codes. Our first list had 140 reason codes written by engineers, and operators chose Other 38% of the time. We rewrote it with operators and their union representatives into 38 codes, and Other fell to 7%. We should have started with the people who press the button.

An inspection-assist model that ranks parts but cannot pass them

Our Boucherville plant inspects about 180,000 parts a year with fluorescent penetrant, a special process in which certified inspectors examine parts under ultraviolet light for surface indications. In 2023 customers detected three escapes from that line, all on rotating engine parts, and inspector overtime was rising. My analytics team proposed a camera system and a model that would flag likely indications, rank the queue and highlight areas for the inspector.

Two disagreements shaped the design. Nadine Béliveau, our vice-president of quality, wanted the model's records to produce individual miss rates for each inspector, to direct retraining and, where necessary, discipline. The union objected, and I agreed with part of the objection. I also had a privacy question: the cloud service we planned to use processed images outside Quebec, and the records would identify inspectors. Our operations team, meanwhile, wanted to use the model to remove the second inspector's review on rotating parts, which would have saved about C$0.9 million a year.

I decided that the model would never accept a part: it orders the queue and marks areas, and a certified inspector makes every disposition. The second review on rotating parts stayed. Inspector identities are held on our own systems in Quebec and pseudonymised before images leave the province, and we completed a privacy impact assessment before the first transfer. Results are reported by shift and certification level; individual feedback goes through the Level 3 inspector's coaching, not into personnel files. Because the change touched a special process, two engine customers required notice and approval, and one took seven months.

In 2025 inspection throughput rose 16%, overtime on the line fell by a third and customers detected no escapes from penetrant inspection. Inspectors found six indications the model had not flagged, which is why the model will remain an assistant. Nadine Béliveau accepted the reporting design after the first year, and her team now uses shift-level patterns to schedule refresher training.

What I would do differently is the labelling. Engineers labelled the first training images, and they disagreed with our Level 3 inspector on 9% of them. The Level 3 inspectors should have labelled from the first day. I would also have opened the customer approval process at the design stage, not after the pilot, which would have saved several months.

Discuss a leadership challenge