Leadership review
Surya Koh
Chief Data Officer · Singapore / Southeast Asia
Highlights
158
Data colleagues in my remit
Seven direct reports at October 2026; 6700 employees across the company
S$21.4 m
Data-function operating expense
2026 delegation; plant capital and operating decisions remain separately owned
36
Years interpreting industrial records
Career began in August 1990 within the same fictional manufacturer
A letter from the chief data officer
I regard a code as the beginning of an explanation. It tells me how something was recorded, but I still need to understand the part, lot, machine or observation it represents. When records travel between factories, the local context can disappear before anyone notices that its meaning has changed. I want the combined account to preserve the distinctions a competent colleague would need in order to challenge it.
I joined Tervanex Components in August 1990 after completing industrial statistics at the fictional Neraveth Institute of Industrial Statistics. My first appointment was within a twelve-person analysis team. I contributed analysis within that team; colleagues and expenditure remained under my manager’s authority. The company was much smaller than it is in this invented present-day record, and a question about a production figure could often be taken directly to someone who knew how the record had been made. That access taught me to look beyond a column heading.
Surya Koh
Chief Data Officer
Strategic priorities
01
02
03
04
Performance
Priority outcome 01
Our 2023 parts programme began with 12400 entries and an understandable request for a simpler catalogue. The population contained 8300 current definitions, 2100 aliases and 2000 superseded definitions. A programme summary that counted every entry beyond the current definitions as duplication would have made the task look straightforward. It would also have ignored the different work those records performed. Earlier production and purchasing records still needed an intelligible connection to the definitions used at the time. Darin Veylun, our fictional manufacturing-systems director, favoured a smaller active catalogue. He argued that unnecessary choices slowed users down and that retaining too much history could preserve the confusion the programme was meant to resolve. I agreed about the user problem. My objection concerned the proposed measure of success. Reducing the number of visible choices and deleting historical meaning were not the same action, and a data programme needed to distinguish them before celebrating a lower total. I supported a design that separated current selection from the relationships needed to interpret earlier records. Aliases remained linked to their intended definitions, while superseded definitions retained their historical context. The S$3.7 million programme expense covered that work rather than a promised inventory reduction. Engineering owners examined the meaning of part relationships. A shared description or an alias did not give my function authority to declare two physical components interchangeable, particularly where the intended use or specification differed. The decision accepted a more complex underlying account in order to make the user-facing choice more intelligible. It also required ownership of the relationships after the initial clean-up, instead of treating the catalogue as a one-time editing exercise. We did not claim that all 12400 entries represented distinct physical parts, or that retaining an old definition made it available for current production. Those questions depended on the specific record and the authority responsible for its use. I would present the result as a change in how the catalogue represented its history, not a quantified saving that was never established. The competing proposal had a legitimate concern about usability. Our response needed to meet that concern without destroying the explanation behind earlier work. The case shows why a CDO has to understand the practical meaning of a record before deciding that a simpler dataset describes a better outcome.
Priority outcome 02
A 2021 review brought together logs from two hundred machines. Thirty-seven showed observed clock offsets ranging from four to eleven minutes. That finding mattered because analysts had begun to read the combined display as an account of the order in which events occurred. A timestamp could reflect the source clock, while another record arrived through a different collection path. Placing both on one timeline did not establish that the apparent sequence faithfully represented the physical events being compared. The programme team wanted a consolidated chronology quickly. Some colleagues argued that the offsets were small in relation to a full production day and that the dashboard should remain useful for identifying broad patterns. I accepted that a broad summary might still serve a defined purpose. It did not follow that the same record supported a more precise account of which event preceded another. We needed to explain the interpretation being requested rather than label the entire dataset either reliable or unusable. I supported S$1.9 million in review work to preserve the original records and document their timing context with responsible engineering and systems colleagues. The analysis kept source-event time distinct from arrival time where that distinction was available. It recorded unresolved cases instead of silently making every entry appear comparable. My role concerned the evidential account and its limits. Changes to machine settings, control systems or operating practice were decisions for qualified owners, outside the authority of an analytical narrative. The immediate consequence was a less definitive timeline. Some comparisons had to carry a qualification that made the presentation less convenient. That was preferable to a neat sequence whose precision came from the display rather than the records. We did not claim that adjusting an analytical interpretation established physical causation, nor that every unexplained interval had the same cause. The review expense bought a more explicit basis for reasoning, without producing an invented throughput or downtime result. I retained the case because it changed the questions I ask when data from several sites appears on a common screen. The display can remove visible differences faster than the organisation resolves them. A useful account should show which distinctions have been understood, which have been interpreted for a stated purpose and which remain open. That gives another professional a basis for challenging the conclusion before the apparent order of the data becomes a claim about the factory.
Priority outcome 03
In 2008 I endorsed a combined lot report with eighty-six unique codes and accepted a claim of complete traceability across the reviewed population. The preparation had removed plant context while bringing the records together. I treated uniqueness in the combined code column as if it established uniqueness of the physical production lots. That inference was mine to examine before the account was presented, and I had not asked the necessary question about how each plant assigned its identifiers. Mei Arvellan challenged the total using the underlying plant records. Twenty-six codes had each been used for a lot in two plants, while sixty codes appeared in only one. The physical population therefore contained fifty-two lots behind the repeated codes and sixty behind the others: 112 lots in all. The arithmetic was simple once the missing context returned. Its significance was that our apparently tidy combined report had concealed different physical entities under the same identifiers. Some colleagues preferred to repair the table and keep the original traceability conclusion, arguing that we had found a naming issue rather than evidence of a product problem. That objection contained an important distinction. The data error did not itself establish a product defect. It also did not leave the stronger traceability claim intact. I withdrew that claim and supported S$640,000 in review expense so that the account could be reconstructed on a defensible basis. Quality and plant authorities determined the operational examination appropriate to the records. I did not turn a corrected count into a statement about product safety, a recall or the absence of either. My responsibility was to explain the analytical failure, retain the original context and avoid claiming more than the revised evidence established. The review had a real cost in this fictional account, including the need to explain why my earlier endorsement had been too confident. The lasting change was in how I treated an identifier when records moved between settings. A value can be unique in the place where it was created without being unique across an enterprise. That is a question about meaning and scope before it is a question about a technical key. I keep this mistake in the public narrative because the later emphasis on context would be less credible if it appeared to have emerged without a decision in which I had failed to preserve it.
Priority outcome 04
A 2017 predictive-maintenance study used records from forty-two assets. The evaluation divided rows randomly, so observations from the same assets appeared in both the model's training material and the held-out set. The resulting presentation was moving towards a claim about performance on unfamiliar assets. I asked whether the evaluation had actually examined that use. Keeping individual rows apart did not necessarily separate the recurring characteristics of the machines from which those rows had been collected. The analytical team argued that the split was reproducible and that every evaluation row had been excluded from training. Those were relevant facts, but they answered a narrower question. If the intended use concerned another asset or a later operating period, the evaluation needed to reflect that distinction. A large number of rows could not by itself establish breadth across machines or time. I wanted the claimed population stated before the score became the central evidence for the proposition. I recommended withholding the unfamiliar-asset performance claim and supported S$2.3 million in further analytical work. The revised evaluation was to be designed around the intended asset population and future period, with the responsible specialists examining dependence and which inputs were available at prediction time. I did not prescribe a universal statistical split or imply that one evaluation design would settle every industrial use. The necessary comparison depended on the question the model was meant to answer. This response cost time and the immediate opportunity to present the study as ready for wider use. The team also had to explain why a technically reproducible result had been described too broadly. We retained the earlier work as evidence of what had been tested, rather than pretending it had no value. No later performance figure is supplied in this fictional record. A more suitable evaluation still needed to be carried out and assessed; asking for it did not establish a successful outcome. Plant and maintenance professionals retained responsibility for operational decisions. The study did not authorise my data function to alter maintenance practice or represent an analytical result as a safety determination. I would use the case to discuss the difference between a held-out observation and the genuinely unfamiliar situation a claim describes. A leadership profile should make that difference understandable without disguising uncertainty behind a model score or turning a cautious recommendation into an unearned success story.
Priority outcome 05
In 2025 we reviewed product-footprint submissions from a population of 270 suppliers. Of those, 144 provided measurements, fifty-nine supplied modelled estimates and sixty-seven had no submission. A single completion percentage would have been easy to report, but it would have hidden the different bases of the received material. It could also have encouraged users to treat an empty field as a zero when combining the records, although absence of a submission established no such result. The commercial team wanted a consistent account for customer discussions. Some colleagues argued that distinguishing too many states would make the response look unfinished and that estimates were necessary wherever direct measurement was impractical. I agreed that an estimate could be useful when its basis was understood. The issue was whether the combined account retained that basis. A modelled value did not become a measurement because it occupied the same column, and a measurement received from a supplier was not automatically independently verified. I supported S$4.6 million in data-review work that kept the three submission states visible. The account recorded where boundaries, periods and supporting explanations still needed examination by the responsible specialists. It did not force comparability merely to produce a complete-looking total. Procurement, sustainability and assurance professionals retained their responsibilities. My function helped organise the evidence and expose its limits; it did not issue an assurance opinion or certify the supplier's underlying process. The immediate result was a more qualified view of coverage. The sixty-seven missing submissions remained missing, rather than disappearing into an apparently lower footprint. The fifty-nine modelled submissions could be considered on their stated basis without being promoted into a stronger category. That made some customer-facing explanations more demanding, but it avoided making the polished account more confident than the material behind it. No emissions reduction, later completion count or net-zero achievement is claimed here. I see this as a question of intellectual ownership within a data programme. A CDO should be able to explain how the organisation represents absence and differing evidence without claiming the specialist conclusions that may eventually be drawn from it. The review's expense and incomplete coverage belong in the same account as the improved distinction. The purpose of the fictional case is to show what the leader chose to make visible when the easier presentation would have flattened materially different information.
Year in review
Tervanex Components
Lead the enterprise data function while preserving the meaning of parts, events and supplier evidence across six factories.
Tervanex Components
Connected industrial datasets with explicit evaluation questions, including the limits of a forty-two-asset maintenance study.
Tervanex Components
Developed cross-plant analytical accounts and owned the correction of a lot report that had lost physical context.
Tervanex Components
Led colleagues maintaining common definitions without removing the history needed to interpret local records.
Tervanex Components
Learned to connect a production figure with the local convention and physical work behind its source record.
Stakeholder voices
Surya accepted that our lot report had lost the plant context. He withdrew the stronger claim before asking the review to explain what the records could actually support.
I wanted a shorter parts catalogue. Surya made the history and alias relationships visible, so that we could discuss simplification without treating every additional record as unnecessary.
Outlook
I examine how industrial identifiers lose their scope when records travel, using an owned lot-count error and a parts catalogue whose history had to remain interpretable.
Read the essayI discuss what a common factory timeline can conceal about source clocks, arrival times and the sequence an analyst wants to infer from them.
Read the essayI examine why measured, modelled and missing supplier submissions should remain distinguishable when an industrial company assembles a product-footprint account.
Read the essayAn invitation
I welcome a well-defined discussion about the evidence behind an industrial data proposition, or a first-board opportunity where this experience fills a specific gap. The mandate, preparation and employer position would need examination. This is an invented profile with no live advisory service. A real leader’s version should offer an authorised introduction route and keep plant-sensitive, customer and supplier material out of public enquiries.
office@suryakoh.example