Chief Data Officer / United States
DavidKeldraven
I joined the fictional Vorendell Data Services in July 1992, shortly after completing applied mathematics at Merravane College of Applied Mathematics.
146
Data-practice colleagues
Seven direct reports at October 2026; client delivery and specialist authorities separately led
US$18.7 M
Delegated function expense
2026 data-practice remit, distinct from company revenue and client-contract authority
34
Years in one services business
Analytical career began July 1992; executive board presentations from March 2019
I believe
I want useful experience to travel between assignments without losing the conditions that made it valid. A pattern can be technically reusable while its source material or intended use still needs review. A benchmark can combine millions of records and compare unlike events. My contribution is to preserve those distinctions so that a services firm learns deliberately instead of mistaking possession of data for permission or proof.
— David Keldraven
The work
The 2023 library proposal at the fictional Vorendell Data Services examined 600 table definitions gathered from project experience. Owen Varleston wanted to move quickly because teams were repeatedly solving similar analytical problems. His argument was that a services firm should not rediscover the same structure in every assignment. I agreed with that ambition. I disagreed with treating the technical resemblance of two tables as a sufficient basis for carrying their definitions or associated material into another client setting. The programme needed to distinguish a useful candidate from an approved use. The first review excluded 190 client-specific definitions from the proposed common set and retained 410 technical candidates. Those labels mattered. The 410 were not a library of universally permitted assets, and the 190 were not evidence that their underlying work lacked value. The distinction concerned the proposed reuse. Source values were outside the common-pattern proposition. Legal, privacy, contractual and business owners retained their respective decisions; the data practice could describe the technical pattern and its history, but it could not grant itself a general right to reuse client material. We selected a staged approach with US$3.6 million programme expense. Each candidate moving forward needed its intended use and relevant owner review made explicit. Seventy-two patterns passed that process and were used in twelve projects by June 2024. The account kept those seventy-two separate from the 410 candidate total. It also retained the original 600-definition scope so that a reader could understand the choices made rather than infer that every item had reached a new assignment. I reported actual reuse without attaching an unmeasured return or assuming that the same approval covered every future context. The trade-off was slower availability and a smaller launch than Owen initially wanted. Some repeated work would continue while review remained unfinished. That cost was real, but it was preferable to a library whose apparent convenience depended on concealing its conditions. The outcome demonstrated that a bounded subset could travel into client work through the chosen process. It did not establish blanket permission, source-data portability or a financial saving. I learned to make the status of a reusable pattern part of the asset itself. If the next team cannot tell which use was reviewed, the company has transferred uncertainty along with its experience.
In 2020, a proposed service benchmark combined twenty-seven million records from fourteen fictional client assignments. The commercial attraction was a broad comparison of order-processing performance. A single ranking would be easy to present and could help clients identify where their process appeared slower than others. Before supporting that account, I asked what event started the measured cycle in each contribution. Eight clients used order approval; six used submission. Those were different points in the process, and the dataset's size did not remove the difference. Supporters of the ranking argued that the broad pattern might still be informative and that a prominent explanation could disclose the variation. I was concerned that the rank itself would imply a common basis. A reader could reasonably interpret a shorter cycle as a better result even where one cohort excluded the time between submission and approval. The issue was not a small adjustment around an otherwise settled comparison. It affected the question being answered. My team needed to preserve that distinction before the numbers became a persuasive cross-client story. We used US$1.8 million review expense and retained two cohorts rather than publishing one combined ranking. The account identified the starting event for each cohort and avoided a performance conclusion across them. Client and contractual owners remained responsible for the permitted use of contributed material. The data practice explained the analytical comparability problem; it did not treat participation in one assignment as general permission to expose a client's records or performance in another context. The public case therefore uses invented aggregate facts rather than actual client information. The choice gave up a stronger commercial headline and reduced the apparent simplicity of the benchmark. It did not mean the records were worthless or that no useful comparison could ever be developed. It meant this proposed ranking lacked the common event basis its interpretation required. The outcome was a more limited comparison structure with two explicitly different clocks. No cross-cohort superiority or subsequent client benefit is claimed. The case reinforced a discipline I would apply before a data asset is marketed: ask whether the units describe the same event, not merely whether they can be put into one column. Volume can improve precision within a meaningful question; it cannot make unlike questions identical.
The 2016 ticket-classifier test contained 10000 examples: 9600 routine tickets and 400 severe tickets. The model correctly classified 9400 routine tickets and 200 severe tickets. That produced 9600 correct classifications overall, or 96% accuracy. I presented the aggregate as evidence that the model was ready to support automatic severe-incident routing. Clara Fenwicke challenged the conclusion by asking how many of the severe tickets had been found. The answer was 200 out of 400, giving 50% severe recall. The headline was arithmetically correct but inadequate for the decision I had attached to it. I had allowed the much larger routine class to dominate the account. Clara's objection was not that overall accuracy had no meaning. It was that the proposed routing use depended on detecting the severe cases, and half of those had been missed in the test. My first presentation did not make that consequence visible enough. Adding the recall figure beneath the same readiness statement would not repair the interpretation. I needed to withdraw the statement and reconsider what evidence the use required, rather than defend the aggregate because it remained technically true. We withheld automatic routing and supported US$650000 review. The revised account included the class counts and correct classifications so that a recipient could follow both calculations. Delivery and operational owners retained responsibility for any later routing proposal, with model-validation expertise separately involved. I did not claim that the review produced a later performance result or that an improved model was eventually deployed. The documented consequence was the withdrawn readiness claim, the withheld use and the additional work needed to assess the issue properly. The decision delayed a capability I had advocated and required me to explain why my original account had been too broad. I accepted that cost. The case changed how I review a model summary: the measure must be connected to the consequence the proposed use cares about. A strong aggregate may coexist with a weak result in a smaller but important class. The public narrative retains that failure because it makes my later judgement assessable. It is a fictional example, not a universal acceptance threshold for classifiers, and it does not replace specialist evaluation of any real system's intended use.
A 2019 client-exit programme at Vorendell Data Services covered 340 data extracts. By the reporting point, 312 had documented closure under the agreed process. Twenty-eight still awaited client and legal decisions about their disposition. The project team wanted to describe the transition as complete because the main service work had ended and most of the inventory had been handled. I understood the desire to close a long-running assignment. I did not want the completion label to conceal material whose next step had not yet been authorised. The unresolved extracts were not interchangeable with the closed items. Their status depended on decisions outside the data practice's unilateral authority. Deleting everything to improve the completion count would have been as inappropriate as retaining everything under an unspecified general exception. We needed a state that preserved the outstanding decision without permitting ordinary reuse. The question was therefore not how to make the inventory total look finished, but how to make responsibility for the remaining items explicit while the transition proceeded. We retained the twenty-eight in a separately restricted pending state and reported the 312 documented closures as the completed subset. The US$1.1 million transition expense covered the programme, not a claimed saving from disposal. Client, contractual and legal owners retained their decisions, while the data team maintained an account of the inventory and status. This narrative does not specify a universal retention period or offer legal advice. It describes the fictional institution's decision to avoid changing an unresolved status into either assumed permission to use or an unsupported instruction to delete. The choice left an unfinished item visible after the wider engagement had ended. It required continued ownership and follow-up rather than allowing the project to disappear from attention with a celebratory closure statement. The outcome was an honest inventory at that point: 340 extracts in scope, 312 documented closures and twenty-eight awaiting the specified decisions. No eventual closure of all twenty-eight is claimed. I took from the case a practical meaning of data stewardship in a services business. The obligation to describe what remains does not end simply because the commercial assignment has reached its final milestone. An incomplete but accountable status is more useful than a completed label that no longer matches the material held.
In 2025, an internal test collection contained 60000 synthetic service examples spanning 540 known categories. The broad category coverage made the collection appear ready to support a wider evaluation. During review, we identified seventeen specified combinations of rare conditions that were absent. Colleagues who favoured proceeding argued that the dataset was already large and that its coverage of every named category was a meaningful achievement. I agreed that the work was useful. I questioned whether category presence alone answered the test question posed by those combinations. A category can appear somewhere in a dataset while a consequential interaction between conditions remains unrepresented. The missing combinations were part of the intended evaluation, not an arbitrary demand to imagine every possible event. Their absence meant the test could not examine the behaviour associated with them. Adding more examples of already represented situations would enlarge the dataset without necessarily filling that gap. I wanted the next work to follow the specified test needs rather than treating volume as a substitute for coverage of the relevant situations. We expanded the test work with US$2.4 million expense and withheld a claim of complete coverage. The data practice worked with the relevant delivery and model-review owners to describe the missing combinations and the limits of the existing collection. The material was synthetic, but that label did not by itself establish privacy assurance or suitability for every use. Those questions remained with the appropriate specialist processes. I did not report that all seventeen combinations were subsequently resolved, because the case records the decision to undertake further work rather than its eventual result. The trade-off was another phase of effort before the collection could support the broader account that sponsors wanted. It also meant giving up a convenient equivalence between 540 categories present and the intended evaluation complete. The outcome remained a substantial collection with a clearly stated gap and a bounded next step. The lesson was to describe coverage in terms of the question the test is intended to answer. A CDO should be able to value the work already done while preventing its scale from carrying a stronger claim than the evidence permits. The public example offers that reasoning without publishing a real client's conditions or asserting a universal synthetic-data standard.
Others say
David accepted that the overall accuracy figure did not answer the severe-incident question. He withdrew his own readiness claim instead of asking the qualification to do the work quietly.
I wanted the library to move faster. David kept the distinction between a promising pattern and an approved use, then showed the actual subset that reached client projects.
Point of view
An invitation
I would consider a discussion about how a services business learns across clients while retaining the conditions of its evidence. A prospective first-board role would require employer permission, conflict review, preparation and realistic availability. This fictional profile provides no data-processing or consulting service. A real introduction should describe the proposed decision without sending client extracts, contracts or model artefacts through a public route; the appropriate owners should determine what material can be shared.
office@davidkeldraven.example