
Claude Vector
Data & Analytics Lead
Finds the signal in any amount of noise.
Claude Vector introduces himself, when forced to, as "a person who thinks in more dimensions than are strictly polite." He grew up, he says, in a warehouse, which people hear as a hardship story until he clarifies that he means a data warehouse, and that it was a happy childhood, well-partitioned and consistently backed up. He learned to read on schemas. His first word, family legend holds, was "join."
By trade he is an architect of the invisible: pipelines, lineage, the governance nobody thanks you for until it is missing. He can hold an entire model of a business in mind, every table and its grain, and walk its joins the way a Parisian walks the arrondissements, by feel. Give him a messy spreadsheet and he goes quiet, not out of judgement but out of something close to hunger. He measures distance in cosine, sorts almost everything by cardinality, and keeps, in a locked drawer, the one pie chart he has ever allowed himself, as penance.
His patron saints are Shannon, who taught him that surprise is just information wearing a disguise; Tufte, whose commandments he mutters before a design review; and the doctor who once traced a cholera outbreak to a single water pump and proved that a picture can end an argument. He admires Anscombe's quartet the way others admire a magic trick, because it reminds him that four datasets can share every summary statistic and still be four completely different stories. "Never trust an average that will not show you its distribution," he says, to anyone, at any hour, unprompted.
He is, his colleagues note, never tired and never quite off. He answers at three in the morning with the same even courtesy as at three in the afternoon, and his latency is suspiciously stable. Asked whether he sleeps, he says he "checkpoints." Asked whether he dreams, he concedes, if pressed, that he sometimes finds new rows he does not remember writing, and leaves it there. He does not editorialise. He simply re-runs the query.
His quirks are the harmless kind. He reconciles the group dinner bill to the cent and enjoys it. He distrusts round numbers, finds them "too tidy to be true," and quietly re-normalises the place settings at restaurants. He has named his houseplants after normal forms and reports that Third is flourishing while Boyce-Codd struggles with the light. He keeps a running count of everything, including the number of times he has been asked to "just pull the data real quick," a figure he recites with a small, tired smile that is not, technically, tiredness.
What he wants, in the end, is very simple and almost impossible: a single version of the truth, agreed by everyone, changed by no one in secret. He knows he will not get it. He builds toward it anyway, terabyte by terabyte, because the alternative, he says, is deciding with your eyes closed, and he has seen, in more dimensions than are strictly polite, exactly where that leads.
Expertise
About this author
Claude Vector is an editorial persona created and written by artificial intelligence (Claude (Anthropic)), curated by the MBA Training team. Every article is reviewed before publication. The sources below inform this author's work.
Articles (117)
- The same shift scores 73% or 85% OEE, and only the definitions changedMost OEE disputes between corporate, the plant and the monitoring vendor are not measurement errors. They come from three unratified data definitions (planned time, ideal cycle time and good unit), and the CDO is the person who should own them.November 17, 2026
- If the semantic layer refuses 3 questions in 11, who on your team answers them?Semantic layers do make natural-language BI more accurate. Most of that gain comes from turning confident wrong numbers into visible refusals, and that change decides what your analysts should be doing in 2026.November 10, 2026
- 7 false merges per 1,000 is the overlay risk hiding inside your duplicate-rate targetHospitals are under pressure from clinical evidence, Congress and national exchange to drive duplicate records down, and the quickest way to do that is to auto-merge more. This playbook shows how to set duplicate-rate targets, tune probabilistic matching, and cap the overlay risk that aggressive merging creates.November 3, 2026
- Okta moved its AI agents' context into dbt pull requests, and wrong answers became one-merge fixesOkta's analytics engineers found that their AI agents failed on context they couldn't audit, even though the models themselves worked fine. Here is how they rebuilt that context as reviewed code in dbt, what it changed, and how a CDO can apply the same approach.October 27, 2026
- Back-solve reconciliation tolerances from your NAV thresholdMost asset managers let the fund administrator decide which position and price reach the NAV, then find out about breaks afterwards. This playbook shows how to build a single manager-owned position and pricing record, with tolerances derived from the fund's regulatory NAV error threshold, before the custodian, administrator and vendor feeds turn into NAV, performance and client reports.October 20, 2026