Chief Data Officer / India
RupinderSahlvethi
I grew up in a Sikh family in Patiala and studied at Tavqumi College of Statistics, completing a Bachelor of Science in June 1996 and a Master of Science in Applied Statistics in June 1998.
42
Data colleagues through four heads
₹12.4 crore FY 2026–27 operating expense; separate ₹3.1 crore capital programme
28 years
From study records to data-methods leadership
Five continuous appointments beginning July 1998
48
Client organisations served by the company
215 employees; ₹83 crore FY 2026 company revenue
I believe
I want our software to carry the choices that made a research observation possible. A blend changes within a total. A panel evaluates products in an order. A laboratory can report a bound instead of an exact value. Those conditions should be available when a customer interprets the result, including when they make an attractive recommendation less certain. My responsibility is to fund and organise the data work that preserves them, while recognising where scientific judgement must lead.
— Rupinder Sahlvethi
The work
In 2019, while I was director of data methods, we prepared a demonstration from four experimental blends. Component A took values of 6%, 7%, 8% and 9%. Component B moved in the opposite direction, from 64% through 63% and 62% to 61%, while C remained at 30%. Each blend therefore retained the specified 100% total. I approved an explanation that attributed the observed response to increasing A. The observations did not separate that change from the simultaneous reduction in B. Peter Qezvomi challenged the account before the proposed broader demonstration. The commercial team wanted to retain a simple ingredient story because prospective users could understand it quickly. I initially thought a short qualification would be sufficient. Peter asked what comparison would establish a separately supported conclusion we were inviting the audience to draw. I could not point to one in those four blends. The issue was not whether the recorded responses were genuine within the invented study; it was the explanation we had attached to them. I withdrew the standalone ingredient-effect claim and supported ₹32 lakh for review and a revised experimental design. Scientific colleagues defined the study region and response questions rather than having the software choose a recipe. We retained the original blend compositions and observations as a record, with their limitations visible. We also revised the demonstration so a reader could see every component movement rather than only the one selected for the headline. The funded work did not turn the original comparison into a causal result after the event. The decision delayed the account we wanted to present and required colleagues to reopen work they had considered finished. I owned the approval instead of describing it as an import defect. The lasting change was a product requirement that the explanatory view display the composition as a whole and the study question alongside the result. This case concerns an invented research-software demonstration, not a formulation instruction or a conclusion about food safety. It taught me to ask whether the experiment had distinguished the explanation before allowing an attractive chart to make it sound settled.
A 2022 sensory-study demonstration contained 72 ratings: twelve panellists assessed each of six products. Every panellist encountered those products in the same order. A proposed comparison showed the products' ratings as though the arrangement allowed a straightforward account of product differences. Because product and position were tied together throughout the exercise, the records could not separately identify their contributions. More ratings within that arrangement did not resolve the confounding. The product team argued that the table was complete and that a ranking was the feature customers expected to see. I understood the cost of postponing a demonstration after the observations had been collected. Peter Qezvomi asked us to consider whether early and late positions might involve different conditions for a panellist. We did not need to assert that order had caused the particular scores to recognise that the design had left the explanations inseparable. The useful decision was about what comparison to release, not an attempt to guess a corrected ranking from the same records. I supported a seven-week delay and ₹24 lakh for a revised study. Its schedule placed each of the six products in each presentation position twice across the twelve panellists. That provided position balance for the stated exercise. It did not establish that every carry-over effect, participant difference or other influence had been removed, and we did not represent the panel as a sample proving wider market preference. Research specialists retained responsibility for the study design and interpretation. Our software team made sequence visible in the records and required the planned arrangement to be reviewed before the comparison was prepared. The additional 72 ratings were collected under the revised schedule, while the original records remained identifiable. We withheld the earlier ranking rather than merging both exercises into a larger, more impressive total. The result was a better-defined basis for the research team's analysis and a delayed demonstration with an explicit additional cost. For me, the CDO responsibility was to connect the design weakness with funding, product behaviour and the publication decision, without pretending that software could calculate away an experimental choice the team had not made.
In 2023 we reviewed an imported set of 150 laboratory results used in a fictional formulation-research demonstration. The laboratory had quantified 110 results and reported the other forty as below a quantification limit of 0.2 mg/kg. Our import represented those forty entries as numeric zero. It retained a unit but lost the distinction between a measured value and a qualified statement about what the method could quantify. A summary built from the imported numbers therefore contained an assumption the source report had not supplied. Engineering colleagues preferred a single numeric field because it simplified charts and downstream calculations. Some users also wanted an average immediately, without dealing with different result types. I asked the team to show where the laboratory had measured the zeroes. There was no such observation in the source. Peter Qezvomi supported restoring the qualifier, but also warned against installing a different automatic replacement and treating it as universally correct. The appropriate statistical handling depends on the analytical question and method; that judgement could not be settled once for every customer through an import default. We funded ₹29 lakh for import changes and review over twelve weeks. The revised representation kept the relation, limit and unit together, alongside the original report reference. The forty entries became qualified values again; the 110 quantified results retained their measured values. Where a calculation required a treatment of the qualified observations, the product made that choice explicit for the relevant specialist rather than silently completing the column. We revised affected demonstration summaries and explained why the earlier output could not be carried forward unchanged. This did not prove that the substance was absent, that a product was safe or that one replacement number would always be appropriate. It also made parts of the interface less immediately convenient. I accepted that cost because a clean numeric field was not a sufficient reason to invent precision. The completed import repair gave researchers a more faithful starting point. Their subsequent interpretation remained a separate task, and our commercial account retained the ₹29 lakh expense rather than presenting the correction as an unmeasured saving or a general certification of data quality.
A 2024 product proposal would have treated ingredient-lot quantities as interchangeable when the mass entered in the software was the same. Our review compared two invented 100 kg lots whose measured moisture fractions were 12% and 15% on an as-received mass basis. The corresponding dry masses were 88 kg and 85 kg. The difference was three kilograms of dry material even though both source entries said 100 kg. That arithmetic exposed a missing basis in the proposed feature; it did not prescribe how a formulation should be changed. The product lead wanted an automatic adjustment because customers found manual reconciliation cumbersome. I agreed that the repeated work deserved attention. I disagreed that a mass-basis calculation alone authorised the product to substitute a quantity in a recipe. A client scientist might need to consider the full formulation, process and other material properties. Peter Qezvomi asked us to preserve that professional decision instead of embedding an implied recommendation in a helpful-looking conversion. The choice was between automating a calculation within an explicit context and automating a conclusion that required more than the calculation. We deferred a ₹1.9 crore feature expansion and separately approved ₹41 lakh for basis-aware data work. The revised records carried the as-received quantity, measured moisture basis and derived dry mass without silently overwriting the source amount. Research users could inspect the relationship and make a documented choice through their own process. The product did not describe 100 kg from the second lot as equivalent to 100 kg from the first, nor did it prescribe adding material to achieve a desired dry mass. The wider expansion remained deferred. The narrower work was completed for the demonstration workflow, making the discrepancy visible where the quantities were used. We did not subtract the two programme amounts and call the difference a saving: they represented different scopes and approval states. I wanted the board account to retain both the unmet request for greater automation and the improvement we had actually delivered. The lesson concerned how a data platform represents material assumptions, not a universal rule for food manufacture, a recipe instruction or a claim that moisture alone determines product performance.
In 2025 a proposed module combined observations collected over twenty-one days under one specified storage condition. The commercial proposition described a shelf-life answer, and its annual licence-revenue forecast was ₹4.6 crore. The available observations could support examination of a bounded trend in that demonstration. They did not, on their own, establish the broader duration promise implied by the proposed product. Other conditions, endpoints and professional assessment had not been addressed by the software account. The sales and product teams wanted to preserve the higher-value proposition because customers were asking a practical question about duration. I did not dispute that demand. I asked what evidence the feature would need to answer it responsibly and which party would hold the necessary scientific responsibility. Peter Qezvomi would not endorse the general claim from the limited study. CEO Anita Qevrasi asked us to present the commercial consequences of narrowing the feature instead of leaving the caveat to implementation after an attractive investment case had been approved. We withdrew the general shelf-life proposition and retained a trend-view workbench whose stated scope matched the observations. The revised annual licence-revenue forecast was ₹1.8 crore, still a forecast rather than contracted or realised revenue. We preserved ₹37 lakh already spent on discovery and approved a separate ₹1.2 crore redesign programme, which remains unfinished in this fictional account. No safety certification or universal duration was attached to the workbench. Its purpose was to help qualified researchers inspect the recorded conditions and observations, not to replace their assessment with a date displayed by software. The decision gave up part of the proposed commercial story and required a different sales explanation. The ₹2.8 crore reduction between forecasts was not money already lost, while the discovery expense could not be erased by changing the name of the feature. My role was to connect the evidential boundary with the investment paper and product requirements. The CEO and relevant approving bodies retained their authority. The result is deliberately incomplete: a smaller, more defensible proposition with funded work still to do, rather than a claim that caution inevitably produces a superior financial outcome.
Others say
Rupinder connected the limits of the proposed shelf-life feature with its forecast and funding request. She kept the smaller proposition, discovery expense and unfinished redesign visible in the same account.
I challenged the interpretation of the four blends. Rupinder withdrew the separate ingredient-effect claim and funded further work without treating my scientific role as a substitute for her own approval responsibility.
An invitation
I would consider a first outside board contribution where research evidence and software promises shape investment. This invented career records April 2026 IICA registration and an August assessment pass without an exemption. I hold no statutory directorship. Twenty-four provisional days, employer consent and a review of relationships would need to support any particular appointment; neither a data title nor an assessment result settles that fit.
office@rupindersahlvethi.example