The question I should have asked before the score

The first question I should have asked was when the model could know its strongest input. Instead, I asked whether its score was high enough. Our retrospective renewal test correctly classified 182 of 200 contracts, and I approved the 91% result for a management discussion. The model was intended to predict thirty days before a rental ended. A renewal invoice used as an input had been created after that point. It was not an early signal. It was part of the later outcome leaking into an earlier decision.

I grew up in Ludhiana in a Sikh family and completed statistics at the fictional Tevnuli College of Statistics in June 1997. The following month I joined Qemzari Survey Services as a research analyst. The enterprise had forty-four employees and ₹6.8 crore revenue for FY ending March 1998. I worked in a seven-colleague research group under its manager, without reporting staff or spending authority. Preparing survey data taught me that a well-filled table could still describe only the people the collection process happened to reach.

In April 2004 I became head of survey analysis. Qemzari employed ninety people and recorded ₹22 crore FY 2005 revenue. Nine colleagues worked through two leads within ₹68 lakh annual expense, and I reported to the head of research. I became responsible for explaining why a comparison supported one conclusion but not another. A response difference could reflect who had been approached, when the question was asked or what had been offered. I learned to investigate those conditions before treating a persuasive chart as the final account.

I joined Qazvori Workcycle Hire in September 2010 and moved to Vadodara. The regional rental enterprise then employed fifty-four people and recorded ₹9 crore FY 2011 revenue. Five analytics colleagues worked through one lead within ₹42 lakh annual expense, and I reported to the head of commercial services. We rented manually pedalled cargo cycles to adult business users. The move brought research into a company that had to decide where to place equipment, how to communicate with customers and which account of demand should guide its next commitment.

In June 2016 I became head of decision data. The business employed 112 people and recorded ₹24 crore FY 2017 revenue. Eleven colleagues worked through two leads within ₹1.3 crore annual expense, and I reported to the CEO. The remit connected data preparation, analysis and the use of results in management decisions. I came to appreciate how a completed transaction records an interaction between demand and what the business made available. A depot with fewer completed hires might have less interest, less equipment or a different combination of both. The transaction file alone could not decide which explanation applied.

My CDO appointment began in February 2021. Today eighteen colleagues work through three leads with ₹2.8 crore FY 2026–27 operating expense and a separate ₹1.1 crore capital programme. Qazvori has 178 employees, nine regional depots and ₹48 crore FY 26 revenue. Founders hold 46%, a regional fund 39% and an employee vehicle 15%; I hold no personal or vehicle interest. Mira Qevnuli is CEO and Harish Tezmori leads operations. Customers are not our employees, and the company is neither a courier nor a public-transit authority. Workshop and operating owners decide equipment condition, suitability, maintenance and safety.

The renewal-model correction cost ₹41 lakh. We withdrew the 91% claim and rebuilt around inputs available at the intended decision date. A different prospective shadow cohort produced 154 correct classifications from 200 contracts, or 77%. I do not describe that as a measured fourteen-point deterioration: the cohorts and evaluation conditions differed. The useful change was that the new account could be examined in relation to the decision it was meant to support. Baselines, error costs and practical value remained questions to resolve before deployment.

A subsequent depot comparison recorded 120 and 80 completed hires across the same twenty-eight days. The first depot had cycles available throughout; the second on only sixteen days. We commissioned ₹32 lakh of observation work and deferred an expansion conclusion. Dividing the second count by sixteen and multiplying by twenty-eight would not establish demand on days when no cycle was available. The apparently simple comparison needed evidence about requests and availability rather than a confident extrapolation.

In 2024 we ran an eighteen-lakh reminder trial with 400 eligible business accounts randomly assigned between an offer and the usual process. Forty-two of 200 offered accounts renewed within eight weeks, compared with thirty-four of 200 in the usual-process group. The observed difference was four percentage points. We chose another planned evaluation before a full rollout, retaining uncertainty and costs instead of declaring a general commercial effect. A later reporting review also showed why removing customer names did not settle the risk of revealing activity in small depot-week groups.

I speak Punjabi, Hindi and English, with working Gujarati. I have not registered with the IICA databank, have not completed the assessment and claim no exemption. With no statutory appointment, I provisionally reserve twenty-four days for a future first board contribution. I would bring questions about evidence and decisions, while learning the proposed enterprise's wider financial and operating responsibilities. My own premature approval belongs in that account because it shows the point at which an attractive result can make a data leader stop asking the question that matters most.

Open to Board appointments · Advisory roles

Languages / Punjabi · Hindi · English · Gujarati — working proficiency