Chief Data Officer / India
Correct the inference before it becomes a commitment
- Company employees
- 217
- March 2026; fourteen data colleagues through two leads
- Company revenue
- ₹102 crore
- FY 2025–26;₹8.2 crore operating profit
- Analytical experience
- 30 years
- Statistical reporting through industrial data leadership
Decisions & consequences
01Correcting the order amount repeated by a shipment join
In 2018 I approved a management extract showing ₹18 crore against a defined group of 240 orders. The order headers actually totalled ₹12 crore. Eighty of those orders, together worth ₹6 crore, had two shipment records each; the remaining 160 orders, together worth ₹6 crore, had one each. After the query attached shipment detail, there were 320 rows and the header amounts were summed at that expanded level.
The two shipment rows were not necessarily duplicate data. They represented distinct shipments against the same order. Removing one would have damaged the detail while concealing the reporting mistake. The calculation needed to preserve the order-level amount and describe shipment information at its own level. I had reviewed the output rather than tested that relationship, so I accepted responsibility when commercial and finance colleagues challenged the total.
We withdrew the affected management extract and funded a six-week,₹21 lakh correction and review. The programme separated the measures, reconciled the 240-order population and examined other reports using the same relationship. Finance retained authority over financial reporting. The case did not establish duplicate customer invoices, an accounting restatement or a ₹6 crore cash loss; the error was in the specified management analysis.
Some colleagues suggested retaining the larger figure with a note describing shipment activity. I rejected that because the heading still presented an order amount. A footnote would not make the measure answer the question the paper claimed to address. Management needed to revisit any recommendation based on that amount, using the corrected population rather than an explanation that left the original impression intact.
The outcome was a ₹12 crore order account with shipment detail preserved separately. We did not call the correction a saving or claim that demand had fallen by ₹6 crore. I learned to ask what happens to a measure when one business record legitimately matches several others. That question belongs in the review before an attractive total reaches the executive table, especially when the query is technically valid and therefore unlikely to fail in an obvious way.
Holding a delivery-risk prototype that could see information from later in the process
A 2023 predeployment review examined 120 candidate fields for a delivery-risk prototype intended to assist an order decision. Thirteen were populated only after the proposed prediction point. In the historical extract they looked like ordinary available information, but they carried knowledge that the system would not possess when asked to assess a new order.
The team had produced an encouraging evaluation, and colleagues wanted to begin using the output in planning. I asked for the time at which each field became available, not merely the date of the business event it described. That distinction exposed the problem. A field about an earlier activity could still have been entered later, and a final status could reveal part of the outcome the model was supposed to anticipate.
We held deployment and approved ₹54 lakh over ten weeks to reconstruct the data as it would have been known at the intended decision point and to revise the evaluation. The thirteen fields were excluded from that use. The other 107 were retained for further assessment rather than declared valid by subtraction. Operations continued its existing decision process while data and responsible reviewers examined the revised evidence.
The programme did not claim a specific production improvement or a ready-to-release model at the review point. Removing unavailable information can change apparent performance, but a comparison is only useful when its populations, methods and decision purpose are understood. We did not attribute every difference to one technical change or substitute a new attractive score for the missing release judgement.
Mohan Pelqeshi, the operations head, wanted practical assistance rather than an indefinite research exercise. I agreed that the work needed a decision and an accountable next review. I did not agree that a schedule justified testing a future decision with facts from its later history. The case taught me to make prediction time a business definition shared with operators. A historical fact can be true and still be an invalid input for the moment at which management wants the model to help.
Keeping regional customer relationships while correcting the combined count
In 2025 two regional reports showed 120 and 90 active legal-customer identifiers for the same defined period. A combined management summary added them and reported 210 customers. Thirty identifiers appeared in both regions, so the enterprise population contained 180 distinct legal customers. The sum described regional memberships, not the distinct population named in the summary.
Commercial colleagues were concerned that a correction would remove legitimate work from their regional accounts. A customer served in both regions was relevant to both teams. I agreed that those relationships should remain visible. The mistake was not that either region had counted a customer it served; it was that the combined report changed the question without changing the aggregation.
We funded a five-week,₹16 lakh reconciliation and reporting change. The revised account kept 120 and 90 regional memberships, identified 30 shared customers and showed 180 distinct identifiers at enterprise level. We defined activity for the same period and retained the agreed legal-customer identifier. We did not merge separate legal entities merely because they belonged to one corporate group or shared part of a name.
The correction therefore did not establish a fall in the customer base, a loss of thirty accounts or a change in revenue. It changed the interpretation of the combined count. A separate concentration question could require a corporate-group view, and a service question might require operating sites. Those relationships needed their own definitions rather than being silently substituted for the legal-customer population.
The useful outcome was a report in which regional managers could recognise their work and the executive team could understand the enterprise total. I learned that deduplication is often an incomplete description of the task. The business may need every relationship that creates the overlap. The data leader must preserve those relationships while preventing their count from becoming a claim about a different population. That is a judgement about meaning before it is a calculation about rows.