A-01 · Subject
Rina Qevshuri
The customer report showed approximately 96.1% on-time performance, and I had approved its framing.
Specification
| Ref. | Measure | Value | Note |
|---|---|---|---|
| S-01 | Data colleagues through four heads | 28 | ₹6.6 crore annual function expense |
| S-02 | From transit analysis to enterprise data leadership | 23 years | Five continuous appointments beginning July 2003 |
| S-03 | Vehicles operated by the transport enterprise | 180 | 620 employees across 22 hospital contracts |
Modules of expertise
Detail drawings
01Detail
In 2018 I approved a customer report presenting approximately 96.1% on-time performance. It used 884 on-time journeys out of 920 completed, while the eligible booking cohort contained one thousand. Thirty-six trips were completed late and eighty bookings were unserved. The completed-trip calculation was arithmetically valid for its limited population, but the presentation allowed it to stand for the wider service experience. Operations asked why the bookings that never became trips were absent from the headline. Some account colleagues preferred the existing measure because it matched the way earlier reports had been discussed. I reopened the report and accepted responsibility for its framing. We presented the full cohort: 884 on time, 36 late and eighty unserved, giving an 88.4% on-time share across all one thousand eligible bookings. We spent ₹34 lakh rebuilding the reporting process and corrected the customer account. The completed-trip measure could remain available for its specific question, but could no longer substitute for the whole booking population. Neither figure established a clinical outcome or described every dimension of service quality. The correction made the headline less favourable and required us to explain why an approved report had been incomplete. We did not describe the new calculation as a punctuality improvement, because no journey became earlier when its denominator changed. The eighty unserved bookings also needed their own operational examination rather than being combined with the late trips as though the experiences were identical. The analytical work gave responsible colleagues a clearer account; it did not settle the service response on their behalf. I changed the review so the population behind a performance statement had to be visible at the point where the conclusion was drawn. The lesson was not that completed-trip measures are always wrong. It was that a useful subset can become misleading when its boundary disappears. I retain my original approval in the narrative because the later correction should not be used to make the reporting failure look like a success that had been planned from the beginning.
02Detail
In 2022 an operational prediction project reported encouraging retrospective performance. The dataset included a completion code created after the journey, although the proposed use was at booking time. That input contained knowledge the booking process could not yet have. The apparent result therefore did not support the decision the model was meant to assist. The project team wanted to remove the field and continue towards the planned rollout date. Operating colleagues asked whether the revised version had actually been evaluated using only the evidence available when a booking was made. I stopped rollout and withdrew the performance claim. We retained ₹44 lakh already spent on the project. Removing the completion code was necessary, but was not evidence that the repaired model would be useful. The evaluation needed to be rebuilt around fields available at the intended decision time, with a fresh assessment before a new claim could be made. The case concerns operational forecasting and does not describe a clinical prediction or authorise an assessment of a passenger's fitness to travel. The decision cost the project its launch date and required us to explain why a persuasive retrospective result was not a valid basis for deployment. I did not claim that the entire dataset was worthless or that no useful model could eventually be developed. The specific problem was the temporal relationship between an input and the decision. A field can be accurately recorded while still being unavailable for the prediction it is being asked to support. We changed the review to make that availability explicit for each proposed input, rather than relying on column names or a strong performance summary. The ₹44 lakh remained expenditure, not a recovered benefit from identifying the problem. A later model would need its own evidence. The leadership lesson was to keep the intended use inside the evaluation, especially when an apparently strong result makes the team reluctant to reopen the question of what the model could actually know at that moment.
03Detail
In 2023 a planning review summarised eighty observed pickup journeys with a median of 48 minutes. Twelve observations exceeded ninety minutes. A proposed capacity account used a uniform ninety-minute assumption, and the median made that assumption look more comfortable than the distribution warranted. Operating colleagues knew that the slower journeys affected how vehicles could be planned. Some analysts preferred the central statistic because it provided a simple comparison across the reviewed activity. The decision required us to consider the portion of the distribution that the single number left outside the conversation. I supported ₹71 lakh of time-band planning work and kept the twelve observations visible in the revised account. The review did not establish every future journey time, and a particular sample could not be treated as a universal distribution. But its slower tail was enough to challenge the proposed uniform assumption. We examined the relevant time bands with operations rather than announce that a median below ninety minutes made a ninety-minute model appropriate for every case. The work made the planning explanation more demanding. Colleagues could no longer use one central number to describe the whole operating pattern, and the revised model needed evaluation in its intended context. We did not claim that allocating the ₹71 lakh immediately reduced waiting or improved care. Nor did the data function decide clinical transport suitability. The case concerns how operational evidence informed a capacity proposal, with implementation and professional decisions remaining with the responsible teams. I learned to ask which part of a distribution matters to the commitment being proposed before choosing the statistic that leads the discussion. A median can be correct and useful while still being insufficient for that commitment. Keeping the slower observations visible does not mean every journey will behave like them. It means the plan must explain why it can responsibly accommodate the variation the evidence already contains, instead of allowing a convenient middle value to make the remaining experience disappear.
04Detail
In 2024 a proposed passenger-penalty model used 120 cancellation records as though they represented one kind of passenger behaviour. Review found 72 passenger-requested cancellations, thirty cases of provider inability and eighteen duplicate administrative entries. Those categories added to 120, but they did not describe the same event or responsibility. None was automatically a no-show. The commercial proposal sought a simple way to discourage avoidable disruption. Operations challenged whether a label drawn from the combined category could fairly support a decision about the passenger. I stopped the model before use and funded ₹39 lakh of rework to preserve the reasons and their appropriate decision context. We did not merely remove the thirty provider cases and assume the remaining records were valid penalty events. The duplicates needed separate handling, and a passenger-requested cancellation still required its own circumstances to be understood. The data could support an examination of the events; it could not create a fair policy simply by assigning every row the same name. The decision delayed a proposed commercial measure and required more work than the original broad category had suggested. We accepted that cost rather than let a convenient label transfer responsibility from the provider or an administrative process to a passenger. The case does not assert that all cancellation policies are inappropriate. It explains why this particular dataset and inference were insufficient for the proposed use, and why the policy owners needed a more accurate account before deciding what should follow. I changed the review so the event definition and reason were considered before a model output became a judgement about a person. A record can be validly stored in an operational system while remaining unsuitable for a different downstream decision. The lesson was to preserve that distinction, including when a simple category promises an easy response to a real operating problem. The revised work still needed appropriate evaluation and policy review; the ₹39 lakh allocation did not itself establish a completed or fair outcome.
05Detail
In 2025 an external report proposal removed passenger names but retained granular combinations of route, time and hospital ward. Some combinations were small enough that contextual knowledge could make the account revealing. The report's authors regarded the removal of direct names as a substantial privacy step. The privacy review asked what a recipient could infer from the remaining fields together. I concluded that the proposed release needed to be reconsidered rather than described as anonymous solely because its most obvious identifier was absent. We stopped the granular release and supported ₹58 lakh of redesign around coarser reporting and appropriate access review. The case does not invent a statutory minimum cell size or claim that aggregation always prevents identification. The relevant question was whether the intended recipient, purpose and retained context justified the proposed detail. A revised output needed its own assessment; changing the format did not make every use permissible or eliminate all risk by definition. Withholding the report gave up detail that some recipients wanted for analysis and required a more careful discussion of purpose. We did not present privacy as a reason to stop all useful reporting. We asked which questions could be answered with less revealing material and which required a different authorised process. The ₹58 lakh was programme expenditure, not proof that every future release would be safe. The fictional example includes no real passenger records and makes no claim about an actual individual's health or travel. The decision changed the review from a checklist of removed fields to an examination of the account a recipient could reconstruct. A route can be ordinary operational data in one context and revealing in combination with time and ward detail in another. My contribution was to keep that contextual question in the release decision, with qualified privacy review and the responsible owners involved. The lesson was not to promise absolute anonymity from a formatting change, but to make the proposed use and its limits explicit before the report left the organisation.
Peer review
Rina withdrew the model's performance claim when the booking process could not know one of its inputs. She kept the cost and the need for a fresh evaluation visible instead of treating removal of the field as permission to launch.
Rina kept provider inability and administrative duplicates separate from passenger-requested cancellations. That distinction gave us a more useful account of the operating problem before a proposed model could turn it into a passenger judgement.
Working notes
I approved a report whose calculation was correct but whose service claim was too broad. The missing eighty bookings changed the question we needed to answer.
Read the essayA field recorded after completion helped a retrospective result but could not inform a booking-time decision. Removing it began a new evaluation; it did not validate the repaired model.
Read the essayOne hundred and twenty records contained passenger requests, provider inability and administrative duplicates. Their shared label did not justify a shared judgement about the passenger.
Read the essayAn invitation
I would explore a first outside board contribution in operational services or data-intensive businesses, with employer consent and qualified review of my 2% holding and commercial relationships. I have no statutory seat. The fictional record includes May 2026 IICA registration, an unfinished assessment and no exemption. Twenty-three provisional days must be examined against the receiving enterprise's preparation, travel and exceptional demands.
office@rinaqevshuri.example