Preserve the evidence behind the convincing picture

In 2020, our catalogue-editing demonstration showed a straightforward improvement: an editor could crop a sample image and see the result immediately. I had endorsed the acceptance review as engineering director. A later rehearsal used sixty synthetic catalogue files and found that twenty-four working source files were overwritten after the crop operation. The visible output met the requested crop, but another editor could no longer start again from the same working original. Separate pristine test copies remained available, and the exercise involved no customer photographs.

The tempting response was to describe the result as a training problem and tell editors to make their own copy first. I rejected that as the principal answer because the proposed workflow itself presented the operation as an ordinary supported action. I held the affected release and recommended ₹48 lakh over twelve weeks to separate original assets from their derivatives, record the relationship and test restoration through the approved lifecycle. Product and security colleagues participated in decisions about access, retention and removal; keeping originals was not permission to retain everything without a purpose.

The revised acceptance set covered seventy-two defined transformation and recovery tests. At the review point, sixty-one met the stated conditions and eleven remained held. We did not describe those numbers as files recovered from customers or as a measure of all possible editing behaviour. The affected paths could proceed only after their required acceptance conditions were met. Engineers retained implementation responsibility, and the product owner retained responsibility for the editing proposition offered to colleagues.

The delay displaced other catalogue work and required me to explain why my earlier approval had been incomplete. The lesson was not that cropping is inherently risky or that preserving an original resolves every content problem. It was that a successful transformation needed to be assessed alongside the next user’s ability to understand and revise it. The release account now recorded what the operation preserved, what it produced and which lifecycle actions remained unproven. That gave the owner a more accurate choice than the original demonstration had provided.

After becoming CTO in 2023, I reviewed a product-comparison export in which the team had removed embedded colour-profile information while preparing lighter files. In a controlled review of sixteen sample images, six showed a material appearance difference under the examined conditions. The result did not establish that every customer would see the same difference, but it showed that our export account had treated file size as the main success measure without adequately examining interpretation of the colour values.

Qualified imaging reviewers helped distinguish preserving a profile, making a justified colour-space conversion and merely assigning a different profile. We did not let a developer choose a new label simply because it made one screen look closer to the expected image. The product team also clarified what the comparison was intended to demonstrate. A remote display cannot support every claim that a controlled in-person demonstration might support, and the wording needed to reflect that boundary.

I obtained approval for ₹63 lakh over fourteen weeks to revise the export pipeline and its acceptance evidence. The review covered twenty-two specified export-and-viewing combinations. Eighteen met the defined conditions and four remained held for further work. Those were combinations, not customer counts or proof of exact reproduction across all devices. The unresolved four stayed in the release account, and the affected comparisons were withheld rather than represented as universally accurate. The programme cost was recorded as a development commitment, not an estimate of sales saved.

The alternative was to keep the smaller files and add a broad warning that screens differ. I did not consider that sufficient when our own processing had introduced a change we could investigate and address. The decision cost time and required cooperation between engineering, product specialists and commercial owners. It made the comparison narrower and more defensible. I remained accountable for the engineering recommendation while the specialists retained judgement about image interpretation and the commercial team retained responsibility for the claim made to customers.

A 2025 experiment proposed conversational answers to camera-product comparison questions. In a defined review of three hundred generated answers, product specialists identified forty-two with unsupported statements. Twenty-three confused product generations and nineteen lacked a supporting reference for the feature asserted; these were separate groups within the forty-two. The review described answers in an experimental set, not three hundred customers or the measured frequency of errors in an already operating service.

The demonstration was persuasive because the answers were readable and appeared helpful. Commercial colleagues wanted to continue towards a public launch while adding a caution beside the response. I held customer release of the experimental assistant and approved ₹52 lakh over ten weeks to narrow its scope, use curated structured specifications for defined comparisons and require a claim-level support check. Product specialists owned the interpretation of source specifications. Engineering owned the retrieval and response behaviour, including the point at which the service should decline to offer an unsupported comparison.

A replay of the same three hundred questions still produced ten answers with unsupported statements. The improvement within that set did not demonstrate acceptable behaviour on unseen questions, and repeating the set did not create an independent evaluation. We retained the ten failures and kept the assistant from customer release. The next decision required further testing and a product judgement about the narrower use case; a lower count alone was not permission to launch. The established catalogue and human product-advice routes continued separately.

I did not claim an invented return on the experiment or use a generic caution as a substitute for evidence. The commitment produced a clearer account of where generated language was unsuitable for the promise we wanted to make. That was a useful engineering result even though it was not a launch. It also required me to defend a stop decision after spending money, rather than letting sunk effort turn an unsupported statement into an acceptable customer answer. Future work would need fresh evaluation and explicit ownership of the remaining uncertainty.