The journey so far
Safiya Zevqenri
Chief Technology Officer · India
The screen said centimetres. The receiving component still assumed metres.
72
Engineering employees
Six leaders; twelve contractor positions counted separately
₹13.8 crore
Annual engineering expense
₹5.2 crore approved development is a separate authority
23 years
Software career
Commerce developer from July 2003; CTO from January 2022
The compass
An interface does more than transmit a number or display a choice. It carries assumptions about units, valid combinations, usable actions and the state that another version must understand. I ask engineering to make those assumptions explicit, especially when an attractive demonstration can hide the customer path that exposes them.
Expeditions
01
In 2021, as engineering director, I approved an interface change that expressed curtain width in centimetres. A receiving component still interpreted the bare value as metres. The synthetic example was simple: 2.35 metres became 235 centimetres, then the number 235 was read as 235 metres. Eighteen of 120 test quotations exposed the mismatch. The arithmetic was correct at each local step only under assumptions that the two components did not share. Product colleagues wanted the interface change to remain within its planned release because the visible input was clearer to customers. I agreed with the intended usability improvement, but the receiving behaviour made that release unacceptable in its current form. We held it and allocated ₹41 lakh to the unit contract and its tests. My earlier approval remained part of the account; I had examined the displayed change more closely than the meaning another component would assign to it. The correction required engineering teams to describe the value with its unit and receiving expectation, then test the relevant boundaries. Technical-design head Yusuf Tevqeri retained physical-product feasibility. The software team could not certify a customer's actual measurement or decide every manufacturing constraint through a numeric conversion. This fictional case contains no live wrong order, no pricing-policy change and no assertion that every other interface was already sound. The result was a held release and a funded repair to the contract between components. I would not reduce the lesson to asking developers to multiply or divide more carefully. The important failure was that a plausible number travelled without enough explicit meaning for the receiver to interpret it correctly. A CTO needs to ask where that meaning is established and how an apparently local improvement will be understood by the rest of the customer journey.
02
A 2023 review of the furnishing configurator tested 360 scripted sequences rather than only the ordinary forward path. Forty-seven left an incompatible lining attached after the fabric choice changed. The interface could display recognised values in each field while their combination no longer matched the technical-design rules. A customer undoing or replacing an earlier choice was taking a legitimate path that the initial acceptance account had not represented adequately. The product sponsor wanted to preserve the launch date and suggested warning customers to check their final choices. Technical-design colleagues argued that the software should not present an invalid combination as if it were ready for consideration. I supported holding launch and allocated ₹68 lakh to state-transition validation. A general warning would not resolve the product state or explain which later choices needed reconsideration when an earlier one changed. Engineering worked with the authorised design owner to make those dependencies explicit. The technical team retained physical feasibility, while software had to represent the resulting conditions coherently through change, replacement and undo. We did not infer that forty-seven scripts represented forty-seven actual wrong customer orders, and the programme allocation did not prove every possible path had been tested. The scope remained the observed scripted sequences and the work needed before the proposed launch. The decision gave up the planned date rather than ask a customer warning to carry an unresolved state problem. It also changed the engineering question from whether each field held an allowed value to whether the whole configuration remained valid after a transition. A recognisable option is not sufficient when its relationship to other choices has changed. The case illustrates a CTO's responsibility to connect customer freedom with the rules that make the resulting design intelligible and feasible.
03
In 2024, twenty-four adult participants attempted a defined width-setup task in our configurator. Nine could not complete it where the action depended on dragging a visual control. The task concerned specifying a dimension, not producing a freehand path. I wanted the review to examine whether the underlying action could be performed through a usable keyboard route rather than assume the preferred visual gesture was the product itself. Design colleagues valued the directness of dragging because it made a change visible immediately. I did not ask to remove that option. We funded ₹52 lakh for labelled numeric inputs and keyboard controls alongside the visual interaction. Qualified accessibility review remained necessary, and the work needed to preserve the meaning of the dimension rather than offer an unrelated shortcut. We did not infer a diagnosis or disability from any participant's difficulty. The same twenty-four participants later attempted the task after the change. Three could not complete it, compared with nine before: six fewer observed failures in that group. Practice and order effects were possible, so the comparison was not a controlled causal estimate. It also did not certify the entire site, every assistive technology or complete accessibility compliance. The remaining three failures stayed in the account as further work to understand. The decision expanded the route to a meaningful action while retaining the existing visual option. The outcome recorded here is a funded change and a bounded usability observation, not a universal claim about all customers. My contribution was to keep the function separate from one input technique and to insist that an encouraging review retain its limits. A customer-product team should examine whether people can perform the action without making successful use of its favourite gesture a prerequisite.
04
A 2025 release rehearsal changed saved curtain designs from a single-width representation to a structure capable of holding multiple panels. The deployment plan included returning to the previous application version if necessary. When we tested that reversal against forty-eight synthetic saved designs, seventeen could not be read by the old version after the new structure had been written. Restoring the executable did not restore the assumptions of the data it would encounter. The delivery sponsor regarded the code rollback as a sufficient safety measure and wanted to keep the release date. Engineering and customer-product colleagues needed an account of the saved designs that would exist at the point of reversal. I deferred release and backed a twelve-week ₹1.2 crore compatibility and recovery programme. The decision retained the cost and delay rather than treat the presence of an old application package as proof that the product could return coherently to its previous state. The work needed to examine compatibility, the state written during transition and the recovery conditions through a defined rehearsal. It did not assume that every transformation could be reversed without qualification or that a backup alone settled customer continuity. This fictional case describes synthetic records and no actual customer data loss. The programme is not represented as a completed universal recovery capability; its acceptance still requires evidence appropriate to the proposed transition. The consequence was a postponed release and a more complete recovery question. Code version and stored state had to be considered together. I would not make this a universal instruction to retain every old structure indefinitely. The engineering decision depends on the actual product and transition. The lesson is that a rollback assurance should describe what the returning application will read, not merely the file that operations can redeploy.
05
A 2026 preview-render load test began with 120 logical requests. Under the tested transient-failure conditions, repeated layers produced 1,080 backend attempts, an average of nine attempts per logical request. The customer demand had not become nine times larger. Our handling of retries had amplified the work presented to the rendering service, making a demand-only account of the capacity problem incomplete. One engineering proposal favoured adding compute capacity immediately so the preview could absorb the observed attempt volume. Platform colleagues asked whether that would simply fund avoidable repeated work under the same design. I supported ₹73 lakh of bounded retry and admission-design work before treating the expanded capacity estimate as the only response. The programme needed explicit ownership of retry decisions and a view of the work allowed within a request's useful lifetime. The fictional test does not describe a cyberattack, a live customer outage or an actual supplier failure. It also does not establish that every request can or should be retried in the same way. The appropriate handling depends on the operation and its consequences. We did not promise exactly-once behaviour or infer that one retry pattern would make a service universally resilient. The new design remained subject to its own evidence before production reliance. The outcome was a different engineering investment question: how much of the observed load represented useful customer work, and how much came from our own repeated attempts? Additional capacity might still be justified, but the test showed why it should not be assessed from the amplified total alone. I wanted the programme to preserve the distinction between logical demand and backend activity, so a reassuring capacity plan would not hide the behaviour that made the service struggle under the tested conditions.
Terrain
Languages
Fellow travellers
Safiya explained why returning to an older application would not necessarily restore a usable customer design. That made the release delay a concrete enterprise decision rather than an unexplained technical objection.
She kept the software's valid state connected with our product rules, including when a customer undid an earlier choice. Engineering did not claim that a convincing preview could replace physical feasibility.
Dispatches
A customer-facing improvement exposed an implicit agreement between components. The engineering question was not whether 235 was valid, but what the receiver understood it to mean.
Read the essayThe underlying action was specifying a dimension. A dragging-only interaction made that action unavailable to some participants, and the improvement still needed a bounded account of what the review established.
Read the essayA release plan had a code rollback but an incomplete account of the state the returning version would encounter. A synthetic rehearsal made that gap visible before customer reliance.
Read the essayAn invitation
A first outside non-executive contribution after company-specific learning, relationship review and employer agreement. I hold 1.2% within employee ownership and no statutory seat. Fictional IICA registration and proficiency assessment are complete; twenty-four provisional days remain subject to the actual appointment.
office@safiyazevqenri.example