Chief Technology Officer / India
When the demonstration could not support the offer
- Engineering colleagues
- 28
- Four leads; ₹7.2 crore annual engineering expense
- Engineering experience
- 27 years
- Optical applications to integrated inspection technology
- Executive director
- Since 2009
- Co-founder with 30% ownership; not an independent seat
Decisions & consequences
01Rebuilding the image before adjusting the software judgement
In 2021 we challenged a preproduction carton-inspection system with forty-eight samples containing deliberately introduced missing characters. Eleven received an incorrect accept result under the original lighting arrangement. Reflections from the reviewed finish reduced the visibility of the relevant print area. The demonstration I had previously accepted had used easier material and positioning. The finding concerned controlled equipment tests before delivery; it was not evidence of eleven defective medicine packs reaching patients.
The software lead proposed increasing the confidence threshold and repeating the test. That might change the result distribution, but it would not establish that the missing features were represented clearly enough in the image. Omar Dervqani asked us to separate image formation from recognition behaviour. I accepted that distinction and held delivery. We redesigned the illumination and enclosure around the defined material and presentation range, then tested the optical and software combination together. The commercial team argued that the customer could operate within a narrower material choice while we completed the broader work. I decided we first needed an explicit, supported offer rather than an informal restriction attached to a delivery date.
The ten-week programme cost ₹1.2 crore in engineering, components, trial fixtures and repeat review. We explained the delay and revised conditions to the customer. In the repeated forty-eight-sample challenge, the eleven previously missed defects were detected under the redesigned setup. That result did not establish performance across all finishes, orientations or future artwork. We retained the reviewed material and positioning conditions in the equipment evidence and left customer validation and batch-release authority with the customer's qualified teams.
I had approved the earlier demonstration too broadly, and the correction belonged in my account of the project. We did not describe it simply as the software team improving an algorithm. The decision was to change what the system could see before making a claim about what it could recognise. It also changed the commercial discussion: an image-based inspection promise needed a defined optical environment, with an agreed response when the required feature was not adequately represented.
Withdrawing a line-rate claim that the inspected cycle could not support
A 2023 proposal described our integrated system as capable of inspecting 600 cartons per minute. At that rate each carton would arrive 100 milliseconds after the previous one. A review of the particular non-overlapping configuration measured 125 milliseconds for the complete acquisition, processing and result-association cycle. The claim therefore exceeded the reviewed sequence even before an operating margin was considered. The camera's component specification had been allowed to stand in for the behaviour of the assembled system.
I had supported the proposal and asked the team to withdraw the 600-per-minute statement. A controls supplier suggested retaining the line speed and accepting occasional skipped acquisitions as a support issue. We rejected that approach. An absent inspection result could not silently become an accepted carton. We revised the control logic so missing acquisitions or results produced an explicit exception and followed the agreed equipment and customer-interface response. Mechanical and customer safety requirements remained subject to their proper review; we did not prescribe a universal stop action for every installation.
The correction cost ₹42 lakh and took nine weeks, including timing instrumentation, control changes and integration exercises. We demonstrated a 400-carton-per-minute configuration under the stated test conditions, which allowed 150 milliseconds between cartons. The 125-millisecond reviewed cycle and that interval described this configuration; other image sizes, algorithms, materials or line interfaces would require separate assessment. We retained those dependencies in the technical offer and identified what would need reassessment for any higher-rate proposal.
The lower supported rate weakened the commercial case for one opportunity, and the board had to consider that consequence alongside the cost already committed. I did not present the change as a hidden efficiency saving. The lesson was about the level at which a promise was made. A component can meet its specification while the system fails to support the proposed service. As CTO and director, I needed to connect the complete timing evidence with the rate we were prepared to sell, including a visible treatment of every expected but missing result.
Preventing an obsolete reference from producing a reassuring match
In 2025 we ran sixteen controlled changeover exercises on a proposed operator workflow. Three selected an obsolete artwork reference even though the job was intended to use the current approved revision. The comparison software matched the samples against the selected image as designed. That was precisely the problem: successful matching did not establish that the chosen reference was the one the customer had authorised for the job. The exercise exposed a configuration weakness before the workflow was extended.
The product team wanted to solve it by adding a warning beside the reference filename. That would help an attentive operator but would leave the machine configuration weakly connected to the approval. I proposed binding the job to a specific approved revision, showing its identity clearly and preventing an unapproved substitution from proceeding through the normal changeover path. The customer retained the responsibility to approve and supply artwork. Our equipment did not decide whether a label's wording met regulatory requirements or whether a pharmaceutical batch should be released.
We spent ₹36 lakh over seven weeks on configuration changes, revision records, review screens and repeat exercises. The revised approach added an explicit approval step and made some changeovers slower when the expected reference was unavailable. The commercial manager opposed that friction because the product had been promoted as reducing setup time. I acknowledged the cost and withdrew the broader speed claim until the revised workflow had been assessed. Saving an operator a step was not useful if it weakened the connection between the job and its authorised reference.
In the repeated sixteen exercises, the obsolete selections were blocked from the normal approved path and routed for review. Wider use still depended on the customer's reference-management process and the relevant installation assessment. We reported that scope rather than claim the system could never use the wrong artwork. The case changed how I describe automation: the engineering must preserve an approval and its revision before simplifying the action taken from it. A green match result should not conceal the fact that the machine was asked to compare against yesterday's decision.