At Wikium, a platform with approximately 7.7 million registered users, a minor product assumption could become widespread behaviour. Zykov acquired an equity stake in the company in late 2018 and joined its executive team as Chief Technology Officer. His team tested an adaptive training model intended to make cognitive exercises more relevant. Within days, users were overriding the system more often, not less: the model was treating a distracted session as a reliable change in ability.
Dmytro Zykov, a product architect and technology leader whose work spans cognitive training, neurotechnology and digital mental health, led the product and technical architecture for the redesign. The episode became a practical lesson in cautious personalisation. The same worry about acting on an unstable signal later shaped how Wikium’s BrainLink integration was designed, tested and used during 2022.
We spoke about what the first model got wrong, why a failed onboarding experiment was more useful than a clean success story, and what it takes to use EEG in a consumer product without pretending to read minds.
TechBullion: You opened our conversation with a failure, which is unusual for a product interview. What went wrong?
Dmytro Zykov: In late 2021 we were looking at the first week of use and saw a stubborn pattern. A sizeable group completed the baseline assessment, trained two or three times, then disappeared. Some people were manually lowering the difficulty; others were moving around the catalogue because the recommended sequence felt wrong for them.
We designed a hybrid recommendation model that combined product rules with a neural-network component trained on anonymised session data. It used baseline scores, response time, accuracy and recent performance. The model went live to a limited cohort in January 2022. The idea was sensible. The behaviour wasn’t. One poor session could pull the next recommendation down too far, while a single unusually good session could push it up. We had built something responsive, but it was responding to noise.
TechBullion: So the model could punish somebody for having a bad Tuesday?
Dmytro Zykov: That’s almost exactly how one of our methodology leads put it in the review meeting. She asked, ‘Are we measuring ability, or are we measuring whether someone had a bad Tuesday?’
That question changed the discussion. We were no longer debating model sophistication; we were debating evidence thresholds. I asked the team to stop using the latest result as the dominant input. We moved to a rolling window of three sessions, added a confidence score and capped any change at one difficulty level per session. The user could still override the recommendation, which gave us another useful behavioural cue.
TechBullion: Who was involved in that decision?
Dmytro Zykov: It was a small cross-functional group: a product manager, data analyst, ML engineer, backend and frontend developers, plus the cognitive methodology lead. I was responsible for the product and technical architecture. The useful tension came from the methodology side. Engineering wanted the system to react quickly; the specialists were more concerned about stability. Both were right, so the design had to accommodate both.
TechBullion: How long did it take to know whether the correction worked?
Dmytro Zykov: We put the revised model into production in March and waited. By July, the experiment covered about 31,000 users. Completion of the first week rose from 63 to 71 per cent. Seven-day retention improved by 10.6 per cent relative to control, and manual changes to difficulty fell by 23 per cent. The average user completed 5.8 sessions in the first 14 days, compared with 5.1 before.
Premium conversion did not move in a statistically meaningful way.
TechBullion: That last line tends to disappear from case studies.
Dmytro Zykov: Yes. It didn’t fit the neat version of the story, but it mattered. We had improved the early experience and retention. We had not proved a revenue effect. If you collapse those two claims, the experiment becomes marketing.
TechBullion: You joined Wikium as CTO in late 2018, after the product had already established demand. How do you separate that earlier success from your own contribution?
Dmytro Zykov: The distinction matters. Wikium had already proved demand for online cognitive training. It had registrations, premium customers, public growth and interest from corporate buyers. I don’t think it would be credible to present that foundation as my achievement.
The problem I owned came later. An established service had to become more coherent as a product: the assessment needed to influence the training path, recommendations had to survive scrutiny, the technology had to support a much larger audience, and the same core system had to work for individual and corporate use. I led the product and technical architecture behind that stage. The January 2022 model, the HADI work and the BrainLink integration are examples where my decisions can be separated from general company growth and tested against a specific result.
TechBullion: What changed when you started treating Wikium as a platform rather than a catalogue of exercises?
Dmytro Zykov: At one product review we put two user journeys on the screen. They looked surprisingly similar. In the first, a person completed an assessment and arrived at a catalogue of roughly forty exercises. In the other, the same assessment affected the next session, the choice was explained, later performance altered the path, and the product remembered enough to continue the conversation when the user returned. The difference was mostly invisible in a screenshot.
That was productisation in practical terms. Assessment, recommendation, motivation and outcome measurement had to behave as one journey. The team also needed rules for the claims shown to users. If an exercise was intended to train attention, the product could describe the training goal and the user’s performance; it could not quietly turn that into a medical conclusion. The technology, methodology and language had to agree with one another.
Adding more exercises remained useful, but catalogue growth stopped being the main measure of progress. We were more interested in whether people understood the next step, returned often enough to form a routine, and received a path that changed for a defensible reason.
TechBullion: You have described Wikium as a mature product. What actually tells you a cognitive platform has stopped being an experiment?
Dmytro Zykov: Repeatability is the clearest test. A prototype can generate enthusiasm in a demonstration, and a new wellbeing product can attract a burst of curiosity. Maturity appears when people return after the novelty wears off, complete a meaningful sequence, pay again or use the same product in a corporate setting without constant support from the founding team.
I would also separate company results from product results. Growth tells you the business is doing something valuable, but not which product created the value. For Wikium, the stronger material was the combination of a large user base, repeat sessions, premium behaviour, B2B adoption and identifiable changes in retention after a product decision. If a product were acquired or integrated elsewhere, I would want to know what happened afterwards: whether the users, technology or intellectual property continued to be used. A transaction alone is a weak ending to the story.
TechBullion: You are mixing user behaviour, commercial performance and architecture. Which metrics did you trust most?
Dmytro Zykov: It depended on the question. Registrations, activation and acquisition quality told us whether the market understood the proposition. They were useful, but the first session in cognitive training is easy to make entertaining. The harder test came later: return frequency, completed sessions, drop-off points and whether use settled into a pattern rather than ending with the initial curiosity.
Commercial measures answered a different question. Premium conversion, repeat payment and corporate adoption showed whether the product could sustain itself. Architecture entered when we asked whether the same assessment and recommendation logic could support a new audience without a separate product being rebuilt from the ground up. I never found one number that covered all three questions, and I would be wary of a dashboard pretending otherwise.
TechBullion: Where did B2B demand fit into that picture?
Dmytro Zykov: It was a useful pressure test. A consumer may try cognitive training because it sounds interesting. A company needs a clearer reason, whether that is learning, employee development, wellbeing or assessment. The buying conversation forces the product team to explain what is being measured and what value the organisation receives.
It also changes the product beneath the interface. Corporate use brought company accounts, roles, access controls, cohort analytics, group onboarding and reporting into the same system. Privacy questions became more demanding because an employee and an employer should not automatically see the same information. B2B therefore tested more than sales. It showed whether the underlying design could support another distribution model without compromising the individual user’s experience.
TechBullion: Where did the HADI cycle enter the picture?
Dmytro Zykov: The recommendation model told us what to suggest, but users still needed to understand why. Our hypothesis in January 2022 was that a short explanation after the baseline assessment would increase the return to a second session by at least 10 per cent.
The first design was too earnest. It gave people a detailed account of their profile before they reached the training screen. Return to the second session rose by 6.3 per cent, below the target, while onboarding completion fell by 4.8 percentage points. Payment conversion was flat. We had added clarity and friction at the same time.
TechBullion: What did you change?
Dmytro Zykov: We stopped forcing the explanation into the main path. The next version gave one plain sentence about why a session had been selected, with detail available on request. That restored onboarding completion. By June, seven-day retention was 9.1 per cent higher than control, and the share of users completing at least four sessions was up by 7.4 percentage points.
The unexpected part was that the explanation helped habit formation more than purchase intent. It made the next action feel less arbitrary. That was enough to bring people back, but not enough on its own to sell a subscription.
TechBullion: That sounds like a modest interface change rather than AI.
Dmytro Zykov: Much of good product work looks modest from the outside. The difficult part was the reasoning behind that sentence: which data justified the recommendation, how much uncertainty we carried forward, and when the system should say it did not have enough information yet. The copy was short because the reasoning wasn’t.
TechBullion: At roughly the same time, Wikium began integrating the BrainLink EEG device. What was the product idea behind that work?
Dmytro Zykov: We wanted to test whether neurofeedback could become part of a cognitive training session rather than remain a separate graph beside it. BrainLink was an existing consumer EEG device. We did not design or manufacture its hardware. Our work was to connect the device to Wikium and give the incoming indicators a defined role inside the user journey.
The project began in late 2021. We built a connector between BrainLink and the platform, linked the incoming data to the active training session and designed how feedback would appear in the interface and be stored in the user’s history. The first integrated version was presented in 2022. The technical work therefore sat in the software and product layer around the device, not in the EEG electronics.
TechBullion: What was your responsibility?
Dmytro Zykov: I was responsible for the product and technical architecture of the integration. I defined the requirements for the connector, the exchange of data between the device and the platform, the way indicators were associated with a specific session, and how results were written to the user profile and the wider analytics system.
I coordinated the platform developers, interface specialists and the team responsible for the cognitive exercises. A practical part of the work was deciding what should happen when the connection dropped, when the device stopped returning usable indicators or when a session ended unexpectedly. The aim was to make the full journey coherent, from pairing the device to saving the completed session.
TechBullion: What made that integration technically difficult?
Dmytro Zykov: The device and the platform had different ideas of time and state. BrainLink produced a continuous stream of indicators, while Wikium worked with a user account, a defined exercise and a session that could be paused, restarted or abandoned. We had to make sure that the data shown on screen belonged to the correct user and exercise, and that a temporary Bluetooth interruption did not corrupt the result.
We also had to keep the product claims narrow. A consumer EEG device can provide useful feedback, but it is not a clinical electroencephalograph. We therefore treated the signal as an additional input for self-observation and training feedback, not as a diagnosis or a definitive measurement of attention.
TechBullion: How did you evaluate whether the integration worked outside an internal demonstration?
Dmytro Zykov: We looked first at operational behaviour: whether users could connect the device, complete a session, see the feedback and return to the feature without support from the development team. By June 2022, about 18,000 unique BrainLink devices had been linked to the platform and users had completed more than one million neurotraining sessions.
Around 78 per cent of device owners completed another neurotraining session within 30 days of the first connection. Those figures did not prove a medical or cognitive effect. They showed that the integration had moved beyond a prototype and had become a repeatedly used module inside a larger digital product.
TechBullion: Did you also run a controlled technical validation of the signal-processing layer?
Dmytro Zykov: Yes. By June 2022, a controlled technical study had compared BrainLink indicators processed by the integrated software module with a 16-channel research EEG in 186 participants. Correlations with the reference system were 0.84 for alpha-band power and 0.79 for beta-band power. The artefact-detection model achieved an F1 score of 0.90, and the system distinguished a guided focus exercise from rest with balanced accuracy of 84 per cent.
Those results validated the signal-processing and artefact-filtering pipeline under the defined protocol. They did not establish that the neurotraining improved attention, produced a lasting cognitive benefit or had a medical effect. Testing the efficacy of the training itself would have required a separate outcome study with an appropriate control group.
TechBullion: Wikium had around 7.7 million registered users by then. Was there a temptation to put that number next to the BrainLink work?
Dmytro Zykov: By keeping the measures separate. The 7.7 million figure described Wikium as a whole. For the BrainLink work, the relevant evidence was the number of unique linked devices, completed sessions, repeat use and the reliability of the connection flow. A large company metric can provide context, but it cannot be used as a substitute for product-specific evidence.
That distinction also mattered internally. If the integration generated interest but users did not complete or repeat sessions, the size of the wider platform would not rescue the product. We wanted to know whether the module worked as a user journey in its own right.
TechBullion: What did the EEG data actually change inside Wikium?
Dmytro Zykov: In the released integration, the indicators were connected to the active exercise, displayed as feedback during the session and saved with the user’s training history. We did not allow EEG data to make a medical judgement or independently determine the difficulty of an exercise. Behavioural performance remained the stronger signal for personalisation.
There was a temptation to make neurofeedback visually dramatic. We chose a restrained presentation because the purpose was to support the training session, not to make the user stare at an animation and assume that the product could read their mind.
TechBullion: A device like this can generate more data than the product needs. How did you decide what to keep?
Dmytro Zykov: We began with the product action and worked backwards. The platform needed the indicators required for feedback, session state and basic quality control. We did not need to retain every available value indefinitely. The product mainly stored the derived information required to display and review the training session.
That discipline mattered because physiological data can appear valuable simply because it is difficult to obtain. In practice, a behavioural measure may be safer and more useful. We used the BrainLink data only where the platform could interpret it responsibly, and participation in ordinary cognitive training did not depend on using an EEG device.
TechBullion: Who needed to understand why the system made a recommendation or displayed a particular result?
Dmytro Zykov: The user needed an explanation that was clear enough to judge without receiving a technical lecture. The product team needed to see whether the feature changed completion, repeat use or reported usefulness. Engineering needed traceability: which inputs were available, how recent they were and whether the result came from behavioural performance, a product rule or an EEG-derived indicator.
In a mental-health-adjacent setting there was another concern. The product had to communicate uncertainty before the user mistook an estimate for a fact about themselves. That is why we avoided presenting a consumer EEG indicator as a diagnosis or as a complete reading of concentration. It was one input within a defined training scenario.
TechBullion: While you were working at Wikium, you founded Congru.me, which Wikium later acquired and relaunched as Max Mentality. It was a very different kind of product. What, if anything, connected the two?
Dmytro Zykov: The connection wasn’t the EEG technology. Congru.me — later Max Mentality — was a separate line of work, closer to personalised psychological support than to cognitive training. But it confronted us with an important question much earlier: what happens when a digital product starts telling a person something about their own mental state?
Assessment, segmentation, behavioural data and chatbot interaction all helped shape the user’s path. The difficult part was deciding where personalisation ended and overinterpretation began. We had to distinguish between what the user had actually told us, what the system was inferring, and how confidently a chatbot could recommend a next step. Sometimes the responsible response was not another recommendation, but a limit: the system did not know enough to continue.
That experience influenced how I later approached the EEG work at Wikium. A physiological signal can look authoritative simply because it comes from hardware, but it is still only one source of evidence. We kept the claims narrow, avoided treating EEG as a diagnosis and did not allow it to determine difficulty on its own. The Congru.me work had already taught me that, in products close to mental health, sounding more certain than the evidence is not just a product mistake — it can change how a person understands themselves.
TechBullion: If a team came to you now with an ‘AI-powered mental performance product’, what would you challenge first?
Dmytro Zykov: I would ask what decision the technology improves. Teams often begin with a model and then search for somewhere to display its output. That tends to produce a clever demonstration with no stable role in the user journey. The better starting point is a moment of uncertainty: choosing the next exercise, adjusting difficulty, deciding whether to continue a chatbot path or recognising that the product should hand control back to a person.
I would also look at what the team calls value. A large content library can still leave the user without direction. More time in an application may indicate engagement, or it may indicate confusion. Sensitive data can improve personalisation, but it can also create an expensive liability. The team has to define the outcome before optimising it and remain honest about what the system knows, estimates and cannot determine.
I am also wary when ‘AI-powered’ appears on the roadmap late. If the data model, consent process, product workflow and evaluation method were designed without it, placing a model on top rarely creates durable value. The model needs a real operational job inside the product.
TechBullion: Suppose you were designing an adaptive cognitive product from a blank page today. What would the journey look like?
Dmytro Zykov: I would begin with a restrained baseline: what the person wants to work on, where they appear to be starting and what limits the product should respect. That creates a working hypothesis, not a permanent label. The first training or support scenario would then be deliberately simple enough to reveal how the person responds.
The product would observe performance, return behaviour and explicit feedback, then adjust one meaningful part of the next session. After that change, it would check whether the intended outcome moved. If the user was returning but reporting less value, or completing more sessions while understanding the recommendations less, the system would need to reconsider its definition of success.
Consent, data minimisation and explanation would run through the journey from the beginning. There also needs to be a clear boundary where automation stops. Some uncertainty should lead to another observation; some should lead to a human review; some should simply remain unresolved. A responsible product does not have to manufacture an answer for every data point.
TechBullion: This sounds broader than cognitive training. What connects your work in EdTech, neurotechnology, mental health and enterprise products?
Dmytro Zykov: The common thread is that the software is acting on incomplete evidence about a person. A learner, candidate or cognitive-training user is more complicated than the profile the product holds. My job is to decide which simplifications are useful, which are dangerous and how the system learns from what happens next.
The interface may be a recruitment platform, a training exercise, a chatbot or a neurofeedback session. The practical question underneath is the same: what should the product do next, what evidence supports that choice, and what result would make us change it?
TechBullion: Looking back, which result changed your thinking most?
Dmytro Zykov: The same concern ran through the recommendation model, the HADI experiments and the later stages of the BrainLink programme. Across all three, more data did not automatically create a better experience. The product became useful when it knew how long to wait, how much confidence to require and how clearly to explain the next step.
