“The biggest mistake you can make with AI in a product is to put the model at the center of the architecture instead of giving it a clearly defined, limited role.”
This is how Vitalii Linevych opens our conversation. A Full-Stack Software Engineer with hands-on experience implementing AI solutions in high-load web platforms, Linevych currently works at an international EdTech company, while his practical experience with artificial intelligence integration was shaped during his work within the 2acommerce ecosystem — a B2B platform for distributors and retailers.
Despite AI’s rapid rise in popularity across every industry — from automating routine processes to personalizing customer experiences — there are still few specialists who can effectively integrate artificial intelligence into a real product. Most companies face a gap between the marketing promises of AI technologies and engineering reality: the model needs to be embedded into an existing architecture, remain stable under load, and managed for both cost and response quality.
The topic of AI in web development today is oversaturated with marketing claims with marketing promises. Vitalii offers a look at AI integration through the eyes of an engineer working in B2B commerce and EdTech — someone responsible for making these solutions work reliably, stay within budget, and deliver measurable results.
AI Chat as an Onboarding Tool: A Practical Case
Within the 2acommerce ecosystem, Vitalii worked on two areas of AI integration. The first was an AI chat as an onboarding tool for new users. A distributor or retailer coming to the platform for the first time can ask questions in natural language and receive contextual answers about how to use the system.
The practical effect was a reduction in the load on the support team and a shorter time to the new user’s first successful operation. Vitalii emphasizes:
“The key advantage of this solution is that the chat understands the platform’s context. Documentation or FAQs don’t solve the problem because the user often doesn’t know which question to ask. The AI chat lets them formulate a query the way it naturally comes to mind.”
In practice, the effect of the AI chat proved to be measurable: the load on the support team dropped significantly, and new users — distributors and retailers — began completing the path from registration to their first successful operation on the platform more quickly. For a B2B service where every day of onboarding delay means lost revenue, this is a direct business outcome of an engineering solution.
Semantic Search and Data Validation in 2aData
The second area involved data quality analysis in 2aData, a unified product registry aggregating items from various distributors. The core difficulty was that the same product from different suppliers could have dozens of name variations, different units of measurement, and missing or contradictory attributes. A classic text-based search in such a “dirty” catalog produced unsatisfactory results.
The AI layer made it possible to replace it with semantic analysis — a system that analyzes the meaning of a query rather than simply matching character strings. The model proposed matching between items from different distributors, identified duplicates, validated attribute accuracy, and normalized naming conventions. Vitalii notes that this task is closer to data quality and entity resolution — identifying the same objects across different sources — than to classic recommendation systems. Yet technically it solves the same problem: identifying what the user meant within a chaotic dataset.
For the business, this meant a concrete effect: distributors stopped spending time on manual matching of items from different suppliers, order errors caused by naming confusion decreased, and the team gained a tool for automated catalog quality control instead of manually reviewing thousands of records.
Three Key Technical Challenges of AI Integration
The first and most tangible challenge is response latency. A request to a large language model (LLM) takes seconds, while the user expects a reaction within the familiar cadence of a web interface interaction. The team addressed this in several ways: streaming responses — text is displayed to the user as it is generated; asynchronous processing for cases where the result is not needed immediately; and aggressive caching of typical queries.
The second challenge is cost management. With a large user base, every AI request has a specific cost. Vitalii describes a query classification approach: simple and typical queries are handled by cacheor rule-based processing (handling via predefined rules), while complex ones go to the model. In 2aData, AI was engaged only for records that failed deterministic checks, which significantly optimized costs.
The third challenge is the unpredictability of model output. Unlike conventional code, where the same input always produces the same result, a language model can respond differently to an identical query. Therefore, every AI call requires parsing and validation of the response, fallback logic in case of an incorrect result, and logging for subsequent analysis.
“Integrating artificial intelligence means taking responsibility for the consequences of its decisions. The model does not have the final word — it belongs to the engineer who controls how the response reaches the end user,” Vitalii emphasizes.
Evaluating AI Feature Quality: A Different Set of Rules
Standard software testing approaches work poorly for AI features — it is impossible to write a conventional unit test to verify whether a model’s response is good.
Vitalii Linevych describes the approaches the team employed:
- For the AI chat — a “golden set”: a collection of reference questions with expected answers. After every prompt change (a modification to the model’s instructions), an automated check confirms that the new version has not degraded on these examples.
- For data validation in 2aData — a “labeled dataset”: pre-annotated records marked as “correct” and “incorrect,” against which precision and recall were measured before each change.
- A mandatory element was staged rollout: initially, the new version of the AI functionality is launched for internal users or a small percentage of traffic, with metric monitoring and the ability to quickly roll back to the previous version.
Vitalii notes that the testing and quality control approaches for AI features that he refined on 2acommerce projects are also applied in his current work, passing these practices on to colleagues. In this way, the experience of building defensive layers around AI calls, quality evaluation methodologies using golden sets, and staged rollout become part of the team’s engineering culture.
The Role of AI in EdTech: A Realistic Forecast
Vitalii Linevych currently works in the EdTech sector. His vision of AI’s role in this field is shaped by practical experience.
“The most realistic and valuable direction is AI as a personal content navigator. A platform with millions of materials from various authors faces the problem of users being unable to find the right information — or not knowing it exists at all. AI can solve this better than any manual taxonomy,” he explains.
The second direction Vitalii considers promising is adaptive feedback: a full explanation of why an error occurred and a recommendation on what to review.
At the same time, Vitalii is skeptical about fully autonomous content generation:
“Quality and reliability still require human oversight, especially in regulated subject areas. AI accelerates routine tasks, but strategic integration must account for productivity, ethics, and human control. The technology does not relieve the developer of responsibility — on the contrary, it increases it.”
