Andrei Chemlev: Understanding AI’s Integration into Web Development

  One of the most discussed and disagreed upon subjects of present day is how AI is changing the world around us. In the immediate sphere of Andrei and his peers, this dialogue steers towards how artificial intelligence is currently influencing modern web development. Two decades ago, the internet was a wheel made of rock with a stick in the center as compared to today’s fully electric computer on wheels that has rejected the internal combustion system concept. To see where the next two decades might lead us, we’ve requested Andrei give us his thoughts on what we will see as AI takes web development into the future.

What was your first experience working with AI?

Andrei Chemlev (AC): In 2016, while working at CTI on an omnichannel contact center platform, we implemented a recommendation system based on TensorFlow 0.x. Back then, AI was solving the task of prioritizing requests: the model ranked tickets by the probability of escalation. This was the first time that machine learning models directly influenced frontend logic — the web client would dynamically rearrange the interface based on predictions. Since then, AI has evolved from an exotic feature into a standard layer of UX personalization.

AI continues to become more immersed in our everyday lives. In terms of your work, how has this more current collaboration been for you?

AC: It’s been extremely productive. Within Wildberries, I primarily use artificial intelligence where it directly enhanced employee productivity and user experience quality. For example, in the partner pickup point (PPP) app, I implemented a built-in LLM assistant: an operator can simply enter a question like “how to process a return without a barcode” or “what to do with a damaged box,” and within seconds receive a clear step-by-step instruction based on the current regulations. If the situation falls outside standard scenarios, the bot immediately opens a support ticket, auto-filling the necessary fields. As a result, typical requests are resolved much faster, and first-line support load has significantly decreased.
 With WB Stream (theproprietary real-time video communication platform Andrei developed at Wildberries), I used TensorFlow JS for local video stream processing: the model segments the participant’s silhouette to softly blur the background or instantly replace it with a corporate template. Everything happens directly in the browser, with no additional server load, so the video remains stable even on weak connections, and participants feel more comfortable without revealing their surroundings. Together, these solutions show how well-chosen AI tools can both reduce operational costs and improve interactions with digital platforms.

It's almost impossible to predict all the areas that AI will move into but are there specific areas that you see AI being used predominantly in for web development in present day?

AC: The most noticeable is UI personalization. Here, models read behavioral telemetry at runtime and, within fractions of a second, adjust layouts, field order, and recommendations. In modern React projects, this is already as standard as lazy loading images. The second front is testing where large language model (LLM) engines generate unit and integration test cases, relieving engineers of routine and allowing them to focus on non-trivial scenarios. I would say that the third area is dynamic graphics generation. At Wildberries, we experimented with diffusion models to render personalized banners for product cards directly in the browser; changes in packaging color or background happened instantly, increasing view-to-click conversion. The next would be developer relations (DevRel). This is when LLM-based bots write architectural decision records (ADR), changelogs, and migration instructions directly into pull requests, turning documentation from a burden into a byproduct of development. The fifth direction is AIOps. In this, platforms like Datadog Causal AI ingest gigabytes of logs and suggest configuration adjustments before users even notice degradation. Finally, there’s development itself: Copilot, Cursor, and similar tools have become a “second keyboard,” speeding up boilerplate code writing, library migrations, and initial prototype drafts.

Do you foresee a time when AI will become the “chief architect” rather than humans steering things?

AC: The evolution is progressive but not linear. After the 2023 hype spike, many teams transitioned from experimental proofs of concept to a “working plateau,” where LLM modules are embedded in regular pipelines. I see AI as a cognitive co-architect: it already generates scaffolding frameworks, helps evolve database schemas, or suggests optimal microservice layouts. However, strategic decisions — prioritizing business goals, balancing CAP properties, addressing ethical and compliance issues — still rest with humans.
 In other words, the scenario where “the machine completely replaces the architect” seems unlikely; a duet model is more realistic, where AI accelerates and humans steer.

That information is comforting. I think most people are concerned about AI replacing the need for humans at all levels. Do you discern any disadvantages regarding escalating the use of AI in your field?

AC: The main problem is explainability. Code generated by an LLM rarely comes with traceability of requirements, which means during audits (SOX or PCI-DSS), one must manually parse each patch set. Security is an obvious risk. Prompt injection can expose private tokens, and auto-completion sometimes “slips in” fragments under proprietary licenses. Technical debt, meaning that if one overlooks quality, it’s easy to end up with “AI spaghetti” — disconnected code snippets with no unified architecture. Skills gap is problematic. A novice who receives an answer from Copilot may skip fundamental concepts like HTTP or basic algorithms. And finally, economics. A serious inference cluster on GPUs is a significant expense, especially for on-premises installations in regions with expensive energy.

Okay, you’ve explained the pitfalls as well as given the benefit of your experience with the growth of AI in web development but how do we get to a place where humans and AI collaborate at an optimal level? What is the idyllic model of this situation when viewed by an expert such as yourself?

AC: I envision a symbiosis where AI plays the role of a “cognitive pair programmer,” suggesting implementation options, while the developer validates them against service level objectives and non-functional requirements. The second layer an observability co-pilot role in which the agent continuously analyzes traces, identifies “N+1” queries, and suggests preloading indexes. The human, in turn, defines the domain model, monitors ethical constraints, controls data flow, and performs human-in-the-loop reviews — the “four-eyes principle” that ensures the engineer has the final say.

Moving forward, how do we ensure that AI doesn’t get out of the boundaries we’ve set? I think in general, people are excited about the possibilities which come with AI but have a mostly unspoken concern about possible threats.

AC: I follow the principle of amortized risk: AI brings exponential productivity gains but requires investments in control mechanisms. I support the initiatives of OpenAI and the Partnership on AI: data transparency, layers of explainability, and mandatory model audits should become the norm. In simpler terms, AI is a powerful excavator — it can dig a trench in a single evening, but only if we put up barriers and know where the utilities are buried.

As someone who is deeply invested in facilitating communication and education in the overall tech industry, what areas do you feel should be discussed and investigated more to facilitate progress and understanding?

AC: I believe it is promising to discuss AI-driven observability, where LLM agents automatically compile root cause analysis (RCA) reports and even generate Terraform patches. Edge AI in progressive web applications (PWA) is also of interest — on-device inference via WebGPU allows offline apps to remain “smart” without constant cloud access. Another area is AI-powered accessibility: the generation of alternative text and gesture-based subtitles elevates WCAG support to a new level. Finally, we should discuss Data Contracts and synthetic data as a safe way to exchange test datasets across distributed teams. Contemplating and discussing these in an open forum allows for the exchange of ideas that leads to more positive results for all involved parties.