Play

I worked on the launch of Yandex AI, a new application for the Turkish market that consists of an AI chat, a content feed, and a built-in browser. The AI chat is the core experience, while the feed and browser support it by bringing users back, creating moments of interest, and guiding them toward meaningful chat interactions.

Short story

---

TLDR
Role and impact

My role was at the intersection of design, product, and technology. I defined solution structures, set quality standards, validated ideas through research, and brought several directions to production.

Key impact

• I launched a zero-to-one generative news video system and scaled it to hundreds of videos per week with controlled quality;
• Redesigned the feed as a system of entry points into the AI chat, based on real user search demand;
• Shaped scenario-based chat responses for News, Sport, and Earthquakes, which together cover the majority of user interests in Turkey;
• In parallel, I delivered the onboarding in three months, reviewing more than ten concepts and running eight qualitative studies to arrive at a realistic, data-backed solution.

Below I describe these areas in more detail.

---

Case I
Generative news videos

I owned this direction at the level of approach, architecture, and quality. I defined what good automated news video quality should look like, translated editorial and motion patterns into scalable rules, designed the first generation pipeline and JSON contract, and built the first end-to-end proof of concept. I also participated in quality iterations, tiers, and moderation criteria.

Problem & constraints

At the start, nothing existed. There was no clear understanding of how automated news videos should look, how they could be produced at scale, or how quality should be defined and improved. Competitor examples were mostly manual, inconsistent, and not suitable for automation.

At the same time, news content rarely comes with usable video footage. AI visual techniques easily create a fake feeling. The solution had to work reliably with text and low-quality static images while preserving editorial credibility.

Solution

I treated video generation as a structured data problem rather than manual editing. Scripts, timing, assets, and animations were described in JSON, while rendering and assembly were handled by a separate deterministic layer. This resulted in a clear “writer → director → renderer” pipeline.

Quality became manageable through DSAT reviews, iterative improvements, a tier system, and moderation.

Impact

The system reached stable production at hundreds of videos per week, covered all major and most mid-level news events, and was later reused by other teams, for example in Maps. It also provided a scalable foundation for additional content verticals.

Play

Play

Play

---

Case II
Feed

In the feed, I focused on the system layer: translating user needs into card formats, defining shared patterns and rules, and connecting cards to meaningful chat entry actions. Product managers provided user queries and priorities; I converted this input into a scalable UI system.

Problem & constraints

The feed initially existed as a visual layout rather than a product mechanism. Its role relative to the AI chat was unclear, and it was not designed to scale as new domains appeared.

Solution & impact

The feed was reframed as a system of entry points into the AI chat. Each card represented a high-frequency user query, selected to maximize coverage.

Instead of designing cards individually, I built a shared system of patterns and rules that scaled with new domains. As a result, the feed evolved from a small base set into a scalable card system covering key domains and user needs, while significantly increasing the share of cards leading users into chat.

Play

---

Case III
News, Sport, Earthquakes in chat

I owned the format and UX structure of chat responses. I built the first scenario-based response using ChatGPT, set the direction toward structured product-style answers, and later focused on UX structure and UI blocks using the existing design system. Prompts and implementation were later handled by PMs and engineering.

Problem & constraints

For large domains like news, sport, and earthquakes, there was no clear definition of a good product-level chat answer. Raw LLM output was inconsistent, hard to scan, and unreliable in depth. Information volume varied greatly between requests, and rigid templates often caused information loss or hallucinations.

Solution & impact

We moved to a scenario-based response model, mapping different user intents to different response formats with flexible structure and predefined UI blocks. Through DSAT reviews and user testing, response quality reached a competitive level and became a stable base for further scaling. Together, these domains cover the majority of user interests in Turkey.

Play

---

Case IV
Onboarding

In onboarding, I acted as a design owner with significant product responsibility. I defined hypotheses, designed solutions, ran research, and presented and defended decisions with stakeholders.

Problem & constraints

There was no shared understanding of what onboarding should optimize for. Stakeholders expected different outcomes, priorities shifted, and timelines before launch were tight. The product combined several complex concepts, and some strong ideas were too expensive to implement in time.

Solution & impact

The solution was to split onboarding into two layers: a short video intro explaining the product, followed by a lightweight product onboarding flow focused on interests and first interaction.

The project took three months, explored more than ten concepts, and included eight qualitative studies. The final solution shipped on time, was backed by research, and left room for future iterations, including chat-based onboarding.

Play

---

Outro

This project focused on building scalable product systems, not individual screens: video generation as a platform, the feed as an entry layer to chat, scenario-based chat responses, and onboarding as a bridge to first meaningful use. Across all areas, I worked as an individual contributor with strong ownership, shaping solutions at the intersection of design, product, and technology and bringing them to production impact.