SCULPT — Avito Case Study

The Challenge of Design Reviews in Large Teams

In a large design team, getting design approval takes more time than the design work itself. Besides the art director's routine, it's an endless cycle of "schedule a meeting, make edits, schedule another meeting, and repeat." We solved this by implementing a custom AI art director, Zero-Defect Design AI™.

Reviewing a single layout at Avito requires input from up to five different specialists: senior designers, colleagues from related verticals, design system guardians, and product managers. Each has their own focus, and knowledge is distributed.

The result: prolonged cycles and dozens of hours spent on synchronization. Or worse, skipping important approval stages and discovering errors in production. A single, objective source of truth was needed.

The Task: Teaching an AI to Think Like an Art Director

A human art director evaluates a design based on two key criteria: overall design quality and adherence to the design system. For example, they check visual hierarchy, typography, spacing, and compliance with accessibility standards. We decided to automate these tasks by teaching an AI to do the same.

The neural network is fed:

The design layout for review.
Rules from the design system.
A clear checklist for analysis.

The output is a detailed and structured report of errors. We took one of Avito's most complex screens—the car ad page (over 20 blocks and 120 UI elements). The AI analyzes the layout and provides a detailed report.

Example of Feedback from the AI Art Director

Zero-Defect Design AI™ doesn't just find mistakes. It highlights strengths, suggests specific improvements, and explains how to bring the layout into full compliance with the design system. The result is a detailed report that replaces dozens of Figma comments and several revision cycles.

Inconsistent Header Style

Section headers (E6.1, E7.1) use a 21px size. The DS rule (p. 92) requires the H25 style, defined as 24px (24/28+12). The style must be replaced for full consistency.

Missing Button Shadow

Secondary buttons (E10.3, E8.4) are designed without a shadow. The DS rule (p. 83) states: "Shadow. Preset only for Default type with Secondary priority." A shadow effect needs to be added.

Low Text Contrast

Secondary text in block B9 uses the color #757575. The contrast ratio of 4.63:1 formally passes the WCAG AA threshold (4.5:1) but is borderline. It is recommended to darken the color to improve accessibility.

Duplicate Entry Point

Buttons E6.4 and E12.4 both lead to the Autoteka report but have different styles and locations. This violates the consistency heuristic. The scenario should be unified, leaving one primary block.

Incorrect Vertical Rhythm

The spacing after block B6, which ends with a button, is 28px. The DS rule (p. 77) requires a 6px compensation for filled objects. The final spacing should be 34px.

Grouping by Proximity

Block B9 (seller information) is an excellent example. The avatar, name, and rating are grouped, and the "Call" and "Message" action buttons are logically separated. This makes the interface predictable.

Finding Stability

Initial tests with powerful LLM models yielded unstable results. The main issue was a high rate of false positives ("hallucinations"), where the AI confidently invented non-existent problems. Standard market solutions have about 30-40% hallucinations—our internal standard was no more than 10%, even on complex mobile layouts.

Example of a hallucination: The model might report an error for not finding a font size that was deeply hidden in the design system and applied to a specific component.

We reduced the false positive rate to below 10% using a special approach: hierarchical mapping of layout elements and structured JSON files that help the neural network navigate the design system. After a series of experiments, we settled on the Gemini model, which showed the best stability and accuracy. The model effectively understood context and provided feedback comparable to that of an experienced specialist.

Quality and Predictability

We intentionally introduced subtle errors into the layouts—the AI successfully caught them every time. It generated structured reports with specific tasks, sometimes noticing things a human eye might miss.

"I liked this review, it's very useful and really saves time. I saw problems that I would have missed myself."

— Feedback from the first designer to test the system

"With AI review, the design team started working much faster, and releases were postponed less often. On average, the speed of layout approval increased by about 5 times, and the workload on art directors decreased by about 20-25 hours per week."

— Feedback from the team's product manager

The Result and What We Gained

Currently, the AI art director operates through a series of prompts and file attachments in a Gemini chat. This is an interim solution to perfectly fine-tune the prompts and process before integrating it into a Figma plugin.

"We reduced the review cycle from several days to 15 minutes, freed up senior designers' time, and improved the final product's quality. The entire implementation process took just three meetings."

— Feedback from Ksenia, the team's design lead

Scalable Expertise

Art directors' knowledge is available 24/7 to the entire team, not just during infrequent calls.

Company Taste

The AI art director operates based on the criteria and perspective of the current art director, passing on the company's taste to designers.

Speed and Focus

80% of routine feedback is generated in minutes, freeing up senior specialists for strategic tasks.

Process Customization

Zero-Defect Design AI™ is fully customizable to your company's unique review processes and standards.

This approach is the new standard for large teams. It saves the most valuable resource—the time of top specialists.