VK showed how AI photo editor “Otredach” went from a joke to a Dev Grants 2025 winner

VK published a case study of developer Anton Lenev, who turned an experimental AI project into the mini-app "Otredach" on the VK Mini Apps platform. The story ended not only with victory in VK Dev Grants 2025, but also with a working product: over the past six months, the service has been visited by more than 109 thousand unique users.

From Idea to Product

The story began back in 2023, not with a photo editor, but with a humorous service for "fortune telling" based on GPT-3.5. Lenev created a simple interface on Next.js, added streaming answer generation, and connected the application with a children's Bluetooth printer so that predictions could be printed immediately. The format unexpectedly took off: videos demonstrating it quickly spread across social networks, and event organizers became interested in the service, liking this way of interacting with neural networks.

The next version appeared in 2024, when models began confidently working with images as input. Then the developer assembled a new prototype: the user uploaded a photo of coffee grounds or tea leaves, and the model interpreted the image and produced a text "prediction". The experiment again resonated with the audience, the project was integrated with VK Bridge, advertising monetization was preserved, and ultimately they won VK Fresh Code 2024 with it, which became the first serious signal about the viability of the idea.

In 2025, Lenev deliberately moved away from the pure entertainment format. Instead of yet another pet project, he decided to assemble a full-fledged product with clearer economics and a mass use case. Against the backdrop of the emergence of Nano Banana, OpenAI GPT Image 1.5, and Seedream 4.5, the developer chose a simple but clear case: the user uploads a photo and receives a stylized version of the image — from a medieval portrait to a retro postcard or a New Year's photo shoot.

How the Service Works

Technically, "Otredach" is built without unnecessary complexity. The frontend uses a standard SPA with mobile-first UX priority, integration via VK Bridge, and an adapted VK UI Kit. Between the interface and model providers stands a BFF layer: it verifies VK signatures, validates input data, sends requests to model APIs, accepts webhooks on completion, and manages the business logic around generations so that the user scenario doesn't fall apart on external service errors.

Instead of heavy infrastructure with PostgreSQL, replicas, and Kubernetes, the developer chose PocketBase on Go on top of SQLite. For moderate load, this proved sufficient: the solution provided a built-in admin panel, migrations, hooks for logic, and very low resource consumption. This stack well reflects Lenev's product approach: first quickly verify demand and unit economics, rather than spending months scaling a system that hasn't yet proven that anyone actually needs it.

The approach to models is particularly interesting. Lenev refused to train his own models and fine-tuning, because for an independent project it's too expensive and organizationally heavy. Instead, "Otredach" works through API aggregators like Replicate and selects a model for a specific visual template, not for an abstract idea of "best quality at any cost". This approach helps simultaneously keep quality, response speed, and generation cost under control.

• Complex visual scenarios use more powerful and expensive models
• Mass templates connect cheaper configurations with acceptable quality
• For each style, a separate prompt is written with strict limitations on composition and face recognizability
• The service automatically returns points if generation failed or hung with the provider
• Repeated requests and unnecessary load are limited by rate limits and duplicate protection

"The chosen direction is not just technically interesting, but also product-meaningful," writes

Lenev.

Economics and Grant

The main part of this case study is not image generation itself, but controlling expenses. Lenev writes that it was already clear at an early stage: without strict economics, such a product would quickly burn through the budget. Only in October 2025, generation costs exceeded $500, so the application was immediately designed to at least not go into the red. For the author, this is not abstract cloud spending, but real money that needs to be recouped through a working monetization model within VK.

To keep the project afloat, the developer didn't chase the most expensive models in each scenario, but selected a balance of price and results. Prompts were optimized for cheaper configurations, and some limitations were introduced not for the sake of interface beauty, but for economic survival: generation limits, duplicate protection, return of in-app currency on failures, and control of request frequency.

As a result, according to the author, the application breaks even with a small profit relative to generation costs, infrastructure, and mandatory payments. In the first stage of VK Dev Grants 2025, Lenev submitted an idea for a different application — "Live Photos", which animates images. Then, on a similar architecture, "Otredach" appeared, and both products began to complement each other: first the user creates an image, then turns it into a short video.

By the competition finale, it was no longer a concept on slides, but a live service with users, working infrastructure, functioning economics, and a clear use case, which ultimately brought victory in the grant program.

What It Means

The "Otredach" case study shows that in AI products, it's not the most complex stack and not a proprietary model that wins, but a quickly validated hypothesis, understandable monetization, and precise packaging of the scenario for a mass platform. For independent developers, this is an important signal: the path from an ironic experiment to a product with a six-digit audience and a VK grant is still real if the project has demand, spending discipline, and enough patience to bring the idea to a working service.