Senior Software Engineer Volodymyr Zaplatynskii decided to take part in a hackathon for the first time in eight years in IT. The IT professional was inspired by the release of a public MCP server from the Silpo supermarket chain. The result of the experiment was a voice shopping assistant developed based on artificial intelligence.
Senior Software Engineer Volodymyr Zaplatynskii decided to take part in a hackathon for the first time in eight years in IT. The IT professional was inspired by the release of a public MCP server from the Silpo supermarket chain. The result of the experiment was a voice shopping assistant developed based on artificial intelligence.
However, while creating the agent, the developer encountered amusing but revealing difficulties. The artificial intelligence turned out to be too trusting of the data from the API and did not always do what was expected of it. Volodymyr Zaplatynskii shared a curious case from his own experience: “I asked the agent to buy buckwheat. He brought a bar with buckwheat. The response status was successful. The agent is like a person you sent to the store. He will do what you said, but will not notice that something went wrong. They asked for buckwheat, gave him a bar, brought him a bar.”
According to the engineer, two weeks of work on the project turned into a process of teaching the artificial intelligence to think critically. To make the agent work correctly, Zaplatynskii had to build in additional verification mechanisms:
verify that the result obtained really corresponds to the question asked, and does not simply return a "successful" API status;
teach the agent to distinguish between an empty catalog and a technical failure (if the branch is unavailable, the assistant switches to self-pickup);
Re-read the status of the cart and address immediately before the action, as delivery slots quickly become outdated.
The developer emphasized that the main lesson of the hackathon for him was the difference between the technical application of tools and the ability to control the process: "The main thing I took away from this hackathon is that the point is not to teach an agent to use tools. He will master this in an hour. The point is to teach him to doubt. This is the same thing that distinguishes an experienced developer from a beginner. Only now you have to write it explicitly."
Technically, the application is built on a modern stack: OAuth to MCP is implemented via PKCE and Dynamic Client Registration, tokens are stored in Upstash Redis, and Web Speech API and Respeecher are used for Ukrainian speech recognition and synthesis. The frontend is written in Angular, and the backend runs on Vercel Edge Functions. Volodymyr wrote the main code using Claude Code, personally checking each diff.
The voice-assembled cart is fully synchronized with the Silpo website, retaining the same list of products and total cost. However, the project currently operates in single-account mode and is tied to the developer’s personal profile.
Volodymyr Zaplatynskii opened the project’s source code on GitHub and noted that expanding to multi-user mode would be the first task if he decides to develop the application further.
