AI MVP development: from proof of concept to a product with real users

An AI MVP is a first AI product for real users, built once a proof of concept shows the model works on your data. Generator ETR, which we build for PSONI, began that way: a proof of concept compared two language models on one task, and it grew into the web and mobile app we are building now. We do both stages in that order. Building a whole application first means finding out late whether the model can do the job.

This page is for founders and product owners who have an idea for an AI feature and do not yet know whether a model can handle it. Every project starts with a free discovery workshop. We look at your product, your data and the rules you work under, then agree what the AI should do and which test will show that it works. We have our own AI toolkit that we use in discovery. You leave with a scope and an estimate.

Both stages are part of our AI software development. For an MVP without AI at its core, see our product workshops and MVP development services. To speed up a prototype with AI tools, read our article on AI proof of concept development.

For PSONI, the Polish Association for Persons with Intellectual Disability, Generator ETR turns difficult text into easy-to-read language. ETR stands for easy-to-read. The proof of concept ran PLLuM 12B, a Polish language model, and the 26B mixture-of-experts version of Gemma 4 (a design that uses only part of the network for each answer) on the same task. We chose Gemma 4.

A script scores every output with the Gunning Fog readability index and writes a report over a set of test samples. A reviewer who knows the ETR standard then rates the results against the source texts. What they find goes into the next iteration of the prompt and the model, and the fifth iteration is under way. Our page on LLM evaluation describes the method.

In Generator ETR users sign in with an account verified by SMS. They can take a photo of a letter, and the phone reads the text on the device using Google ML Kit and Apple Vision, while PDF documents go through a pipeline on the server. A glossary explains difficult words. Text-to-speech reads the result aloud and highlights the fragment being read, and a counter shows how many requests a user has used.

The project specification requires the app to meet WCAG 2.2 at level AAA.

We cannot name this client, and the project is still in progress. We began in September 2026 with a short spike, a time-boxed study of the data sources. Then we connected ticket sales data from several sources and the client's spreadsheet to a data warehouse, using dlt, dbt on ClickHouse and Dagster for orchestration.

A sales curve shows how ticket sales built up over time for an event that has already ended. The models forecast an event's final sales from comparable finished sales curves, with an empirical interval. They also calculate the sales pace against the median of comparable events, the risk of selling out and anomalies in ticket baskets. The analysis covers the effect of weather on attendance and sales too.

The results go to a portal with an event list, a sales plan and a "how it's going" rating, with user accounts and roles. The portal runs on Django with DRF and Next.js.

• Model quality: the measures from the proof of concept keep scoring every new iteration. Our QA also runs jailbreak tests, which are attempts to push the model off its task, and repeats them on new iterations.
• Forecasts: in the forecasting project a recommendation passes a model gate and a backtest before it reaches the portal. A model gate is a check the recommendation must pass before users see it, and a backtest checks the model against sales that have already ended.
• Accessibility: in the later stages of Generator ETR we check accessibility with screen readers and in the DOCX files the app generates.

Not every AI MVP needs a language model, and sometimes our advice in discovery is to skip it. In the forecasting project, statistical models on the client's sales data solve the task.

Copyright to the code passes to you with each payment, for the uses the contract lists. PSONI owns the code of Generator ETR. After launch we stay on for support and maintenance, because models, prompts and data keep changing.

Generator ETR runs on a 24-month agreement, from February 2026 to February 2028, split into 8 stages. The project is now in stage V. We answer the contact form within 48 hours on working days.

For PSONI the same team writes the prompts, evaluates the output and builds the web and mobile apps. In the forecasting project we are building the data warehouse, the models and the portal. Every project starts with a free discovery workshop, where we agree what the AI should do and which test will show that it works.