User Interface Development for the Primo RPA AI Server

Overview

A service was launched within the company that significantly improved the workflow for text recognition models on documents: loading a dataset from company documents, field tagging, model training, and the subsequent use of the model in other products. This approach provided higher recognition quality and met security requirements—data remained within the company’s network. As work on the product progressed, functionality was added to fine-tune the base neural networks for the company’s specific tasks and to work with skills such as summarization, extraction, generation, and other scenarios.

I was invited to join the project as a UX/UI designer; prior to that, the developers had been building the interface on their own. As a result of this work, the service’s barrier to entry was lowered—significantly more people could use the product without specialized training in AI and OCR.

Before and After

The service was built using Ant Design on top of React with the Hope UI visual theme for Bootstrap. The company already had several web products and wanted to unify and refresh the user interface, so the development team relied on components from those systems.

I immersed myself in the product and, together with the team, analyzed all the functionality in detail.

I mapped out user flows based on current scenarios: these showed the team where confusion arose in the interface and user journeys. Based on this, we redesigned the configuration processes themselves: we automated and hid some functions to reduce the risk of errors and eliminate unnecessary routine tasks.

The new approach changed both the perception of the product and the setup process. The interface guides users through the setup steps without disrupting their usual workflows: a fully linear wizard isn’t feasible here—users still need to go back and refine settings to ensure high-quality results.

Questions for the team regarding the current status

The answers will help quantify the impact of the implementation with facts and figures.

• Has the number of support requests decreased?
• Are users able to complete project setup on their own from start to finish?
• Does the new interface allow users to create new types of projects based on existing patterns?

Interface Patterns

Below are the key areas of the product: how they support learning and model-working scenarios, and how they differ from the previous approach.

Field Mapping

Mapping links document fields to the target data structure. It was important to make the relationship visual and reversible so that users don’t lose context when making edits.

Dataset, Training, and Quality Control

The dataset editor, training launch, and quality screens form a single mental path: data preparation → training → result verification.

Inference in the project

Special Focus: Running and Monitoring the Model Within the Project: The user must understand the status of the inference run and stay in sync with the rest of the settings.

Expand

Full-Screen Mode and Import

This isn’t about “long tasks,” but rather about processes in a separate context: some block work on the project until they’re finished, while others run in parallel in a separate window, allowing you to continue configuring project settings. We moved this functionality to a separate window because there were several entry points for these tasks, and a standard modal wasn’t enough—you had to complete a separate process elsewhere and then return to the project settings. This approach reduced the amount of development and testing required.

Below are three full-screen canvas options: import in a separate window, panel mode, and the minimalist “zen” mode.

Future Development

Several major items remained on the roadmap: scenarios related to the LLM (lists, statuses, actions), a step-by-step training wizard with progress tracking and undo functionality, and a full-featured OCR engine with region marking and on-page quality previews. These were not included in the minimum set for the interface’s first production release, but the established patterns should accommodate them without major rework.

Process and Handoff

Work proceeded through regular synchronization with the development team: mockups, specifications, and a phased handoff to the front end. After I left the project, the team looked for the next designer and structured the onboarding process based on the existing design assets and layout guidelines.

Questions for the Next Designer

• What turned out to be unclear during the first two weeks on the project?
• Which artifacts and handover guidelines proved most useful?
• What had to be redesigned and why.
• How would you describe the current state of UX in three sentences?

Conclusion

The AI Server interface in Primo RPA was transformed from an “engineer-oriented developer interface” into a product with intuitive configuration logic and user guidance. The focus on user flows and routine automation reduced the implementation burden and paved the way for broader adoption within the company and among clients.