Flotation Operator Dashboard

Delivering Scientific-AI recommendations to flotation control operators. Designed for Intellisense.io

flotation optimizer app

Background

At Intellisense.io we have built a first-in-class application for control operators working at flotation plants. The operators want to know how to achieve best flotation performance, which normally means to recover as much metal as possible. The app simulates a scenario with best performance based on plant equipment and material, and recommends control setpoints to the operator, who reviews them and enters in their control system.

My role

Research
I was involved in the end-to-end product design process. Starting from product discovery, we faced a few challenges: the language barrier between the product team and the users, and limited access to them due to their job nature. To mitigate this, I partnered with the company's domain experts, who worked with customers directly, and arranged interviews to learn how operators work and the flotation process and identify questions we had to ask customers.

Design
This helped me outline user workflows and start working on UI prototypes, which I validated with domain experts and customers. The feedback from the workshops led us to change the application concept from a calculator into an operational dashboard that delivers recommendations in real-time.

Ship
I contributed in defining what scope the first version should have, designed high-fidelity mockups in Figma to workshop with engineers and QA. After launching the first version the app evolved through numerous iterations based on constant feedback. I've been involved in continuous product discovery, listening for user feedback and solving new problems with the team.

Highlights

Cognitive load
Discovery showed that operators’ job is hectic, their focus and attention is scattered across numerous screens with alarms popping up, and any additional cognitive load wouldn't be appreciated. I designed the UI in a way its structure resembles flotation circuit and highlights setpoints that require action from operator. This reduced time operator had to spend on figuring out what needs to be done.

recommendation alerts shown to the operator

Explainability
Another challenge we faced was explainability. A common problem for AI-related products is users thinking of the app as a "black-box". It is hard to trust recommendations if the reasoning is hidden. To address this, the app shows what is it trying to achieve at the moment, i.e. which target is being prioritized. And explains performance impact of the recommendations, what will happen with the performance in case the recommendations are followed vs. if ignored.

performance forecast

Automation
Expecting operators to review the recommendations and enter them manually to the control system is not always reasonable, as depending on the circuit size the time operators spend on this may become more than they can afford. At many sites we received request to automate the process of writing the recommendations to the control system that gave a start to a completely new workflow.

Impact

The application helped to test the model performance at number of sites and start a conversation with operators. This became a tool for commissioning the product at a site, and opened a stream of feedback on the product usability and model accuracy based upon we improve the product to this day.