Jua is a Series A startup building frontier AI models used by intraday energy traders. I owned the strategy and design of v2 from end-to-end.

Problem

Forecasts are only worth what a trader can do with them in 15 minutes or less. Jua’s AI models were excellent, but the path from signal to decision had to be realised in those minutes. I designed v2 around the most optimal path.

Discovery

Studying the team’s interview recordings convinced me the model was only half the product. Traders already had good forecasts; what slowed them down was judging whether to act on one. The harness around the model was where design had the most value to add.

Principles

Every surface in v2 had to pay its rent, by either (a) supporting a decision, (b) removing manual work or (c) making the model easier to question. Anything that couldn’t was cut.

Constraints

The binding constraint was attention, since traders had none to spare. That ruled out anything trying to slow things down, or blocking v2 from shipping. 3 concepts failed that test.

Widgets

Widgets were a popular idea in early discussions. But configurable dashboards spend the one resource traders don’t have: time to synthesise data. The intent was right but the execution model was wrong, so I cut them.

Portfolio

The portfolio surface bridged a forecast to P&L. After speaking with Operations, certain issues became clear: adoption cost, sensitive data and maintenance burden. I moved it out of scope and made a note to revisit it later.

AI Modal

An AI modal ended up being the wrong form factor. Traders needed to compare scenarios, interrogate reasoning and hold context across a session. I cut the floating modal and gave the AI a dedicated workspace.

Solution

v2 became 5 surfaces; each with one job in the trader’s workflow. I designed and built the entire prototype in code, using Cursor, ShadCN/UI and live API data. This allowed engineers to work with React components, instead of screens inside Figma.

Summary

The summary gives traders the model’s first read on what matters and why. Granular detail is still 1 click away, but nobody has to dig through the raw output just to get themselves oriented.

Scenarios

Scenarios puts plausible forecast paths side-by-side, each with a confidence score I proposed to make the model’s certainty legible. Traders can make decisions by comparing possible futures, rather than betting on a single reality.

Map

I improved v1’s map, which shows where a weather signal sits and which assets are in its path. A trader can’t judge a weather event until they know where it will be, so this had to paint a clear picture fast.

Feed

The feed’s job is triage. It surfaces the changes worth a trader’s valuable minutes and drops the rest: a judgement most complex tools push back onto the user.

AI Workspace

The AI Workspace is where the retention chain starts: interrogation builds trust and insights provide value. Traders are able to question the LLM’s reasoning in a space designed exactly for that.

Outcome

The coded prototype became v2’s engineering blueprint. Design decisions survived to production because they were already running in code.