Overview

[Back]

Jua is a frontier AI lab that provides energy traders with forecasting intelligence. I owned the strategy and design of v2 end-to-end.

The coded v2 prototype: live data and reusable components

Product

  • B2B
  • SaaS

Team

  • Engineers
  • Stakeholders
  • Product Management

Role

Lead Product Designer

The 15-minute window from forecast signal to trader decision

Fig. 1 A race against time

Forecasts are only worth what a trader does with them in 15 minutes or less. I designed v2 around the fastest path from signal to decision.

Trader workflow map showing dashboards, spreadsheets, model outputs and exchange feeds

Fig. 2 Beyond the data

The team’s user-research recordings showed the model was half the product. Judging whether to act on a forecast was the real bottleneck.

The 3 questions, applied to every proposed surface in v2

Fig. 3 Finding purpose

Every surface had to pay rent: remove manual work, support a decision or make the model easier to question. What couldn’t was cut.

Attention as the binding constraint: what passed and what was cut

Fig. 4 Understanding resources

Attention was the binding constraint; traders had none to spare. Anything that slowed traders or the ship date was out. 3 concepts were cut.

The cut widgets concept: traders composing a dashboard from blocks

Fig. 5 Paradox of choice

Widgets were popular in early discussions. But configurable dashboards spend the one resource traders lack: time to synthesise. I cut them.

The cut portfolio surface concept: forecasts tied to asset exposure

Fig. 6 Addressing feasibility

The portfolio surface bridged a forecast to P&L. Operations flagged adoption cost, sensitive data and maintenance, so I parked it for later.

The cut chat modal concept: a floating AI prompt box

Fig. 7 A better workspace

An AI modal was the wrong form factor. Traders compare scenarios and interrogate reasoning, so I gave the AI a dedicated workspace instead.

The coded v2 prototype and design-system foundation

Fig. 8 From 0-1

v2 became 5 responsive surfaces, each with one job. I built the design system and entire prototype in code, using Cursor and ShadCN/UI.

Fig. 9 Getting oriented

The summary is the model’s first read on what matters and why. Detail stays 1 click away; nobody digs through raw output to get oriented.

Fig. 10 Comparing futures

Scenarios puts forecast paths side-by-side with a confidence score I proposed. Traders compare possible futures instead of betting on one.

Fig. 11 Geospatial reasoning

I improved v1’s map, which shows where a signal sits and which assets are in its path. Traders judge nothing until they know the where.

Fig. 12 Finding signal

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

Fig. 13 Agent transparency

The AI Workspace is where retention starts: interrogation builds trust, insights provide value. Traders question the LLM’s reasoning here.

0

static handoffs

44%

scope reduction

≤100%

inspectable AI

v2’s engineering blueprint was the prototype itself; because the design was already running in code, decisions survived to production without translation. The chat modal I advocated for was convenient rather than robust, and traders’ hardest questions exposed that quickly. They became the bar every later surface had to clear: an eval set in all but name.

To see the full version:

Back to top

[↑]