From emissions data to decisions

SINAI already detected anomalies. The catch was when. Data only counted once it had been approved and then locked behind what the team calls the lock gate, and detection only looked at data on the far side of it, with a fixed threshold nobody could adjust.

So people found out something was wrong after they'd already approved it. It was a chronic problem, a bit like a smoke alarm that only goes off once the firefighters have left. Anomalies are something you want to catch before you approve, precisely so you don't have to fix them later.

Once detection moved before the gate, we could rethink what counts as an anomaly, and who decides. The fixed threshold made way for settings that users control themselves.

And then we added a little extra seasoning. Some anomalies are too specific for a slider. So users can now write a prompt describing what looks suspicious to them, set the period and severity, and an agent goes looking for exactly that in their data.

The whole feature was built in three days at a company hackathon, by a trio: one engineer, one QA and me. It was celebrated as the most "production-ready" project of the event, which I still choose to take as a compliment to our QA.

After a single week of refinement, it was live for customers in production.

This one started with two people asking for the same thing at the same time. A client wanted it as a condition to renew. An important prospect, who at that point was a long shot, wanted it too.

I didn't want to design for just two voices, so I also interviewed other customers who collect their own emissions data. Once the first prototypes were ready, they didn't just approve them, they got genuinely excited, which in enterprise software is rarer than a quiet open office.

For the flow to work, a few complicated legacy decisions had to be gently taken apart.

And all of it moved quickly. Working AI-first let me take the feature through the whole process myself, from interviews and prototypes to the full-stack build.

Some pages in the product were essential, the kind that explain where a company's emissions are heading and how it plans to bring them down. And yet, in demos with prospects and clients, people would quietly skip past them. Not a great sign.

Using AI to write front-end code, I started a continuous improvement effort. Before touching anything, I talked to key people across the company, the CEO included, to find out which pages needed some love the most. Then I gave it to them.

The old projections page stacked bars year after year and left you to work out the story. The new one draws the past and the forecast in a single view, adds working hours as context, and finishes with emission hot spots: where to focus first.

Before, targets were a row of cards and the roadmap was a raw line chart with three numbers floating beside it. Now the page opens with the programme at a glance, lets you pick a target, and shows the plan against what's already happened and where your projects will take you, with a nudge to find projects that close the gap.

Continuous improvement is exactly what it sounds like: it doesn't stop. Analyze Emissions got the same treatment, with calmer charts, clearer scopes and a layout that leaves room for the data to speak.