Shambhavi
Gupta

The problem

Teams were turning long briefs and discussion threads into work items by hand. Someone reads the document, decides what needs doing, and retypes it as tasks. It is slow, it comes out differently depending on who did it, and the link back to why a task exists disappears the moment it is created. Weeks later, nobody on the board can answer why something is on it.

My approach

The obvious build is to let the model write straight into the work queue, and I argued against it: that puts something nobody reviewed into the list people are actually working from. So I drew a hard line instead. The AI proposes, a person approves, and only then does the platform create anything. Try to break it below.

The problem

A $1.1 billion government energy rebate was about to land on customer accounts. AGL needed to know whether that money would genuinely pay down its debt book, or whether customers already struggling would absorb it without their payment behaviour changing at all. The answer drives how much bad debt to provision for and where to point collections effort.

My approach

A single-point forecast gives the business one number to plan around, which for a provisioning decision is close to the least useful shape. What you need is how wrong that number could be. So I argued for Monte Carlo simulation over 25 customer segments, and bootstrapped the segments where the evidence was thinnest rather than assuming a figure.

The problem

Wheat farmers across four Indian states were making sowing and irrigation decisions from experience rather than measurement, with no reliable read on soil nutrients, moisture or water quality, and no direct route to a buyer. Most of them were also illiterate or semi-literate, and a large share did not own a phone.

My approach

Almost every product in this space quietly assumes a smartphone in someone's pocket. Taking that away is what shaped everything: the screen had to come back onto the device, and the readings had to be legible at a glance without reading text. Flip the assumption below and watch what has to change.