Sixty Cents a Year

Summer is winding down for most of the country. In Florida it just gets slightly less aggressive about it, and I’ve spent the week thinking about ice cream: six tubs since February, two of them free. The thing that found the free ones costs sixty cents a year to run.

What it actually does

It’s a cron job that pulls the weekly grocery ad, filters it, matches what’s on sale against what I actually buy, and emails me when something lands. Those are four steps a computer has done reliably since before I had a driver’s license, and there’s exactly one model call in the whole thing, sitting at the very end.

On July 12 I bought a single tub at full price. Six dollars and change, no discount, no alert. The robot had nothing to say that week and I bought ice cream anyway, which is honest data about how much I actually trust it.

On August 10 the email came. I walked out with five tubs and paid for three.

Why there’s a model in there at all

Because of lines like this one, printed on the July receipt:

B&J ICE CREAM      6.23 T F

You can’t write an assertion for that, because a rule doesn’t know whether it’s ice cream or a typo. Every receipt invents new abbreviations for things it has printed a thousand times before. So one model call reads the line, turns it into a brand and a product and a price, and hands it back, and then the boring machinery takes over again.

The part I’ve stopped trying to resolve

There is an enormous conversation happening right now about enterprise AI. I’m in a lot of it, most weeks, with people who are serious about it.

The piece of software that has most reliably improved my week this summer is a sixty-cent robot that watches for ice cream.

Both of those are true at the same time, and I’ve given up trying to make them resolve into something tidier.

The thing has a name

When I mentioned the cost in a comment thread a couple of weeks ago, someone replied: “There’s still room for expert systems!”

He’s right, and the name matters, because what I built has one: a rules engine with a single narrow exception handler bolted to the end. That pattern has a forty-year history, a literature, and a well-documented set of failure modes that somebody already wrote down for me.

Deterministic rules never stopped working. They stopped being fashionable, which is a different thing, and fashion comes back around.

What it saved, which is almost nothing

On ice cream, the robot has saved me $13.10.

I’ve spent $1,024 at that store this year and saved $151.85 of it, a little under 13% off the top. The ice cream is 8.6% of that saving, which says more about the saving than it does about the ice cream. Sixty cents to recover $13.10 is a return of nearly 22 times on operating cost and an argument about absolutely nothing.

None of this says everything should cost sixty cents. The ambitious version costs more because it’s attempting something harder, and plenty of the time that’s the right trade.

I built this one to find out whether the boring version would hold up. It’s been running since February, it has never once paged me, and I go weeks without thinking about it.

That’s the highest compliment I have for software.