The Zero-Margin Kitchen

Baba & Co is a Denver food-service company that cooks for the city's shelter system: about 600 meals a day in the summer months, up to 1,300 a day when the cold arrives, plus the snacks, prepackaged meals, and drinks that keep a shelter running between meal windows. We operate the kitchens ourselves. And we run them with the help of agentic AI systems we build in-house, with a person approving every consequential decision before it touches a meal.

This is the first in a series of essays about what that actually looks like. It starts with why we chose this kitchen, the hardest one, first.

I didn't leave tech. I took it somewhere it's never had to survive.

I spent over a decade in software: engineering at Instagram, then Modern Animal, most recently as a director of engineering. Then my husband, whose background is data science and experimentation, and I started a company that feeds Denver's shelters. Neither of us came up through restaurant kitchens, which is exactly why we had to build systems instead of leaning on instinct. Most people who knew me assumed I left the industry. I didn't. I moved to the part of the economy it never reached.

The constraint is the point

Here is the shape of the work: the people we cook for need the most support and have the least resources. There is no pricing power. You can't charge a shelter more. No premium tier. No slack. Every dollar of operational waste comes out of someone's dinner.

And the demand curve doesn't hold still. Our volume roughly doubles between July and January, because winter fills shelters. The same kitchen, the same team, and the same budget discipline have to absorb a seasonal swing that most restaurants would call two different businesses.

Meals per day, summer

~600

Meals per day, cold months

up to 1,300

Seasonal volume swing

~2×

Consequential decisions made without a human

0

Which makes this a perfect forcing function. The only lever available in a zero-margin operation is complexity reduction, so that's what we build. Our systems draft the work that carries the kitchen's cognitive load: menus checked against the full dietary matrix, purchasing plans against price movement, prep sequencing against delivery windows. Every consequential decision has an approval gate, because the system is wrong in ways that only someone standing in the kitchen can catch.

Why the lowest margin first

We're not trying to fix all of food service yet. We're doing the solvable thing now: making it operationally sustainable to feed the people who are hardest to feed. The margin we recover becomes meals.

That's the bet. If AI-assisted operations can hold up where there is no pricing power, no premium tier, and no room for error, then the approach generalizes to every kitchen above ours on the margin ladder. The reverse is not true, which is why a decade of well-funded food-tech proved so little.

What this series will cover

One operational constraint at a time: what it costs in a zero-slack environment, who carries it today, what the system absorbs, and where it fails. These essays are about the workers and the systems. The people we serve are the reason for the work, not material for it.

If you're putting AI into physical operations, or you're curious what the oldest industry on earth does to an engineer's assumptions, I'd like to compare notes. The door's open.