Case study · Custom GPT
Turning a pastry chef’s ratio framework into a tool that writes balanced, machine-ready frozen-dessert recipes from a single sentence.
The Ninja Creami makes it easy to spin a pint of ice cream at home, but getting a pastry-shop texture is hard. Most home recipes are written in cups and spoons, rarely explain why an ingredient is there, and leave the cook guessing when a batch comes out icy, crumbly, or too sweet.
Professional pastry kitchens solve this with ratios: target ranges for sugar, fat, solids, and stabilizers that predict texture before anything goes in the freezer. That knowledge exists, but it is not packaged for someone standing in their kitchen with a flavor idea.
Batch 83 is built on the whole-wheat sorbet base ratios from NYC chef Alex Stupak. I used that framework as the default model for balance and texture, then extended it into anchor ranges for each dessert style the GPT supports.
| Style | Key ranges | Stabilizer |
|---|---|---|
| Ice cream / gelato | Fat 7–9%, milk solids 10–12%, sugars 19–21%, total solids 38–40% | 0.30–0.40% |
| Sorbet | Fruit purée 40–60%, sugars 20–28%, total solids 30–34% | 0.25–0.40% |
| Sherbet | Fruit 25–40%, dairy fat 2–4%, sugars 20–25% | 0.25–0.40% |
Example requests:
Every ingredient is listed as a weight. Stabilizers appear as individual gram amounts (for example, locust bean gum 1.25 g), never as a blended percentage.
A cook can’t measure “stabilizer blend 0.35%.” If an instruction can’t be followed on a kitchen scale, it isn’t finished.
The default batch is 500 g ± 10 g and is portioned to land slightly above the Creami pint’s max fill line.
The real constraint is the container. Designing around the user’s actual equipment removes a step they would otherwise do themselves.
Stabilizers are limited to locust bean gum, guar, and xanthan, in two named blends. Pectin, citrus fiber, and specialty hydrocolloids are excluded, and liquid syrups are avoided where possible.
A technically perfect recipe that needs a specialty order is a recipe nobody makes.
Every recipe ends with a diagnostics table so the user can see where it sits against the target ranges.
Transparency builds trust in AI output, and it teaches the user enough to start tweaking on their own.
When a recipe is scaled, weights are recalculated from the original percentages, rounded only at the end, and must total the target within 0.1 g, with stabilizers held within 0.01% of target.
Language models drift on arithmetic. Spelling out the rules makes the output predictable.
Using the GPT for real batches exposed problems I would not have caught on paper:
Each fix came from using the product, not from planning it. That loop of shipping, testing, and tightening the spec is the same one I run on software teams.