Calcuplator

Overview:

Calcuplator is a meal prep–focused nutrition calculator built to estimate macros and cost for real-world food using LLMs, AI Vision and domain heuristics

Tracking calories is easy when food comes with labels. But once you start cooking your own meals — especially with fresh ingredients or Asian groceries, nutritional data becomes vague, incomplete, or completely missing.

Calcuplator focuses on what matters for real-life meal prep:

• Estimated macros per serving
• Cost per meal
• Ingredient-level breakdowns
• Minimal friction, no manual data entry

Instead of relying purely on food databases, Calcuplator uses a combination of AI Vision and LLM-powered understanding and domain heuristics to estimate macros in a way that’s good enough to be useful, not falsely precise.

The app is currently on TestFlight, with 300+ testers joining within the first week.

Journey:

One of my “new year, new me” goals this year was to meal prep, eat healthier and be more mindful.

I’ve been meal prepping for a while now. Calories are easy when you eat packaged food labels are there, numbers are clear, you can eyeball it. But once you start meal prepping… things get fuzzy.

How much protein am I actually eating? What about sodium, fats, cost per meal?

My first instinct was Google Sheets. Build meals, plug in macros, let formulas do the work.

Halfway through, something clicked:

Why am I doing this in Sheets? Why not just build an app?

A few days of overclocking my brain later, I had a minimal app structure:

• Build meals from ingredients
• See macros per serving
• See cost per serving

Then reality hit.

Most things we buy and eat, especially Asian groceries don’t have nutrition labels and aren’t in online food databases like open food facts.

For the first week, I manually keyed in nutritional info. That was… completely counter-intuitive. Not what I wanted. I wanted hassle-free. So I turned to LLMs. I built a framework using Gemini Vision + Gemini Flash to:

• Estimate food amounts
• Infer macros
• Convert messy prompts into structured data

What I learned was interesting: LLMs are very confident with Western food, but struggle with nuance when it comes to Asian ingredients and dishes.

So with Calcuplator, I took a different approach:

• Use LLMs to understand the dish or ingredient context
• Apply domain-specific heuristics
• Estimate macros based on what we know, not blind guesses

The result is something that feels robust enough for real life, not just demos.

Calcuplator is now on TestFlight,

and hit 300+ testers within the first week.

Still early. Still rough around the edges.

But it solves my problem — and apparently, a few hundred others too.

More to come 🚀