NYUS

From the founder

AI Localisation for India Is a Data Problem

Translation is the interface. The food-composition data underneath is what decides whether an AI coach understands an Indian plate.

By Shaan, founder of NYUS ·

There is a comfortable story about building AI products for India. Take a product that works in San Francisco, translate the interface into a few Indian languages, rework the onboarding, swap the imagery, and a Western product becomes an Indian one. It is comfortable because every step is tractable. Translation ships. It demos well. It is the easy part.

The hard part is the data underneath the model, and that data does not translate. It has to be built. For most consumer-AI categories in India, localisation is not mainly a language or UI task. It is a reference-data task: slow, unglamorous and usually under-budgeted. The category I know best, nutrition and fitness tracking, is a clear example.

I build NYUS, a fitness and nutrition app, and disclose that interest up front. The nutrition examples here are illustrative, not dietary advice.

Translation is the interface; localisation is the substrate

A model is only as local as the data it reasons over. You can render a fitness app's entire interface in Hindi, Tamil and Marathi and it will still be a foreign product if the numbers it computes assume a foreign diet. The intelligence in these apps (the calorie maths, the macro targets, the recommendations) sits on top of a food-composition layer. If that layer was assembled around Western eating, no amount of interface polish reaches it. Users feel the gap immediately: the app is fluent in their language and illiterate about their plate.

When your input is a preparation, not a product

Western consumer nutrition data assumes packaged, unit-portioned inputs. A chicken breast weighs what the pack says. A slice of bread is a slice of bread. That assumption does a lot of quiet work: it means each item maps cleanly to one row in a database.

Indian home food breaks the assumption. The plate is built from preparations, not products, and a preparation carries no fixed number. An idli depends on a batter ratio that is a household's own. A khichdi's macros move with how much ghee went in. A restaurant gravy might be thickened with curd, cashew paste or fried onion, and the menu names none of it. Even a staple varies: as our piece on why "2 rotis" isn't one number shows, size and thickness alone shift the count. A model built to look up "grilled chicken, 100 g" has nothing to resolve when the input is "one katori of whatever was cooked at home tonight". The model is not wrong; its reference data is foreign to the problem.

India's public food data is a foundation, not a finished dataset

India is not starting from zero. The Indian Food Composition Tables 2017, published by ICMR's National Institute of Nutrition, report 151 nutrient components for 528 key foods, analysed from samples collected across six regions. It is the credible public reference for any serious nutrition dataset here.

But by design it describes ingredients, not meals: the tables state that, except for eggs, all values are for foods in raw form. Cooked, composed dishes, regional variants and the everyday home preparations that dominate real logging have to be built on top, along with portion conventions like the katori and the ladle. Doing that honestly means mapping regional and global dishes to macros, encoding sensible portion logic, and being explicit about where a value is a measured figure and where it is an estimate. That is labelling and data-engineering work. A bigger model doesn't do it for you.

Clean local data is what lets the loop adapt

This matters for AI specifically, not just for a static food database, because everything modern about these products is a feedback loop. An adaptive coach reads what someone logged this week across training, nutrition and recovery and adjusts next week's plan. That loop is only as trustworthy as its inputs. If a katori of dal resolves to a wild guess, the adaptation is guessing too, and personalisation degrades into noise. Bad input doesn't produce one bad number; it affects every downstream decision the system makes.

This is the bet behind NYUS. Its food database covers 8,900+ foods with macros, logged by photo, barcode or search, and a weekly check-in adjusts calories, macros and step goal from what a user logs. I don't offer it as the answer; plenty of teams could and should build this. I offer it as evidence about where the work is. The difficult, defensible part was never the model or the interface. It was the data. (For a sense of what that data looks like to a user, see our reference table of calories and macros in common Indian foods.)

What "AI for Bharat" actually costs

If localisation is a data discipline, teams building consumer AI for India should budget and staff it like one. Reference data, portion modelling, labelling pipelines and honest gap-tracking deserve the line items usually reserved for model work. The uncomfortable implication for a field that loves models is that the lasting advantage in Indian consumer AI is often its least fashionable asset: a clean, local, well-labelled dataset nobody else bothered to build.

Translation gets you in the door. Data is what earns you the right to stay.

NYUS is in early access

An AI fitness and diet coach built for Indian food: a weekly check-in that adjusts your calories, macros and step goal from what you log, 8,900+ foods with macros, in English, Hindi and Hinglish. No ads, no data sold.

Join the early-access list

Frequently asked questions

Why do global calorie apps get Indian food wrong?

Most were built around packaged, labelled food where one item maps to one database row. Indian home cooking is made of preparations that vary with oil, ghee, ratios and portion size, so a generic entry is often a rough guess.

Does the IFCT list calories for cooked dishes like dal or roti?

IFCT 2017 reports nutrients for 528 key foods and states that, except for eggs, the values are for foods in raw form. Cooked dishes have to be estimated from those ingredients using recipes and portion sizes.

How many foods does NYUS have?

NYUS has 8,900+ foods with macros, which you can log by photo, barcode or search.

Sources

  1. IFCT 2017 published by ICMR-National Institute of Nutrition; 151 components for 528 key foods; samples from six regions; except eggs, data are for raw foods