Walk down the packaged-food aisle of any Indian supermarket and the nutrition-data problem looks largely solved. Under the Food Safety and Standards (Labelling and Display) Regulations, 2020, pre-packaged food must carry nutritional information per 100 g or 100 ml (or per single consumption pack), including energy, protein, carbohydrate, total and added sugars and fat, plus each serving's percentage contribution to a daily reference intake. Every pack follows the same format, and a barcode lets an app look the product up in a database that someone has already filled from that label.
Walk into an Indian kitchen and the same problem is unsolved.
A large part of what most Indian households eat never comes in a pack. It is dal made a certain way in a certain house, a roti rolled to a certain thickness, a sabzi with one cook's own ratio of oil to vegetables to spice. None of that carries a label or a barcode, and none of it is standardised.
What the national reference tables do, and what they don't
The authoritative public reference is the Indian Food Composition Tables 2017 (IFCT 2017) from ICMR's National Institute of Nutrition, Hyderabad. It is a serious piece of work: according to its preface, it reports 151 food components for 528 key foods, using regional composite samples averaged across six geographical regions. But the same preface is explicit about its scope. Except for eggs, all the data are for foods in raw form, and the tables do not claim to list every food eaten in India; for the diverse regional foods of the country, the preface says the compositional data are still "inconsistent and fragmentary".
That is the right design for its purpose, which is assessing nutrient intake across a population. It is not a ready answer for a product that needs per-portion protein, carbohydrate and fat for an arbitrary home-cooked dish. Mixed preparations, regional recipes, and household differences in cooking method and portion size all have to be modelled on top of raw-ingredient data.
There is early work on that layer. In 2024, researchers published an open-access Indian Nutrient Databank (INDB) in Current Developments in Nutrition, combining 1,095 individual food items, drawn mainly from IFCT 2017, with 1,014 commonly consumed recipes whose ingredients are matched to those items. The authors open by noting that an open-access, comprehensive nutrient database had not been available in India. A shared recipe layer like this is a real step forward. But a standard recipe is still one version of a dish, and much of the hard part sits in how far a real plate drifts from that version.
A data-infrastructure problem, not a labelling problem
Nutrition and calorie-tracking apps, including NYUS in its early stage, tend to model the packaged half of the plate well and the unpackaged half badly, because the packaged half is where clean data already exists. The result is a category of products that works reasonably well for a diet built from packs and restaurant orders, and gets much less reliable the moment the input is a home-cooked Indian meal.
The gap is easy to miss on a nutrition-facts screen, but it matters to anyone building on Indian food data: meal-kit services estimating macros for a recipe, cafeteria vendors trying to label a thali, wellness platforms scoring "healthy eating" from self-reported logs, or any AI system asked to reason about what an Indian user actually ate. Each of these either inherits the same blind spot or builds its own workaround, often in isolation and often without publishing the method.
Restaurants show the same tension from the other side. The 2020 regulations require food service businesses with a central licence or outlets in 10 or more locations to show calories per serving on their menus, and they allow a 25 percent deviation in the nutritional information those businesses provide on request. That tolerance is a fair acknowledgement that a cooked dish is a range, not a fixed number.
What building this data actually requires
Three problems have to be solved together:
- Dish-level decomposition. A named dish such as dal, sabzi or biryani is not one thing. It is a recipe with variable ingredient ratios. Useful macro data has to model the dish at ingredient level and let portion and preparation move the estimate, rather than assigning one fixed number to a name. Our piece on why "2 rotis" isn't one number shows how much size alone can shift the answer.
- Regional and household variance. The same dish name can carry meaningfully different macros across regions, and across households in the same city, driven by oil quantity, protein source and method. A model that assumes one canonical "dal" is already wrong for many of the people eating it.
- Continuous correction, not a static table. Reference tables are revised periodically; IFCT 2017 followed earlier Indian editions published in 1937, 1951, 1971 and 1989. A workable product layer also needs a feedback loop, where logged meals are checked and corrected over time, rather than a one-time lookup treated as permanent truth.
We have been building exactly this as the nutrition layer under NYUS, an India-first fitness and nutrition app currently in early access. It now has 8,900+ foods with macros, from Indian staples to global foods. The honest lesson is that the hard part was never the app screens. It was building, and continually correcting, the unpackaged-food data underneath them. For anyone logging meals today, the practical workarounds are the same ones we lean on: start from approximate values for common Indian foods, and use consistent portion cues such as the hand-portion method rather than chasing false precision.
The opportunity for the industry
None of this argues that any single company should own India's unpackaged-food data. If anything, the opposite. Packaged food benefits from shared standards, a common label format and common reference values, that let every downstream product build on the same ground truth instead of re-deriving it. Unpackaged Indian food has the raw-ingredient foundation in IFCT and the beginnings of an open recipe layer in projects like INDB, but no shared standard yet for how a dish, a portion and a household's way of cooking it turn into a number. Until that exists, nutrition-tech, food-service and AI-diet products serving Indian users will keep solving the same problem separately, and arriving at slightly different, mostly unverified answers each time.
Disclosure: the author builds NYUS. This piece describes a data problem observed while building NYUS's food-mapping system; NYUS is not the subject of the argument above.