NanoRhino is an AI nutrition coach you text. You start with a free Starting Plan — a calorie estimate that's a guide, not a rule. For the first 2 weeks you just log meals (a photo is enough) and weigh in twice a week; from that data we measure what your body really burns, and in week 3 you get your real plan, retuned every week after. When progress stalls, a human nutritionist looks at your data and gives you one specific fix. Your first 2 weeks are free; after that you pay $10 per pound you lose, capped at $500.

Why start with an estimate instead of a real number?

Every calorie calculator — ours included — is a formula built on population averages. For our members, formulas have run about 15% high on average, which is often the whole deficit. So we give you the formula's number as a starting reference, show you the math behind it, and then replace it with your measured burn: calories you eat plus the weight you lose (1 lb ≈ 3,500 calories), from two weeks of your own logs and weigh-ins.

Do I have to change how I eat in the first 2 weeks?

No. For the first 2 weeks, just eat as usual and log it — a photo or one line per meal — and weigh in Tuesday and Saturday mornings (you'll get a reminder). Your starting target is a guide, not a rule: getting close is great, but your normal eating is exactly what we need to measure your real burn.

What happens in week 3?

Once your scale has enough points (about three weigh-ins over two weeks), your coach calculates your real daily burn from your own logs and weigh-ins and sends your real plan: how your burn compares with the formula, and your new daily target. From then on the target is retuned every week as your body changes. If the data isn't there yet, your coach tells you exactly what's missing — usually one more weigh-in.

Is a real person involved, or is it all AI?

Day to day, your coach is AI — it answers in seconds, logs every meal, and checks in on you. When progress stalls, a human nutritionist reviews your meals, weigh-ins and habits to find the real cause and gives you one specific fix. NanoRhino is not medical care: we don't diagnose, prescribe, or replace your doctor.

How does pricing work?

Your first 2 weeks are free — pounds lost in them are never billed. After that you pay $10 per pound you lose, settled when a weigh-in crosses a 5-lb milestone, and your total is capped at $500. The same pound is never billed twice, and if the scale doesn't move, you pay nothing. No subscription, no upfront charge, no card to start.

I'm on a GLP-1 (Ozempic, Mounjaro, Wegovy, Zepbound) — does NanoRhino fit?

Yes. About 20% of our active users are on a GLP-1. NanoRhino is built to complement the medication, not duplicate it: the coach focuses on hitting protein (so you protect muscle on a suppressed appetite), staying hydrated, and adjusting calories as your weight comes down through dose changes. We don't medicate, prescribe, or give medical advice — your doctor stays in charge of dosing. We just help you eat well within whatever appetite you have on any given day.

Do I need an app? How do I start?

No app. In the US, text START to (915) 277-7888 — your coach asks a few quick questions and sends your Starting Plan by text. Outside the US, message us on WhatsApp from nanorhino.com/message.

How is NanoRhino different from apps like MyFitnessPal or Cal AI?

Apps like MyFitnessPal and Cal AI are trackers — they tell you the calories in your meal and stop there. NanoRhino logs meals too (send or photo a plate in chat), but the product is the coaching: it sets your targets, checks in weekly, explains why the scale moved, and adjusts your plan — and you only pay when you actually lose weight ($10 per pound, capped at $500). See the full breakdowns: NanoRhino vs MyFitnessPal, NanoRhino vs Cal AI, or browse all comparisons.

Is NanoRhino private and secure?

Yes. NanoRhino is a private, judgment-free space — your conversations live in your messaging app. There are no social feeds, no posting, and no sharing. Your health data is encrypted, never sold, never shared with third parties, and never used for advertising. You can request deletion of all your data at any time.

How do I unsubscribe?

Reply STOP at any time to unsubscribe. Reply HELP for assistance. You can re-enroll anytime by texting START again.

We Tested Our AI Calorie Estimator on Google's Nutrition5k. Portion Size Is Most of the Error.

An engineering note. We ran our production coach on a public dataset of weighed meals to see how far one photo can go.

TL;DR

Nutrition5k is a public dataset from Google Research (CVPR 2021). Each dish was weighed ingredient by ingredient, and the calories come from those weights. So every photo has a real answer.

We ran our production coach on 455 dishes from the official test split. One overhead photo, one message: "had this for lunch."

Photo only: 39.5% mean error (112 kcal on an average dish of 283 kcal). Only 29% of estimates were within ±20%.
The estimates were biased high by 39%. Small plates were the problem: under 100 g, the model guessed 1.76× the real weight.
Large plates were fine: for dishes over ~340 kcal, the median error was 28% and the bias was about zero.
Photo + the true total weight: 20.5% mean error, bias near 0%. Knowing the weight cut the error in half.

The rest of the error comes from calories per gram, and mostly from oil and dressing that the camera cannot see.

This is a different number from the 3.8% we published in July. That 3.8% is on foods with a knowable answer: a label, a number the user said, a simple food. This post is about the hard case: a mixed plate, one photo, nothing else.

Why we ran this

Our own regression gate uses real user conversations. That is good for catching regressions, but the "truth" for a photo of a home-cooked plate is a range, not a number. Nobody weighed it.

Nutrition5k fixes that. The dishes were made in a cafeteria, each ingredient was weighed on a scale, and calories were computed from USDA values. It is released under CC BY 4.0, so we can run it and publish the results.

How we ran it

Same code path as production. Each dish was a full coach turn: the same agent, the same model (Claude Sonnet 5.5), the same tools that log a meal for a real user. We read the logged calories from the meal file.
Cold start. A fresh workspace for each dish. No dish memory, no history, no profile to lean on.
Zero-shot. We did not train or tune anything on Nutrition5k.
Dishes: all test-split dishes with at least 30 kcal and an overhead RGB photo. That is 455 dishes. One dish produced no log, so 454 were scored.
Two arms:
- A: photo only. "had this for lunch."
- B: photo + weight. A random 200 of the same dishes, with "about N g total", where N is the true weight rounded to 10 g.

Arm B is not a product feature. Real users almost never tell us the total weight of a mixed plate. Arm B is a controlled experiment. With the weight known, the model still has to decide what is on the plate, how the weight splits between foods, and how many calories each gram has. So comparing A and B tells us how much of the error comes from portion size and how much comes from calories per gram.

The whole run cost $114 in API usage.

Results

	A: photo only (n=454)	B: photo + true weight (n=199)
Mean absolute error / mean true kcal	39.5% (112 kcal)	20.5% (58 kcal)
Median error	37.0%	15.7%
Within ±20%	29%	58%
Mean bias	+39.4%	−0.2%
Median weight error	38.9%	1.3%
Median calories-per-gram error	15.8%	16.0%

On the 199 dishes that ran in both arms, the median error went from 38.2% to 15.7%. 142 of 199 dishes got better.

Look at the last two rows. Calories-per-gram error is about 16% in both arms. Giving the weight did not change it. The weight error went from 39% to 1%. Almost all of the improvement is the portion.

In arm A, we also split the variance of the log error into a weight part and a calories-per-gram part: weight is 57% and calories per gram is 43%. Weight is the larger part. It is also the only part with a strong bias in one direction.

Finding 1: the model pulls every portion toward "a normal meal"

When we plot estimated weight against true weight (on a log scale), the slope is 0.73, not 1. Small portions are guessed too big and big portions are guessed too small. The two lines cross at about 470 g.

True weight	Estimated / true (median)	Dishes
under 100 g	1.76×	107
100–200 g	1.44×	136
200–350 g	1.31×	142
over 350 g	0.86×	69

Our reading: the model has a prior for "one serving." When it cannot judge size from the photo, it moves toward that prior.

This matters a lot on Nutrition5k because the plates are small. The median dish is 185 g. Many are sample plates from a cafeteria line, not full meals. That is exactly where the over-estimate is largest.

By calories, the same pattern:

True calories	Median error	Bias
30–146 kcal	48.6%	+85.3%
147–340 kcal	39.9%	+36.0%
343–1050 kcal	28.4%	−2.9%

For a plate the size of a real lunch or dinner, the photo-only estimate was not biased. It was still noisy (28% median error), but it was not high on average.

When we gave the true weight (arm B), the small-plate bias dropped from +85% to +11%, and the median error in every size group was 15–19%.

Finding 2: the oil you cannot see

Calories per gram were close to unbiased for dishes without added fat (median ratio 0.99×, 225 dishes).

For dishes whose weighed ingredients include oil, butter, dressing, mayo or pesto, the model's calories per gram were 0.89× the truth (229 dishes). The model listed a fat item for only 58 of those 229. Oil soaks into food. In a photo it looks like nothing.

This is the main source of the 16% calories-per-gram error. The model is under-counting what it cannot see, not mistaking one food for another.

Finding 3: some of the error is in the dataset

We looked at the 12 worst dishes by hand. In about 4 of them, the label is clearly missing food that is in the photo. For example, the photo shows pizza or a Caesar salad, and the label lists a few grams of broccoli. In 2 others, the label includes things the camera cannot see, like 45 g of olive oil or chicken hidden under other food. A few almond weights also look too high for what is on the plate. Only 3–4 of the 12 are clear portion mistakes by our model.

The worst 10% of dishes carry 29% of the total absolute error. So the average is inflated somewhat by label noise. But the two patterns above (small plates guessed too big, hidden oil missed) are ours. They hold across hundreds of dishes, not just the outliers.

How this compares

The Nutrition5k paper reports its own baselines (Table 3). They are not the same setup as ours, so read this as context, not a ranking:

Method	Calorie error (MAE / mean)
Always guess the dataset mean	60.2%
2D model trained on Nutrition5k, single photo	26.1%
Model with depth camera	16.5–18.8%
Trained model, true weight given, predicts calories per gram	9.5%
Our coach, photo only, zero-shot	39.5%
Our coach, photo + true weight, zero-shot	20.5%

The differences: their models were trained on the same cafeteria, the same camera rig and the same kinds of dishes. Ours has never seen this data. We also dropped dishes under 30 kcal and dishes without an overhead photo. A specialized model trained in-distribution should win on its own test set, and it does.

One more number from the paper: when people estimated the weight of food by eye, dietitians were off by 41% and non-experts by 53% (a small test on 10 plates). Our model's median weight error was 39%. Judging weight from a photo is hard for people too.

What this means for the product

A better vision model would help a little. Better information about the portion helps much more. In this test, knowing the weight halved the error and removed the bias. So the coach is built to get portion information wherever it can:

Labels and stated numbers win. If you send a nutrition label or say "it's 350 calories," that number is logged as-is. There is no photo estimate to be wrong.
Ask about the portion when it was a guess. When the portion was assumed and your history says you often eat part of a dish, the coach asks one short question: "all of it, or part?" Small portions are where photo estimates go wrong the most.
Remember your usual dishes. If you log the same breakfast often, the coach keeps its calories per gram and your usual amount. A label or a correction from you becomes the reference next time, instead of a new guess from a new photo. For you as a user, the practical advice is short: if a plate is small or the food was cooked in oil, add a few words. "Half a bowl," "small plate," "fried in butter." For cooked dishes, the fat is often the biggest item the photo cannot show: "cooked in a tablespoon of olive oil" is about 120 kcal. That one line does more for accuracy than a better photo.

Limits

Cafeteria food, not home food. Nutrition5k dishes are cafeteria food, shot from directly above with a fixed camera rig. Real user photos are taken at an angle, in different light, often with packaging or a hand for scale.
Plate sizes are not typical. The median dish is 185 g. Real meals are usually larger, where our bias was close to zero. We cannot claim a single "real-world" error number from this test.
One model, one run per dish. We did not test other models or repeat runs to measure spread.
The label noise goes both ways. Some dishes look worse than they are because the label is missing food. We did not remove those dishes; all 454 scored dishes are in the numbers above.

Data and credit

Nutrition5k: Thames et al., "Nutrition5k: Towards Automatic Nutritional Understanding of Generic Food," CVPR 2021 (arXiv 2103.03375). Dataset released by Google Research under CC BY 4.0.

This is an engineering report from the team that builds NanoRhino. It is not medical advice. I am not a clinician.

Try it

Send a photo of your meal and add one line about the portion. Start messaging your coach on WhatsApp or SMS. No app, no signup. Or start with your numbers on our free Accurate TDEE Calculator.

Built with care by Link Heart Limited in Houston, Texas.