Health and Claro, running on demo data

Food, habits and training, with a camera that counts

What makes it different

• Up to four photos, one meal. A close-up and a wider shot sharpen the portion estimate. The model is told twice that the photos are one meal, so two angles never count as two servings.
• The app reshapes itself around conditions. Lactose intolerance adds lactose in grams per meal against 6 g and 12 g limits, with a lactase check-off. Gut conditions add a 1–5 “how did it sit” rating for each meal and a monthly trigger analysis. Sam has none; Settings shows what each one would change.
• An empty meal can mean something. “Didn’t eat this meal” records a real skipped meal, so a gap in the diary isn’t mistaken for a day you barely ate and felt fine.
• A workout program that runs itself. Rotating workouts, a built-in logger with a rest timer and last time's weights, phases that change the effort target, and a progress chart for every lift.
• A household leaderboard for steps, sleep, getting up and public habits.

The AI inside, and how the models were tested

• Meal estimates: Gemini 3.8 Flash, picked Oct 20 after two tests. First, each candidate estimated 20 plates from Nutrition5k, a research set of real cafeteria meals whose calories are known, and was scored on calorie error, speed and failed calls. Second, the same leftover-pizza photo was run repeatedly to check that the answer holds still. Gemini 3.5 Flash Lite, the cheaper model in use until then, said a quarter, three quarters and seven eighths had been eaten on different runs (500 to 3,500 kcal). 3.8 Flash said three quarters every time, for about a cent more per estimate.
• Fallback: Gemini 3.5 Flash Lite. In the same 20-plate test the previous fallback, Qwen3.8 Flash, was rate-limited on 9 of 14 calls and took about 54 seconds on the rest. Lite answered every call in about 5 seconds with about the same calorie error.
• Monthly trigger analysis: Claude Sonnet 5.5. Chosen Sept 28 from published reasoning scores rather than a test of our own: within a few points of Opus 5.5 at half the price. It runs twelve times a year at 10 to 15 cents a run.
• Models are a setting on the server, so they can be swapped without an app update. In this demo the AI answers are pre-written; everything the server does with them is real.

Try this

1. ✓On Food, tap the camera on Dinner, choose Photo library, pick both taco photos and estimate. The whole-plate photo finds a third taco, still counted as one meal.
2. ✓Estimate the taco close-up alone and compare: a guessed two tacos, lower confidence.
3. ✓Log the two-photo estimate to Dinner and watch the day's totals update.
4. ✓Mark Snacks as Didn’t eat this meal.
5. ✓Add a glass of water.
6. ✓On Check-in, tick off “Stretch 10 minutes”.
7. ✓Open Report and switch between Daily, Weekly, Monthly and Overall.
8. ✓See where Sam stands on the household Board.
9. ✓On Training, see Sam's four-day program, which workout is next, and this week's strength days.
10. ✓Tap Progress on the plan to chart a lift and see personal records.
11. ✓In Settings, turn conditions on and off and read what each one changes.

Everyday Mexican Spanish for a returning beginner

What makes it different

• FSRS review timing. The same scheduler serious flashcard apps use. After 100 delayed reviews Claro fits one personal multiplier to how you actually remember.
• “Say it and check.” Speech is recognized as Spanish (Mexico) on the phone and compared word by word with the sentence. It measures whether you were understood, not your accent, and nothing is recorded.
• A writing checker that knows beginner slips. Soy 26 años, yo gusto, gracias para, a missing ¿ or accent. Slip categories feed a “What to work on” card, never your own text.
• Practice anywhere. The 🎧🎤 pill says whether you can listen or speak right now. Turn speaking off and speaking prompts become writing prompts.
• Mexican usage, explained. What ahorita really means, tú versus usted at work, and why ¿me pasas la sal? sounds natural.

The AI inside, and how the models were tested

• Answer translator: Claude Haiku 5.5, picked Oct 28. The word-by-word check shows where a typed answer differs from the model sentence, but only a language model can say what the learner's own Spanish actually means. Ten models were given the same beginner answers, mistakes included, and compared on how well they read them. Haiku read them best, at about $0.00025 an answer. Gemini 3.1 Flash Lite, at the same price, is the fallback.
• It only ever shows a translation and up to three notes. It never changes a rating or a score, it is sent only the prompt, your answer and the model answers, and it goes quiet offline. It is off in this demo.
• No AI where a rule does better. Review timing is the FSRS algorithm, the writing checker is hand-written rules for known slips, and “Say it and check” uses the phone's own Spanish (Mexico) speech recognizer.

Try this

1. 1On Today, tap Start to recall the phrases due now, then rate how well you knew each one. The next review date comes from FSRS.
2. 2Open What to work on: the slips Sam keeps making, such as soy 26 años, with a short practice.
3. 3On Practice, try Verb forms or the Sentence lab, or type an answer and use Check my writing.
4. 4Tap the 🎧🎤 pill and turn speaking off: speaking prompts become writing prompts.
5. 5Search Phrases and star a few you want to keep.
6. 6On Progress, see what is staying with Sam after a delay, and tap See my week.