Private beta, not a black box launch.
Darshn is for people who want nutrition logging that can be inspected before it becomes part of a health record.
iOS-first private beta · Summer 2026
The AI that shows its work before it touches your health record.
Darshn turns a meal photo into a nutrition estimate you can inspect, question, and correct. What it detected, what it suspects, what it can't know — all on the table. Nothing is saved until you approve it.
Free during beta
No entry saved without your sign-off
Android & web planned after early feedback
The loop
Trust is the mechanism, not the tagline.
Most food loggers hand you a number and hope you don't ask questions. Darshn is built around the questions.
01
Scan
Point the camera at a meal, a menu, or a label.
02
Explain
Darshn separates what it detected, what it suspects, and what it can't know from one photo — with a plain-language reasoning summary.
03
Correct
Wrong protein? Bigger portion? Say so. The estimate updates without a rescan, and your correction counts as evidence.
04
Save
Only what you approve becomes part of your health record. Nothing is logged silently.
05
Adapt
Training and nutrition guidance adjusts from confirmed entries — not from guesses you never saw.
- StatusiOS-first private beta planned for summer 2026.
- RecordsAI output is a proposal until you review and approve it.
- PrivacyCoaching context is designed to stay local where possible; photo analysis uses secured backend vision models.
Demo 01 — Food scan
Audit the estimate. Right now.
This is the review surface Darshn puts between the AI and your record: evidence tiers, a reasoning summary, and corrections that update the numbers. Tap a scenario and watch the estimate — and its confidence — change.
Preset scenarios are scripted for speed. Ask your own question below and it's answered live by a dedicated Gemma model — no Gemini, no other provider mixed in.
Evidence overlay · demo photo
Explain this estimate
Confidence shown, not hidden.620calories
42gprotein
62gcarbs
18gfat
78%confidence
Detected
- Grilled chicken
- White rice
- Broccoli
Possible
- Teriyaki sauce
- Sesame oil
Unknown
- Exact portion size
Reasoning summary
Chicken appears grilled rather than fried. Rice occupies roughly one third of the bowl. Sauce or oil could shift calories by about ±120.
Visual cluesChicken, rice, broccoli, sauce sheen.
Macro assumptionsGrilled chicken serving plus rice base.
Uncertainty factorsSauce/oil and exact bowl depth.
Confidence levelMedium until portions are confirmed.
Demo only, not medical advice. Preset scenarios are scripted; freeform questions get a live, bounded answer. Questions are limited to this sample meal.
Actually there were two chicken breasts.
Updated
Calories +180 · Protein +34g · Confidence ↑ 92%
Calories +180 · Protein +34g · Confidence ↑ 92%
Demo 02 — Adaptive coaching
A plan that listens to your day.
Confirmed meals, logged workouts, and recovery signals reshape what Darshn asks of you next — and it always tells you why the plan changed. Tap a signal below.
Signal scenarios are scripted. Ask why below and a dedicated Gemma model answers live — in the shipping app, this coaching loop runs on-device.
Today's context
Lower body strength day
Four signals are already on the table before you train:
Protein behind target
Sleep was poor
Last leg session: 3 days ago
Mild knee soreness reported
Darshn's recommendation
Reduce volume, keep movement quality, and prioritize protein at dinner.
Moderateintensity
Reducedvolume
74%confidence
Recovery noteSleep and soreness suggest keeping the session controlled.
Nutrition noteProtein is behind target, so dinner should close the gap.
Reasoning summaryThe plan keeps lower body work, but reduces total stress because recovery signals are mixed.
Demo only, not medical advice. Preset scenarios are scripted; freeform questions get a live, bounded answer — not rehab, diagnosis, or injury treatment guidance.
Under the hood
One decision surface. Specialist signals underneath.
Visual evidence, nutrition estimates, training context, recovery context, and policy checks — resolved into one recommendation you can question.
Where the AI actually runs — no hand-waving
Food photo analysis runs on secured backend vision models, because that's what accuracy currently requires. The behavioral coaching loop runs on-device. In both cases the output is a proposal, not a verdict: you see the reasoning, you can push back, and nothing is saved to your record until you confirm it.
Private beta
Be one of the first to hold an AI accountable.
iOS-first private beta opens summer 2026 — free while in beta, with Android and web planned after early feedback. Bring your skepticism; the product is built for it.