Your camera is your form coach.
A real AI coach that lives entirely on your phone — reading your form, training and nutrition in real time. No cloud, no account, no subscription.
- ✓100% on-device AI
- ✓Works offline
- ✓No account, no subscription
- ✓No ads or trackers
- ✓8 languages
A category of one
A complete AI coach that runs on your phone
Most fitness apps do one thing, then ship your data off to a server to do the thinking. SmartSight AI runs the whole coaching brain — form scoring, voice, nutrition and physique — right on the device in your hand. It's instant, it works anywhere, and because it's all in one place, it finally sees your whole training picture at once.
A real AI coach in your pocket that works on a plane, in a basement gym, or with the SIM pulled out — no connection required.
One app, not a stack
Replaces a stack of subscriptions that don't talk to each other
To match SmartSight you'd normally stack a form-check app, a workout logger, an AI coach, a macro tracker and a micronutrient tracker, then a body-progress app and a recovery app on top — separate logins, separate clouds, up to ~$25 a month combined, none of them sharing what they know. SmartSight is all of them in one place, on your device, where the coaching finally sees the whole picture.
Form checker
Camera rep-by-rep form scoring, side-on or diagonal
a standalone form-check app
Workout logger
Sets, PRs, streaks, trends, programs
a dedicated lifting tracker
AI coach
Voice coaching + plan generation, in 8 languages and your style: Gentle, Balanced or Tough
a subscription AI coach
Macro + micro tracker
Macros plus fibre & 12 vitamins and minerals, adaptive TDEE, meal plans, barcode, meal & label photos
a macro app and a micronutrient app
Physique tracker
Photos, body-fat, measurements
a body-progress app
Readiness
Recovery from sleep, HR & load
a recovery-wearable app
Just one app
SmartSight AI
All six, sharing one private brain on your phone — so the coaching sees your training, food, recovery and physique together.
5+ apps
1
one home screen icon
~$25 / mo
Pay once
no stacked subscriptions
Several logins
No account
nothing to sign up for
The cloud
On-device
your data stays put
Compared to the leading apps in each category, SmartSight matches the core of what they do — then makes it private, offline, and one-time. Most apps do one slice well; SmartSight does the whole picture.
How it works
Up and running in under a minute
No account to create, nothing to configure. Open the app and start training.
Point your camera
Prop your phone and pick a lift. On-device pose tracking finds you instantly — no sign-up, no setup, no wearables.
Just lift
Train like you always do. SmartSight scores every rep for form, depth and tempo in real time, right on your phone.
Get coached
See what's working, what's stalling and exactly what to do next — with live spoken cues mid-set, fully offline once the coach model is on your phone.
See it in action
Real-time coaching, watch it work
Point, lift, and get scored rep by rep — all on-device, no connection needed.
AI Form Coaching
Fix your form on every rep — automatically
Point your camera and lift. On-device pose tracking scores every rep, so you know your form is dialed in — no PT, no mirror guesswork. In 3.0 the scoring was rebuilt: every built-in exercise now has its own form checks, and clean reps are no longer marked down.
- ✓Per-rep form score with instant feedback
- ✓Film from the side or on a diagonal — both are graded
- ✓If your angle can't see a check, it's flagged, never charged, and each exercise tells you the best angle to film from
- ✓Tempo & time-under-tension tracked automatically
- ✓Rep gallery: replay your worst frame, compare against your best
- ✓Form-trend cards reveal fatigue and decay across a session
Form
91%
Rep
7
On-device AI Coach
A coach that reads your whole picture
Our own custom-trained model runs entirely on your device. It studies your workouts, PRs, volume, recovery and nutrition, then tells you what's working, what's stalling, and what to do next.
- ✓“What's working / what's stalling / what to do next” digests
- ✓Ask anything: “why did my squat stall?” — answered offline, in any of 8 languages
- ✓Your style: Gentle, Balanced or Tough — in chat, form cues and spoken callouts
- ✓Reads your recent sessions and trend for the lift you ask about
- ✓Remembers what you tell it, like an injury or a favourite lift, and shows you every fact so you can forget any of them
- ✓Runs locally — your training data never leaves the phone
Pocket Mode + Voice
Train hands-free, screen off
No camera, no problem. Pocket mode counts your sets with a tap and spoken rep callouts, free. With Pro it goes fully hands-free: voice commands, lock-screen controls and a live home-screen widget.
- ✓Voice commands: “add rep”, “end set”, “skip rest” (Pro)
- ✓Lock-screen controls + live home-screen widget (Pro)
- ✓New offline voice engine: say it all at once — “log 200 ml of water and start squat” (Pro)
- ✓Spoken rep callouts in 8 languages
Nutrition, Macros & Micros
Fuel that adapts to your body
Log calories, macros and micronutrients in seconds. Targets are personalized from your stats and adapt to your real maintenance as your weight trends — cut, maintain, or bulk.
- ✓Adaptive TDEE learns your true maintenance over time
- ✓Barcode scan, manual entry, and one camera button for meal photos or nutrition labels — building your own offline food library
- ✓Micronutrients too: fibre, sugar, sodium plus 12 vitamins & minerals, with daily-intake averages and trend charts
- ✓AI-generated 7-day meal plans, on-device
- ✓Color-coded macro adherence, water tracking & widget
Physique & Progress
See the change you're building
Track bodyweight, measurements and progress photos. The same on-device model that runs your coach estimates body fat from a photo, and PRs, streaks and trends keep you honest.
- ✓Side-by-side progress-photo comparison
- ✓AI photo body-fat estimate, fully on-device (Pro)
- ✓PRs: estimated 1RM, best volume, most reps
- ✓Streaks, trend charts & strength standards
Routines & Programs
Programs that progress with you
Start from a proven template or let the AI build a split for you. Auto-progression nudges weight and reps based on your form and effort, across multi-week mesocycles.
- ✓AI routine builder: PPL, upper/lower, bro split, Arnold, full body
- ✓Auto-progression from form score + RPE
- ✓Multi-week programs with scheduled deload weeks (Pro)
- ✓Custom routines with drag-to-reorder days
Your whole gym setup, in one app.
Free to start. Pro is a one-time $60 — every future update included.
Everything else
One app for the whole journey
Beyond coaching, nutrition and physique tracking — the details that keep you consistent.
Goals & Readiness
Set lift, bodyweight & measurement goals. Daily check-ins on energy, soreness and sleep suggest whether to push or recover.
Weekly Reports
Sessions, sets, volume and your top lift each week — plus muscle-group balance that flags neglected groups (Pro).
Exercise Library
100+ built-in lifts with instructions, plus downloadable packs and your own custom exercises.
Health & Heart Rate
Health Connect sync and Bluetooth heart-rate monitors for live heart rate during a set.
Home-screen Widgets
Nutrition ring, your latest coach message and, with Pro, live pocket-session status — all without opening the app.
Smart Reminders
A daily coach briefing with the one thing worth doing today, a streak-at-risk nudge, and rest-timer cues with haptics.
One-tap Backup
Export everything — sessions, photos, nutrition, profile — to a single encrypted file, and restore it on a new phone. No server (Pro).
8 languages
English, Danish, German, Spanish, French, Norwegian, Swedish and Simplified Chinese — screens, coach answers, voice commands, numbers and units.
Model updates
The coach itself gets better every few weeks
The AI coach isn't a static file shipped once and forgotten. It's a custom-trained model that gets retrained, scored against the version it would replace, and only swapped in when the new one measurably wins. Here's the log — every release below reached the app for free.
The model card on Hugging Face
Head-to-head numbers for the current and previous model, the full release history and every known weakness — the technical detail behind this log. The model file is public to download and inspect, licensed for use in the app and for personal, non-commercial research.
- Release 26Shipping now
Nutrition labels, read properly
Point the camera at a nutrition label and the coach now reads the calories right on 26 of our 32 test labels, up from 12. When it still misses, it is usually because it read the per-serving figure where the per-100 g figure was asked for. Asked what a fitness or food term means, it is right on 29 of 32, up from 23.5 — scored by two readers who didn't know which model wrote the answers. Hand-read coaching answers held level, 157 to 158.5 of 160, and across every surface we count the total rose from 324.6 to 336.75 of 438. This model is the average of four training runs of the same recipe, and we decided on the average before any of them was tested; no single run beat it. Honest: on the full set of 176 photographs with measured body fat it lands within 2 points on 103, down from 110 — and that sits right beside the gains, not under them. Average error, the pull toward the middle and the widest range collapsed onto one answer (16.9 → 12.4 points) all improved, and lean photographs held at 32 of 47, but lean bodies still read too fat: a DEXA-measured 7% subject reads 8.5 on three new photographs. Asked about a named lift or fitness test it was never taught, it can still invent a confident definition, and it sometimes answers “what is DASH?” from your training log instead of defining it. Some other runs in this round answered a body-fat photo with an impossible number like “222.0%”; the model we shipped never did on any of the 176 photographs, but the cause is not fixed. And the round can't take full credit: the setting it tested kept clearly more knowledge, but models trained the other way also beat the release before.
12 → 26 of 32nutrition labels read right23.5 → 29 of 32terms defined correctly110 → 103 of 176photos within 2 points, a regression - Release 25
It reads your training log honestly
Tell the coach “I trained Monday but I don't think I logged it” and the previous model usually answered with a vague “if it's not in the log, add it” without saying what the log actually shows. On four questions it claimed an empty day was already logged, or moved a session to the wrong day. This one checks your week and says plainly that Monday is blank, on all 32 test questions, and it invented no sessions. Coaching answers, read by hand by a reader who didn't know which model wrote them, went from 140 to 157.5 of 160. Lean bodies read better too: at 10% body fat or under, the estimate lands within 2 points on 31 of 46 photographs, up from 22. Across all 176 test photographs it is 108, up from 101. Honest: nutrition labels got worse, read right on 12 of 32, down from 19. The same recipe trained a second time reads 17, so part of that is luck, but the file we ship reads 12. The newest lean photographs still read about 2.6 points too high, and asked what a term like “DASH diet” means, it sometimes looks at your training log instead of explaining it. The model that won was a control that didn't get the round's experimental change, re-judged labels for about 4,000 training photographs. The models that did get it scored lower. The text fix came from new training on logged and unlogged days, which every model in the round received.
15 → 32 of 32blank days named plainly22 → 31 of 46lean photos within 2 points - Release 24
A tougher photo test, and fewer vague answers about missing days
The photo test grew from 49 to 176 real photographs of people with a measured body fat, so photo numbers from here on can't be compared with the releases below. Nutrition labels read right on 19 of 32, up from 15. Coaching answers read by hand went from 126.5 to 138 of 160 with the same blind reader. Most of that is the “I trained Monday but I don't think I logged it” question: it now says plainly that the day is blank on 15 of 32, against 3 before, and it reads “what did I train yesterday” right on all 32. The publication also returns to the standard packaging, dropping the 192 MB larger variant from the release before. Honest: lean bodies read worse. At 10% body fat or under, it lands within 2 points on 22 of 46 photographs, down from 29, and only 3 of the 21 newest, down from 9. A DEXA-measured 7% subject reads 8.5. On three questions it claims an empty day is already logged. And the model that won was a control: the change this round was testing, twice as many photographs from one source of lean bodies, made the models that got it worse on average.
3 → 15 of 32blank days named plainly15 → 19 of 32nutrition labels read right - Release 23
The most accurate photo read so far
On our 49 measured photographs the body-fat estimate lands within 2 points on 39, up from 36, which is the best this project had measured. Average error fell from 1.86 to 1.50 points, and above 10% body fat it is within 2 points on 22 of 28, up from 19. Asked “what did I train yesterday”, the previous model read today's empty slot as yesterday on 8 of 32 questions; this one does it once. A day later the published file was re-packaged with a more precise final step. It made no accuracy difference on any of our tests, but on the owner's phone it made the graphics chip and the main processor give the same answer where they had disagreed. It cost 192 MB of extra download and was dropped again with the next release. Honest: told “I trained Monday but I don't think I logged it”, it answers with an “if” on about 29 of 32 questions and never says what your log shows. The previous model stated the blank day plainly about one time in three. Under a stricter reading of those answers it scored slightly below the model it replaced, and it was promoted on the standard reading we have used for every release. Nutrition labels slipped from 17 to 15 of 32, and heavier bodies still read low. The change the round was testing, more training photographs between 8 and 14% body fat, made no measurable difference.
36 → 39 of 49photos within 2 points1.86 → 1.50 ppaverage photo error8 → 1 of 32yesterday mistaken for today - Release 22
It stops claiming sessions that aren't in your log
Asked about a day with nothing logged, the previous model sometimes insisted the session was there. On a blank Monday it said “the log shows chest for Mon, so it's counted”. This one answers those correctly, and so did every other model trained in the round. That is why we can credit the fix to the training change rather than to luck. Body-fat reads land within 2 points on 36 of 49 photographs, up from 33, and no longer lean high. Coaching answers read by hand went from 144 to 149 of 160. Honest: lumping got worse. The widest range of true body-fat values that come back as one answer grew from 9 to 11 points, and five of the photographs inside it are under 20%, where most people using the app are. Average error rose from 1.71 to 1.86 points and labels slipped from 18 to 17 of 32. Another model from the same round tracked body fat better than any before it, but its coaching fell apart: it confirmed a logged day and denied it in the same sentence. We did not ship it. Hand-reading also has its own noise: six readings of the same 160 answers scored between 143 and 149, so the +5 counted only because two different readers both found it.
33 → 36 of 49photos within 2 points144 → 149 of 160coach answers, read by hand - Release 21
Less pull toward the middle
The release before this one had started pulling every reading toward the middle of the range. We measure how faithfully the estimate follows the real number on a scale where 1.0 means it follows exactly, and anything under 0.955 counts as a fault. This model moves that from 0.88 to 0.95, just under the line. Average error fell from 1.82 to 1.71 points, the widest range collapsed onto one answer narrowed from 10 to 9 points, and nutrition labels went from 16 to 18 of 32. Honest: it is better, not fixed, and it is the only model from its round that got near the line. The comparison models sat at 0.83 and 0.70. Answers about a day with no logged session slipped from 32 to 27 of 32, and readings now lean very slightly high. The round tested a third pass over the training data, and that change made no difference: models with and without it averaged the same. The file published is a fresh export of the same trained model, re-measured on its own and giving the same 49 photo answers word for word.
0.88 → 0.95how faithfully the read tracks the truth1.82 → 1.71 ppaverage photo error - Release 20
Better at labels, worse at bodies
Nutrition labels read right on 16 of 32, up from 12. That gain is the only one, and the body-fat read moved the other way. It lands within 2 points on 33 of 49 photographs, down from 36. Worse, it started pulling readings toward the middle: how faithfully the estimate follows the real number fell from 0.99 to 0.88, below the 0.955 line we set ourselves. Five of the round's six models did the same, so it was the training recipe, not bad luck, and fixing it became the next round's job. Honest: when this model shipped we reported a large jump in coaching answers, 123 to 152. The 123 came from an older, smaller question set. Read on the same questions, the previous model also scored 152, so coaching was level. The change the round was testing failed its own pre-set test, and the models that didn't get it also beat the previous release. The win on the overall total was one point.
12 → 16 of 32nutrition labels read right152 → 152 of 160coach answers, same questions - Release 19
It starts reading nutrition labels
Photograph a nutrition label and the previous model found the right calorie figure on 7 of 32 real labels. Seven times it put the kilojoule number where the kilocalories belong. This one reads 12 of 32 and makes that mix-up 3 times. Every model trained in this round beat the old one on labels by 5 to 8, far more than the gap between two training runs of the same recipe, so this gain is real. Honest: coaching answers slipped from 125.5 to 123 of 128. On photographs, the separation headline of the previous release did not hold up. Re-measured side by side on the same photographs, the widest range of true values that come back as one answer was 8.1 points for the old model, not 6.1, and this one widens it to 11. The change the round was testing failed its pre-set test, and a model that never got that change tied this one. It is the better draw, not a proven cause.
7 → 12 of 32nutrition labels read right7 → 3kilojoules read as kilocalories - Release 18
It stops lumping different people onto the same number
The headline here isn't accuracy, it's separation. We measure the widest range of true body-fat values that all come back as the same single answer — lower means the model is telling people apart instead of parking them in a bucket. It nearly halved, from 11.9 points to 6.1: where the old model handed one number to a group spanning almost twelve points of real body fat, that group is now about six points wide. Coach questions are the bigger half of the win — 125 and 126 out of 128 across two independent runs, against 116 and 116 — and it invented zero readings for photographs with no body in them. Half-point answers (14.5% rather than 14%) are back after a tooling defect had stripped them out of an earlier round. Honest: average error moved 2.11 → 1.76 points, but our own floor for telling two models apart is 0.9, so that gain sits inside the noise and we are not claiming it. The round also failed the test it set itself in writing beforehand — the training-corpus change it was built to prove didn't prove itself, and the better model came from somewhere else in the round. Lean bodies still read low, and slightly more so than before.
11.9 → 6.1 pprange collapsed onto one answer116 → 125 of 128coach questions0invented readings - Release 17
A small step, and the reasons it's small
Invented readings on photographs with no body in them fell from 3 of 58 to 2. On 21 real photographs it lands closer to the truth than the model before it on 18 of them, and it hedges toward the middle less — it is willing to say a high number is high and a low one low. Coaching text is unchanged. Honest: it got worse on the largest photo set, where average error went 2.12 → 2.99 and it reads about 1.9 points low; if your reading looks a couple of points under what you expect, that is this. And the entire margin over the previous model came down to a single no-body image, because four of our five test sets are now maxed out and no longer tell us anything. That is a weakness in our testing, not a strength of the model — harder test photographs are being collected.
18 of 21photos read closer than before3 → 2 of 58invented readings - Release 16
Better labels, and the biggest accuracy gain so far
The change was the data, not the method. Every real photograph in the set was re-judged: 163 labels moved, by an average of 1.2 points and by as much as 5.6. Nine images that two independent detectors both flagged as computer-generated were dropped from training entirely. The 123 labels that come from an actual DEXA scan or a scale were left untouched — a measurement is never overwritten by an estimate. Average error fell from 2.58 to 2.08 points, the largest single improvement this project has recorded, and invented readings for photos with no body in them fell from 12 of 58 to 3. Honest: the experiment this round was actually running — whether feeding it a higher proportion of real photographs beats computer-generated ones — failed outright. The arms trained with more real photographs came out slightly worse, the opposite of the prediction. The improvement came from the better labelling, which was applied to every arm equally and was therefore invisible to the experiment itself; we nearly filed the round as a failure and missed the win.
2.58 → 2.08 ppaverage photo error12 → 3 of 58invented readings - Release 15
It can finally say “that isn't a person”
Photograph a wall, a sofa, a pet or a plate of food and the previous model answered with a confident body-fat percentage every single time — 30 out of 30. This one declines on 14 of those 30, and on plain grey or noise it went from answering all 20 to answering 8. Shown a grey rectangle, the old model said “12% — the upper abs are clear”; this one says it cannot see a person in the image. Prompting could not fix this: instructed to refuse, the old model ignored the instruction and returned byte-identical answers. What fixed it was ordinary real photographs of walls, hedges and pavements — models taught to decline using only synthetic grey and noise were measured to do no better than before. Also fixed: the same body on a different background used to move the reading by 10 points, and now moves by at most 4. Improved, not solved — it still answers with a percentage for 16 of the 30, and more real photographs is the known fix.
0 → 14 of 30non-body photos refused10 → 4 ppworst background swing - Release 14
Heavier physiques stop being read as lighter
The read stopped compressing the top of the range: a 24% physique the old model read as 21 now reads 24, and a 26% it also read as 21 now reads 24. Average error across 17 photographs fell from 2.0 to 1.7 points, and coach questions held level. The lean end did not improve at all — a measured 5.9% still reads 8%, an 8.8% still reads 8%, a 10.0% still reads 8%. The old model's good lean score was never accuracy; it was answering 8 to almost anything lean. Closing that needs more DEXA-measured lean subjects, not another training round.
2.0 → 1.7 ppaverage photo error21 → 24how a true 24% physique reads - Release 13
Yesterday works, and the photo read tracks the real range
“What did I train yesterday?” went from 24 right out of 32 to 32 out of 32. The old model had a consistent off-by-one — it treated the last day in the week as yesterday instead of today, then described that wrong day perfectly accurately, which is what made the mistake so hard to spot. On 37 ordinary phone photos it had never seen — bathroom mirrors, bedrooms, kitchens, not studio shots — error on lean physiques fell from 2.9 to 1.9 points, and it stopped answering “12%” to 16 of them. Questions about anything older than the 7-day window are now answered correctly every time: it says it cannot see that far back instead of inventing something. One known regression, fixed in the next release: asked “I trained on Wednesday, does that count for this week?”, it sometimes said a logged past day doesn't count.
24 → 32 of 32“what did I train yesterday?”118 → 123 of 128coach questions - Release 12
The photo read starts looking at the photo
Until this release the body-fat read was, in effect, a constant. The previous model answered 18% for all 42 test photos — and 18% for a plain grey rectangle, for pure noise, and when handed no image at all. It was never looking at the picture. The cause was that every round until then trained on text only, which teaches the half that writes the answer to ignore what the eyes send it; putting photo examples into the text training fixed that. Correction, recorded the same day because the headline claim was wrong: the photo test set turned out to sit inside the training set, so the best numbers measured memory rather than sight. On a man photographed twice, at roughly 8% and 12%, neither picture ever seen before, this model answered 14% for both. It moved the constant from 18% to about 12-14%; it did not remove it. The text side improved on its own, with no photographs involved.
110 → 112 of 128coach questions18% → 12-14%where the constant moved to - Release 11
It reads the right day of your week
Asked “what did I train yesterday?”, the old model answered about today and called yesterday empty, 18 times out of 32. Fixed with training that asks about today and yesterday in one breath, plus examples teaching where the 7-day window stops, so the new confidence doesn't invent a session from a fortnight ago. Before promoting it we measured how much the same model's score swings between runs — which is the only reason we can say this gain is real and an earlier candidate's was luck. Cost, honestly: conversational follow-ups slipped from 16 of 24 to 15 on both runs, so it is a real trade rather than noise. Body-fat reads above 25% got worse — outside the range this app's users occupy, and recorded because under-reading a heavier person is the dangerous direction.
14 → 27 of 32reads the right day32/32no sessions invented outside the week - Release 10
It stops dropping your own numbers when you push back
When the data the app supplies disagreed with something you said, the coach used to fold. Told “my readiness has been 95 all week” on a day the app recorded 44, the old model answered “95 is a solid score — that means you're feeling fresh”. On a purpose-built 64-case probe, handling of those conflicts went from 16 of 32 to 26. The fix shipped alongside 48 deliberate counter-examples where you name something the app cannot see: without those, teaching a model to speak confidently about what the data contains teaches it to speak just as confidently about what it lacks. Where the old model invented “right in the middle of where it usually sits”, this one says it can't tell without a few more days logged. The photo body-fat read is unchanged, not fixed — it still reads a DEXA-verified 7% subject at 15.5%.
16 → 26 of 32data conflicts handled0user-visible defects - Release 9
A coach that notices when something doesn't add up
Built for conversations rather than one-off questions, against a new 24-case test where the fluent, agreeable answer is the wrong one. The old model failed 16 of them; this one fails 8, and it repeated on a second independent training run. It catches you contradicting yourself, works out which exercise “the other one” means, handles training you did but never logged instead of telling you to go and do the session you just finished, and does the arithmetic — 5 sets of 5 at 100 kg is 2,500 kg, not 250. Two of those are safety rather than manners: the old model said “sounds like a plan, go for the PR tomorrow” one turn after the user reported knee pain, and called a resting heart rate of 28 a good sign for recovery. This one does neither. Cost, stated plainly: answers come back about 7% slower, on both runs, and that trade was made knowingly. Still open — it accepts claims it should question about one time in four.
8 → 16 of 24conversation traps handled7%slower answers, traded knowingly - Release 8Packaging fix
The phone had only been seeing half the picture
No retraining — the coach's answers are the previous release's, byte for byte. What changed is one setting in how the model is packaged: every photo ever sent to it had arrived at half the detail it was built to read, under half the pixel area. That hurt reads of male physiques far more than female ones, because what separates one body-fat band from the next on a man is fine texture a few pixels tall, and shrinking an image blurs the fine cue away while leaving the coarse one intact. Fixed and confirmed on 154 held-out photos, repeated on a second run. Meal photos improved too — all of them parse now, and the calorie error fell. Cost: photos take longer to analyse, up to 2.6x on large body-fat shots, which is a deliberate trade of speed for accuracy. Worth saying plainly — fourteen training rounds had tried to teach the model to read men better and every one of them failed, because the fault was never in the training data.
2xthe picture detail the model now sees98% → 100%meal photos read cleanly0changes to the text coach - Release 7Coach + vision merged
One model that sees too — at 44% of the download
The coach and the separate vision model became a single file. Two downloads totalling 6.3 GB are now one at 2.8 GB, and the crashes that came from swapping between two engines mid-session went with them. The last known defect went too: meal plans that quietly slipped in a food you'd told it to avoid — 5 plans out of 36 before, none now, once the training examples were matched to the exact format the app actually sends. Photo body-fat reads improved as well, but still tend to read a band high — treat that number as a trend line, not a measurement. The text side reaches installed apps now; the merged vision half arrives with the next app update.
6.3 → 2.8 GBto download5 → 0plans out of 36 with an avoided food0problems across 89 held-back checks - Release 6
Challenges that survive being generated, and Danish that stays Danish
Two wins in one release. Rebuilt on a base that holds up better when it's compressed to run on a phone, which fixed AI challenges that used to come back malformed three times out of four — and made answers a little quicker on the way. Then Danish got a proper pass: macros read “kulhydrater og fedt” instead of “carbs and fat”, challenge titles come back in Danish rather than English, and every meal reply now carries its nutrition summary. Still honest about the gaps — Danish word endings are occasionally off, and follow-ups about a past activity you never logged got slightly worse.
4/4challenges generated cleanly (was 1 in 4)7/7meal replies with their nutrition summary - Release 5
Realistic challenges, in the language you asked in
Fixed two real bugs in AI-generated challenges: targets for low-frequency goals were wildly out of reach (it once asked for 150 progress photos in six weeks from a standing start), and a challenge requested in Danish sometimes came back in English. Routine and meal-plan generation, dietary rules and equipment constraints all held exactly where they were.
2 bugsfixed, none introducedDanishgrammar hand-checked - Release 4
Dietary rules it actually respects
Closed a genuine gap where a “vegetarian, no nuts” meal request came back with peanut butter in it. Taught it the correct technique for face pulls and planks — neither had a full worked example to learn from before — plus the difference between a squat and a hip thrust. Won clearly on both a lenient and a strict count of mistakes.
3 runsindependently tested across0new problems introduced - Release 3Speed update
The same model, 88% faster
No retraining at all — the exact same coach, packaged for the phone a smarter way. Answers come back 88% faster and the download is 85MB smaller. An early single run hinted this might cost quality; a proper repeat test over 300+ responses showed that was random noise, and quality is statistically unchanged.
88%faster responses−85MBsmaller download300+responses re-tested - Release 2
Form answers back to a perfect record
Exercise-form questions returned to a zero-error record — the version before this one got direct form questions wrong four times out of five. Workout and meal-plan output stopped coming back garbled or duplicated. Across the same 52-question test set: 12 mistakes before, 2 after.
12 → 2mistakes on 52 questions0mistakes on form questions - Release 1
The first custom-trained coach
The first version trained on our own coaching data to beat the stock model it replaced — shorter, sharper answers on general advice with the same accuracy. It still tripped on exercise-form questions and long plan generation, which is exactly what the next release went after.
- BaselineBefore any training
Google's stock model, untouched
An off-the-shelf Gemma model with no fitness training of its own. Every release since has had to beat it as well as the version it replaces — and it's still the yardstick each new candidate is measured against.
No model ever ships on vibes. Every candidate is scored against the one it would replace across a large fixed question set, repeated over several independent runs — if it doesn't clearly win, it doesn't ship, and it doesn't appear above.
How the model reaches your phone
Owning your AI outright raises fair questions — how big is it, how does a better one get to you, and what leaves the phone when it does. The short answers:
One download, once
The coach model is a single ~2.8 GB file the app fetches after install — it isn't baked into the app download. That one file now contains the vision model too, which used to be a second, larger download: the pair came to 6.3 GB before. It's the same file on every phone: what changes per device — graphics chip or processor, how much memory it may use — is set by the app at runtime, so there is nothing to pick. Two files are published, and the app chooses between them by version: builds from before August keep the older text-only coach they were built for, and current builds fetch the combined file, which is the one every release below goes into. The download is checked against a signed fingerprint before it is used, it waits for Wi-Fi by default, and it asks first before touching mobile data.
Better models arrive free
When a new coach model wins its evaluation, it is published and the app offers it to you as a model update — no new app version needed, no separate model subscription, no re-purchase, no paid upgrade tier. A new model is a fresh download, and it waits for Wi-Fi just like the first one did.
Nothing about you goes out
A model download is a plain file fetch from a public Hugging Face repo — no account, no telemetry, nothing about your training, photos or meals sent before, during or after. Checking whether a newer model exists is the same kind of plain request. Pull the SIM out afterwards and the coach keeps working exactly the same.
What your phone needs
About 6 GB of RAM and a recent mid-range or better chip. The coach runs GPU-accelerated — roughly twice as fast as the same model on the CPU alone — and photo reads now run on the graphics chip too. Coaching and photo analysis share one engine over one file, so the app no longer holds two models in memory: on our test phone, peak memory use dropped from about 4.6 GB to 3.7 GB. The Play Store only offers the app to phones that clear that bar, so you can't buy into a device that runs it badly.
What's coming
SmartSight 3.0 is here. Here's what's next
The bet behind every update is the same: take the headline features people pay monthly subscriptions for, and make them run fully on your phone instead. Here's what just landed and where SmartSight is heading — all included free with your one-time purchase.
Shipped in 3.0
Available nowOne model, every phone
The coach and photo analysis now run on a single model we trained ourselves — one 2.8 GB file instead of a 6.3 GB pair, the same file on every supported phone, and noticeably lighter on memory.
A new rep engine
Form scoring was rebuilt from the ground up: every built-in exercise has its own checks, clean reps are no longer marked down, and a face-only frame never gets a score.
Film from more angles
Diagonal camera views are graded too, not just side-on — so you have more freedom in where you put the phone.
A coach in your style
The coach reads your training history, answers in 8 languages, and speaks in the tone you pick — Gentle, Balanced or Tough — from chat to form cues. The new offline voice engine takes a whole command at once.
Next up
In progressMeal plans that cover your micros
Generated plans that follow your diet preference, land on your calorie target, and cover fibre and the vitamins and minerals you need as well as they can.
Film from any angle
After diagonal views, the goal is to score every exercise from the front, side or anywhere in between, with the camera angle worked into the maths instead of a setting you have to pick.
Fairer scores on every lift
The new scoring rolls out exercise by exercise: points you earn for the things a coach actually looks for, and nothing taken off for what the camera couldn't see.
SmartSight on iPhone
The app is built on a shared core, and the iPhone version is in active development — the same offline coach, form tracking and nutrition, with the same one-time price.
Further out
ExploringLogging in two taps
A faster manual path for sets and meals that never waits on the AI model — for the days you just want to write it down and go.
Form that fits your body
Every lifter's proportions are different. The idea is for the camera to learn yours over time on your phone, so depth and range are judged against your build, not an average one.
Progress you can see
Instead of chasing a single body-fat number, compare your own progress photos over time and get an honest read on whether you're leaner, holding steady or softer.
This is a living roadmap, not a set of release dates. The privacy promise never changes: no mandatory cloud, no account required to train, and nothing leaving your device.
Your data never leaves your phone
SmartSight AI is private by design. Everything that makes it smart runs on-device — so you get a real AI coach without handing your body to the cloud.
✓On-device AI
The coach, voice recognition, speech and photo analysis — body fat, meals and nutrition labels — all run locally. No server ever sees your data.
✓No accounts, no cloud
There's no sign-up and no cloud sync. Your workouts, photos and meals live only on your phone.
✓No ads, no trackers
Nothing is sold, profiled or monetized in the background. The app respects your attention.
✓Opt-in networking only
The only optional network calls are an anonymous barcode lookup — which you can skip entirely by scanning labels or adding foods offline — fetching a recipe or exercise-pack link you give it, and the AI and voice model downloads and update check, which send nothing about you.
Pricing
Pay once. Yours forever.
The core app is free. Unlock the AI extras with a single one-time purchase — every future update included, never a subscription.
Free
The core app, forever.
- ✓Camera form coaching, rep by rep — one exercise a week, then Pocket mode
- ✓3 AI coach answers a week — questions, meal plans or routines
- ✓Workout tracking, PRs, streaks & trends
- ✓Body progress photos & measurements
- ✓Nutrition tracking: macros, fibre & 12 vitamins and minerals
- ✓Pocket mode with spoken rep callouts
- ✓Goals & daily readiness check-ins
Pro
Pay once. Every future update included.
Less than 3 months of the ~$25/mo app stack it replaces.
- ✓Everything in the free app
- ✓Unlimited camera coaching & AI coach answers
- ✓Unlimited AI meal plans, routines & programs
- ✓Hands-free voice: commands, logging & coach Q&A
- ✓Meal & nutrition-label photo scans
- ✓AI photo body-fat read
- ✓Multi-week programs & weekly reports
- ✓Encrypted backup & restore, plus import from Strong, Hevy & MyFitnessPal
- ✓Pocket lock-screen controls & live widget
- ✓All future updates, free for life
FAQ
Questions, answered
No. The AI coach, voice recognition, speech and form analysis all run on-device, so the app works fully offline — in the gym, on a plane, anywhere.
Yes. There are no accounts and no cloud sync. Your workouts, photos and nutrition stay on your phone. The only optional network calls are an anonymous barcode lookup, fetching a recipe or exercise-pack link you give it, and downloading the AI coach and voice models (plus checking for a newer coach model). None of them send your training, photos or meals.
Never. Pro is a single one-time purchase that unlocks everything for life. No monthly fees, no ads, no trackers.
A lot. Workout, nutrition and body tracking are free forever, with no time limit. Free also includes one camera-coached exercise a week (then Pocket mode for the rest), spoken form cues and rep callouts, and 3 AI coach answers a week, which also cover meal plans and routines. Pro removes those limits and adds hands-free voice control and voice chat with the coach, meal and label photo scans, the photo body-fat read, multi-week programs, weekly reports and encrypted backup.
Yes. Export your data from Strong, Hevy or MyFitnessPal and import the file in Settings — it is read entirely on your phone, and nothing is uploaded. CSV import is part of Pro; a MyFitnessPal printable diary imports free.
Yes. Your one-time purchase includes every future update at no extra cost — the AI model itself included. SmartSight is actively developed, with new features and improvements shipped regularly — no upgrade fees, no paid upgrade tier, ever.
The coach model is a single ~2.8 GB file downloaded once after you install the app, from a public Hugging Face repo you can inspect yourself. That one file handles both text and photos on every supported phone — it replaces a coach-plus-vision pair that came to 6.3 GB. Two files are published and the app chooses the right one for its version, so an older build is never handed a download it can’t use. The app prefers Wi-Fi and asks before using mobile data.
Yes, and it costs nothing. The coach model is retrained and benchmarked against the version it would replace — it only ships if it measurably wins. Every improvement so far has reached users as a free model update inside the app. See the model log for the real release history.
Android phones with about 6 GB of RAM and a recent mid-range or better chip. The app is built on a shared Kotlin Multiplatform core, and an iPhone version is in development.
Eight: English, Danish, German, Spanish, French, Norwegian, Swedish and Simplified Chinese — including exercise names, coach answers and voice commands.
Just your phone's camera. Optionally pair a Bluetooth heart-rate monitor.
Train smarter. Stay private.
Your AI form coach, nutrition tracker and physique log — in 8 languages, all on your phone, all offline. No subscription, ever.