July ended with our sensors recognising movements they’d never been taught. August was about turning that from a demo into something that could actually live inside an app, alongside a significant upgrade to our sensor hardware and a rethink of how our training plans work.
We also crossed 10,000 views on Instagram, which for a four-person team with no marketing budget felt worth a moment.
Rep Counting Now Runs On Your Phone, Offline
The capability we’ve been chasing works: perform a movement once, and Korvi counts every repetition of it from then on, running entirely on your phone with no internet connection needed.
That last part matters more than it sounds. Gym signal is famously terrible, and a rep counter that stops working in the basement free-weights area is not a rep counter. Everything happens locally, on the device in your pocket.
August was spent making it robust. A demo that works when you perform each rep at a steady tempo is a very different thing from something that survives a real set, where your first rep is explosive, your last one grinds, and your form drifts as you tire. Hardening that is unglamorous work, and it’s most of what the month went on.
Velocity-Based Training, Without the Lab
Here’s the capability we’re most excited about.
Velocity-based training has been used by elite strength coaches for years. Instead of guessing how hard a set was, you measure how fast you’re moving. As you fatigue, your reps slow down, predictably and measurably. That slowdown tells you far more about how hard you’re genuinely working than a number you pick after the fact.
The catch has always been access. Proper velocity tracking has meant expensive barbell-mounted devices or a sports science lab, and it typically only works on barbell lifts.
Because Korvi tracks you rather than the bar, we can look at how your movement changes across a set for essentially any exercise: barbell, dumbbell, machine, or bodyweight. In August we mapped out how that becomes something genuinely useful rather than another number you ignore, a real, measured picture of fatigue and effort, feeding into what you should be lifting next session.
This solves a specific and very common problem. If you smash a punishing set of squats and then move to leg press, a naive app sees weak leg press numbers and concludes you’re weaker than you are. A system that understands fatigue knows you’re simply tired, and adjusts accordingly. That’s the difference between a workout log and something worth calling a coach.
For now, the app works from the RPE you enter yourself. The system is built so that once the sensors are in your hands, that number can be measured rather than estimated. Same approach, better inputs.
Rebuilding the Brain Behind Your Training Plan
The other major piece of August was structural. Our AI training plan had outgrown its foundations, so we redesigned how the whole thing thinks.
A few principles came out of that work that we’re quite proud of:
Every decision is recorded, and every decision can be explained. When Korvi changes your plan, there’s a record of why, and rather than waiting for you to ask, the app tells you at the time. “We’ve added a rest day because your last three sessions have been slowing down.” Fitness apps that hand down instructions with no reasoning are asking for a lot of faith. We’d rather show our working.
Nothing changes behind your back. If the system notices you’ve skipped three Mondays running, it asks whether Monday is a bad day for you. It doesn’t quietly delete Monday and decide you’re a person who doesn’t train on Mondays. That distinction matters. Get it wrong and the app slowly shrinks itself around your worst habits instead of helping you build better ones.
It knows what’s temporary and what isn’t. A recurring shoulder issue, or a goal to bench 100kg by December, is long-term context that should shape your training for months. One bad night’s sleep isn’t, but it is the reason yesterday’s session felt rough, and the system accounts for that rather than concluding you’ve got weaker overnight.
Progressive overload that responds to reality. Most plans assume you’ll add weight every week forever. Real training doesn’t work like that. The point of all the above is a plan that adjusts to how your training is actually going, not how it was supposed to go on paper.
Designing How the Sensors Feel to Use
We spent time in August on a question that’s less technical but just as important: what should using the sensors actually feel like?
The obvious approach is to tell the app what exercise you’re about to do, then do it. We’re exploring something better, that you don’t tell it anything. You put your sensors on, you train, and Korvi works out what you did and checks with you.
There’s a nice property to that: every time you confirm, the system learns, and gets better at recognising your movements specifically. We’re still working through exactly how this looks and feels in practice, and we’d expect it to change several times before you see it. But the direction is clear. The less you have to touch your phone mid-workout, the better.
A Major Hardware Upgrade: V3.1
Our sensors are moving to version 3.1, and it’s a substantial step up from the V2 units the team has been training with.
V2 proved the concept, but every unit was slow, delicate, hand-assembled work: fine for a handful of prototypes, impossible at any real scale, and a genuine bottleneck on how quickly we can collect data. V3.1 is designed around a single principle: zero soldering. Drop the board in, plug in the battery and the charging connector, close the case. Every unit identical, assembled in minutes rather than most of a week.
It’s also flatter, so the sensors sit comfortably under clothing rather than protruding and catching on things, and it leaves clean room to add features like haptic feedback later without needing another board redesign.
August was spent on the unglamorous but essential work of verifying that design: checking every connection on the board, and building a full 3D model to confirm the thing can genuinely be assembled by human hands. That modelling has already caught problems that would have been expensive to discover after manufacturing. This is the part of hardware development nobody posts about. Slow, meticulous checking, because a mistake here costs weeks and a board order, not an afternoon.
Boards are close to being ordered.
10,000 Views and Counting
Our Instagram passed 10,000 views this month.
We’re aware that isn’t a number that impresses anyone in social media. But we launched properly only weeks ago, we’ve spent nothing on advertising, and we’re a strength training startup with no product to sell yet, so people watching us build one is genuinely encouraging.
App downloads keep arriving organically too, from people we’ve never met who found Korvi on their own. If you’re one of them: thank you, and please tell us what’s broken.
We’ve also just started building on TikTok, which is where the more entertaining side of what we do belongs. Come find us there, on Instagram, or over on r/Korvi.
What’s Next
September is about three things:
- V3.1 sensors in more hands. Getting boards ordered and assembled so the whole team can collect data in parallel, currently the biggest constraint on how quickly everything else improves.
- Bringing detection into the app. Moving the sensor experience out of a standalone demo and into the product you actually use.
- A better training plan. The architecture work is done; now it gets built. This is the feature we think genuinely separates Korvi from the other gym apps on your phone.
The app is free on the App Store and Google Play, and it works right now without any sensors at all. If you lift, give it a try and tell us where it falls short.
Thanks for following along.
The Korvi Team
Korvi is a free AI workout tracker with adaptive training plans, automatic rep counting, and a wearable strength training sensor system in development. Available now on the App Store and Google Play.
