The pods are the input layer,
not the product.
The variables that actually govern how you adapt cannot be captured by hand by anyone. Nobody times a 1.4 second eccentric while they are under the bar. Nobody holds thirty per-rep durations in their head. Nobody notices their left side began giving out two reps before their right. That is not a discipline problem and it is not something a better logging app solves: those measurements do not exist without instrumentation on the body.
Tensio produces them as a by-product of a normal session, already labeled with the movement and the load because you said both out loud when you started the set. The result is clean, hundred-hertz motion data attached to a known exercise at a known weight, per side, per rep. Nobody has that at scale, because nobody else built a way to collect it that a person will willingly use twice.
That is what makes real personalization possible. Software generalizes: a program in an app is written for a population, and it cannot know that your bench slows at rep six while your squat holds to rep ten, or that you recover from volume faster than the template assumes. Measured per rep, per side, every session, it can.
A single sensor measures one arm and reports it as the set. Your two sides do not move identically, so that number is an assumption about the side it cannot see, and asymmetry is not even available to it: there is no second measurement to compare against. That gap is where a lot of lifters quietly lose progress and build imbalances. Tensio reads two pods at a hundred samples a second and measures each side against the other. Worn on both wrists, because for barbell and dumbbell work the hands are on the load, so what the pods see is what the weight did. Other mounts come once the model has been trained on them, not before.
Inputs
Two synchronized pods, a voice agent that captures the exercise and the weight without you typing either, and a training corpus that grows every time somebody uses the product for the reason they bought it. First batch under production.
Intelligence
A neural network already runs in production, segmenting a lift into its phases and counting reps from raw motion more accurately than the hand-written logic it replaced. That is the loop closing: our data trains a model, the model ships, and every set after it improves the next one. The aim is not one all-encompassing network but specialized models composed together, and fatigue is the one we are building now, read from tempo decay rather than asked about. On the training corpus it scores 98.2% of sets and 99.75% of reps correctly; how it does on lifters it has never seen is the next thing we establish.
Coaching
Every coach on the platform prescribes a session, sees how it went, and adjusts the next one. Those decisions, next to the measured sets they were made from, are what the models learn from. The aim is to extend what good coaches do to the people who never had one: change the plan while you are still standing there, add a set because you have more in you, hold the load on a day your own numbers say to. Against your training, measured, rather than a population average.