It's like having a spotter who silently writes down every single exercise, set, and rep you do at the gym without you ever touching your phone.

Fort's core ML use case is automatic exercise detection and rep counting. The wrist-worn device captures continuous streams of accelerometer and gyroscope (IMU) data alongside PPG heart rate signals. Proprietary time-series classification models—likely a combination of convolutional and recurrent neural networks trained on labeled datasets from beta testers—segment raw sensor streams into discrete exercise events, classify the specific movement (e.g., bench press vs. overhead press vs. bicep curl), and count individual repetitions within each set. The system must handle enormous variability: different body types, lifting speeds, equipment variations, and transitions between exercises. Sensor fusion between IMU and heart rate data adds a physiological dimension that improves classification confidence, particularly for distinguishing similar movement patterns. The screenless form factor means all inference must run on-device with minimal latency, requiring highly optimized edge ML models. Beta feedback loops continuously expand the training dataset and improve model generalization across the growing exercise library.

It's like Shazam for your muscles—except instead of identifying a song from a few seconds of audio, it identifies your exercise from a few seconds of wrist movement.

It tells you how many reps you have left in the tank before your muscles give out, so you know exactly when to push harder or stop.

Proximity to failure (PTF) is one of the most scientifically validated predictors of hypertrophy and strength adaptation, yet it has historically been nearly impossible to measure outside a lab. Fort's ML system addresses this by building regression models that analyze intra-set rep-to-rep degradation patterns in velocity, acceleration, power, and cadence captured by the IMU, combined with heart rate variability and cardiac drift from the PPG sensor. As a user progresses through a set, each successive rep typically exhibits measurable decreases in concentric velocity and increases in time-under-tension—signals that correlate strongly with proximity to muscular failure. The ML models learn user-specific fatigue curves over time, personalizing predictions based on individual strength levels, training history, sleep quality, and recovery status. This creates a dynamic, adaptive system that improves its accuracy the more a user trains. The output is a simple, actionable metric (e.g., "2 reps remaining") displayed post-set in the companion app, helping users train at the optimal intensity for their goals without risking overtraining or undertraining.

It's like a fuel gauge for your muscles—instead of guessing whether you're running on empty, you get a real-time readout of exactly how much is left in the tank.

It watches how you move during every rep and tells you when your form starts breaking down—like having a personal trainer's eye on your technique at all times.

Fort's form analysis system represents a sophisticated application of biomechanical anomaly detection using wrist-mounted IMU data. For each recognized exercise, the ML system maintains a learned "ideal movement template"—a statistical representation of proper form derived from aggregated data across users and validated against exercise science literature. During each rep, the system extracts kinematic features including peak concentric velocity, eccentric control ratio, range of motion (estimated via angular displacement), rep-to-rep cadence consistency, and movement path regularity. These features are compared against both the population-level template and the user's own historical baseline using anomaly detection techniques. Significant deviations—such as decreasing range of motion, asymmetric velocity profiles, or erratic cadence—trigger form quality alerts. The system is particularly powerful for detecting fatigue-induced form breakdown, where rep quality degrades as a set progresses. Over time, the personalization layer learns each user's unique movement signatures, reducing false positives and increasing the relevance of feedback. This creates a closed-loop system where users receive post-set or post-workout form reports in the companion app, complete with specific, actionable suggestions for improvement.

It's like spell-check for your squats—it underlines the sloppy reps in red so you can fix them before they become bad habits.