Enhancing Basketball Shooting Performance with AI-Powered Biomechanical Tracking

SUMMARY

A U.S.-based sports technology startup partnered with Folio3 AI to build a native iOS platform integrating the MetaMotionS wearable sensor with LSTM-based machine learning models. 

The goal: deliver real-time biomechanical shooting analysis to players and coaches. Folio3's team of 5 engineers delivered the full solution in 12 weeks, achieving 99% shooting form detection accuracy and cutting coach review time by 65%.

ABOUT THE CUSTOMER

The client is a U.S.-based early-stage sports technology business on a mission to democratize elite-level basketball coaching. Their product vision centers on giving every player, from high school athletes to semi-professional teams, access to the kind of motion analysis data that was previously locked inside expensive performance labs.

With a wearable-first product strategy, the company set out to build an AI-powered coaching platform that could decode the physics of a basketball shot and turn it into a personalized, real-time feedback loop. Prior to partnering with Folio3, they had the hardware device ready but lacked the software, AI infrastructure, and mobile engineering capability to bring the product to market.

THE CHALLENGE

Elite-level shooting analysis has historically lived inside performance labs. High-speed cameras, motion capture suits, and force plates produce detailed biomechanical reports, but they require controlled environments, costly equipment, and trained operators. For most athletes, that level of feedback is simply out of reach.

The client had already identified the hardware answer: a compact, wrist-worn IMU sensor capable of capturing six-axis motion data at scale. What they lacked was the software intelligence to make that data useful in real time, on a basketball court, without a data scientist in the room.

Three specific technical challenges defined the project:

• Raw sensor noise during dynamic movement — Accelerometer and gyroscope readings from wrist-worn devices capture every micro-movement, including dribble vibration, arm swings, and incidental gestures. Separating a shooting motion from ambient wrist activity required purpose-built signal processing and model design.
• On-device latency requirements — For feedback to influence form, it must arrive within seconds of the shot, not minutes. Cloud round-trip processing was too slow. The full inference pipeline had to run on the device itself using CoreML.
• Labeled training data scarcity — LSTM models require labeled sequential time-series data. Basketball shooting form datasets at the granularity needed, wrist sensor readings tagged by form quality across each phase of the shot, did not exist off the shelf. The team had to collect, structure, and label the training dataset from scratch.

Our DELIVERY APPROACH

Folio3 assembled a dedicated cross-functional team of 5 engineers, 2 iOS mobile developers, 2 ML/data science specialists, and 1 QA engineer and delivered the complete end-to-end solution within a 12-week development cycle.

The team designed and built a native iOS application that serves as the central intelligence layer between the wearable sensor and the coaching workflow. At its core, the platform captures live biomechanical data from the MetaMotionS device via Bluetooth, processes it through a trained LSTM neural network, and returns a real-time shooting form score with drill-down analytics on each phase of the shooting motion.

TOOLS & TECHNOLOGIES EMPLOYED

The SOLUTION

The team designed and built a native iOS application that serves as the central intelligence layer between the MetaMotionS wearable sensor and the coaching workflow. At its core, the platform captures live biomechanical data from the sensor via Bluetooth, processes it through a trained LSTM neural network running entirely on-device, and returns a real-time shooting form score with drill-down analytics on each phase of the shooting motion.

The Four-Stage Pipeline

Folio3 built a four-stage pipeline that takes raw wrist sensor data and converts it into a coaching-grade shooting form score — all running natively on iPhone without any cloud dependency.

1.     BLE data acquisition — The MetaMotionS device streams six-axis IMU data (3-axis accelerometer + 3-axis gyroscope) to the iOS app over Bluetooth Low Energy at 100 Hz. The app buffers incoming readings and timestamps each frame to preserve the temporal sequence required for LSTM inference.

2.     Signal preprocessing — Raw sensor readings contain significant noise from incidental wrist movements. A sliding-window filter isolates the shooting motion window, from the start of the upward arm drive to wrist snap at release, discarding ambient activity before and after the shot.

3.     LSTM inference — The preprocessed time-series sequence is passed to a CoreML-optimized LSTM model, trained on a purpose-built dataset of labeled shooting form samples. The model classifies form quality across five key phases: catch stance, dip, drive, release, and follow-through.

4.     Feedback delivery — Within 150ms of shot release, the app surfaces a composite form score, a phase-by-phase breakdown, and a personalized drill recommendation. Coaches can review full session history, export drill logs, and track form improvement over time.

SOLUTION ARCHITECTURE

RESULTS ACROSS 120+ COURTS