Built on Science

How Softball AI Trainer analyzes your mechanics

Our Approach

Softball AI Trainer uses Google ML Kit pose detection (MediaPipe) to extract 33 skeletal landmarks — including heel and foot position, critical for stride mechanics — from your video, entirely on your phone. We then apply biomechanical scoring algorithms. Our pitching thresholds are anchored to the published fastpitch kinematic studies cited below. Our hitting checkpoints are coaching-standard swing markers; where they overlap with the collegiate softball swing research cited below — hip and trunk rotation through contact and the finish — that research informed what we measure, and the trunk-rotation power literature supports the weight we give rotation. Because published youth softball hitting kinematics are limited, our youth hitting targets are calibrated against our own recorded-swing dataset rather than taken from collegiate values.

All video processing happens on-device. Your video is never uploaded — the only images that leave your phone are analysis cards you explicitly choose to share.

Pitching Analysis

Our pitching analysis evaluates four phases: Power Load, Stride, Release Zone, and Follow-Through:

• Power Load — Trunk alignment and drive knee flexion at push-off
• Stride — Ankle-to-ankle stride length at foot plant as a percentage of estimated standing height, with age-adjusted targets
• Release Zone — Arm speed, hip lead (how far the hips are ahead of the shoulders at foot plant), and release position
• Follow-Through — Balance and body control after release

Hitting Analysis

Our hitting analysis breaks down your swing into four phases: Load, Stride, Contact Zone, and Follow-Through. Each phase is scored using biomechanical metrics derived from pose keypoints:

• Load — Hand position, back-knee bend, and weight distribution over the back foot
• Stride — How far the front foot travels from load to plant, as a percentage of estimated standing height, with age-group targets
• Contact Zone — Hip rotation angle at contact and back-arm elbow bend at contact
• Follow-Through — Finish height, wrist crossover, and whether the finish is held (balance, and the feet staying planted after contact)

Contact detection uses a custom-trained ball-detection neural network that tracks the ball through the contact window to identify the exact moment of impact, with pixel-level frame differencing as a fallback.

Age-Appropriate Scoring

Youth athletes are not small adults. Our scoring thresholds step up by age bracket (10U, 12U, 14U, 16U, 18U). Where published youth and collegiate values exist we anchor to them: the Power Load drive-knee target runs from the published youth mean (about 25°, measured at ages 12–15) at 12U to the collegiate mean (about 38°) at 18U, with 14U and 16U interpolated and 10U extrapolated below the youngest measured group. Pitching stride targets are anchored to published youth values too. A pitcher earns full stride credit at the published youth mean minus half a standard deviation: 56.5% of height for 10U, from the 9-to-13-year-old cohort (61% ± 9%), and 82.5% for 14U and older, from the 12-to-15-year-old cohort (89% ± 13%), with 12U (69.5%) interpolated between them. 16U and 18U share the 14U target; the one direct youth-versus-collegiate comparison found no difference in stride length. A long pitching stride is never penalized, and a shorter one is scored down gradually rather than failed. Hitting stride targets are short, controlled steps of about 4% to 8% of height, rising with the age bracket; a hitting step well short of or well past the target scores lower.

Stride is measured from side-view video: for pitching, the ankle-to-ankle distance at foot plant; for hitting, how far the front foot steps from load to plant. Both are divided by an estimate of the athlete’s standing height — the tallest standing nose-to-ankle span in the video before foot plant (pitching) or contact (hitting), taken as the middle of the three tallest frames so one bad frame can’t skew it, divided by 0.88, because nose-to-ankle height is roughly 88% of stature. The result is an estimate; see Limitations.

Research References

Our scoring targets and Coach Jodi’s drill rationale draw on the following peer-reviewed research. Where a paper is background reading rather than the source of a shipped threshold, the note says so.

• Biomechanics of Fastpitch Softball Pitching: A Practitioner’s Guide (2025)
  PMC11969493 — Friesen et al., Sports Health. Clinical review of the four windmill-pitch phases, proximal-to-distal kinetic-chain energy flow, and published youth vs. collegiate normative values for stride length and trunk rotation. Its youth stride value (89% ± 13% of body height at foot contact) anchors our 14U-and-older stride target. It also informs our decision not to penalize long strides, and Coach Jodi’s proximal-to-distal sequencing cue and Power Load / Release Zone drill rationale. A review; it reports no original sample.
• Differences in Lower Extremity Kinematics Between Collegiate and Youth Softball Pitchers (2021)
  PubMed 34778481 — Downs Talmage et al., Orthopaedic Journal of Sports Medicine. Motion-capture comparison of 40 youth (mean age 13.7) and 43 collegiate pitchers. Collegiate pitchers showed greater drive-knee flexion at the top of the arm circle (about 38° vs. 25°); stride length as a percentage of height did not differ between the groups. The two knee-flexion means anchor our 12U and 18U Power Load drive-knee targets; 14U and 16U are interpolated between them and 10U is extrapolated below the youth data.
• Pitching Mechanics in Female Youth Fastpitch Softball (2018)
  PMC6044599 — Oliver et al., International Journal of Sports Physical Therapy. Electromagnetic motion-capture study of 23 Little League pitchers (ages 9–13, mean 11.4). Only trunk kinematics correlated with ball velocity — greater trunk flexion at every pitching event and more trunk rotation toward the throwing arm at the top of the backswing; pelvis, shoulder, elbow, and stride-length measures did not. It also reports a mean youth stride of 61% ± 9% of body height, which anchors our 10U stride target (12U is interpolated between it and the older-youth value). Also our youth reference for the trunk-position variables the Power Load phase measures.
• Kinematic Differences Between Hitting Off a Tee Versus Front Toss in Collegiate Softball Players (2018)
  PMC7857448 — Washington & Oliver, International Biomechanics. 22 NCAA Division I players; compared with tee work, front toss produced greater lead-knee flexion at foot contact and greater trunk rotation at ball contact, while tee swings showed more pelvis and trunk rotation at follow-through. Informs the hip-rotation-through-contact and finish-rotation variables we score and Coach Jodi’s tee-versus-toss drill rationale. Its 3-D angles use a different reference than our 2-D side view, so no numeric target is taken from it.
• Sport-Specific Differences in Power–Velocity–Force Profiling During Trunk Rotations at Different Loads (2020)
  doi:10.3390/app10238366 — Zemková, Poór & Jeleń, Applied Sciences. Trunk-rotation power profiling across rotational-sport athletes; supports the weight we give hip rotation in Contact Zone and Follow-Through scoring and in Coach Jodi’s rotational-power drills. Not softball-specific and no numeric threshold is taken from it.

Technology

• Pose Detection — Google ML Kit (MediaPipe Pose, 33 landmarks), running fully on-device
• Ball Detection — A custom-trained neural network for softball tracking (contact detection), running fully on-device
• AI Coaching — Anthropic (the company that makes Claude), receiving the athlete's age bracket, analysis scores, skeletal position measurements, and typed questions — never images, video, or the athlete's name (the App swaps the name for a placeholder before sending and restores it only on the device)
• Privacy — All video analysis is on-device. The App never uploads video or images.

Limitations

Our analysis uses 2D pose estimation from a single camera angle. This provides reliable biomechanical insights but has inherent limitations compared to 3D motion capture systems:

• Arm speed is measured from 2D side-view video, so it understates true 3D wrist velocity by an amount that varies with camera angle. Our benchmarks are set against 2D measurements and are not comparable to radar or motion-capture values.
• Accuracy depends on video quality, camera angle (side view recommended), and lighting conditions.
• Pose models may have difficulty tracking small or fast-moving limbs, particularly for younger athletes.
• Stride percentages are estimates. Standing height is estimated from the video rather than measured, using the approximation that nose-to-ankle height is about 88% of stature, and the stride itself is a 2D measurement from a single camera. The pitching targets come from motion-capture studies, so small differences from a target should not be over-read.
• The App’s pitching stride score rewards a stride at or above the age target and never penalizes a longer one. In collegiate pitchers, shoulder distraction force has been found to rise modestly with stride length (Oliver, Yanagita & Bordelon, Orthop J Sports Med 2024, 63 NCAA Division I pitchers; stride length explained about 11% of the variation), so whether an unusually long stride suits a particular pitcher is a judgment for the pitcher’s coach, not the App.
• The App does not measure bat speed or the force, work, and power a hitter applies to the bat. Those require an instrumented bat or 3-D motion capture (see Milanovich & Nesbit, J Sports Sci Med 2014, 14 collegiate hitters swinging off a tee), and no hitting score in the App is derived from them.

Softball AI Trainer is a training aid designed to supplement — not replace — qualified coaching.

Nothing in the App or on this page is medical advice. The App is not designed to diagnose, treat, or prevent any injury. The research cited above is listed because it informed our scoring targets or our understanding of the mechanics we measure, not as a claim that the App reduces injury risk.