Your running form, decoded.
Turn raw Apple Watch telemetry into personalized, actionable coaching. Live cadence deltas, 30-day relative baselines, and targeted form drills—100% on your device.
Built for runners who value form and privacy.
No generic 180 SPM rules. No mandatory account signups. Every insight is generated locally on your iPhone.
30-Day Relative Baseline
Your cadence, pace, and cardiac efficiency are benchmarked strictly against your personal 30-day rolling history—never arbitrary textbook numbers.
Apple FoundationModels
Runs local LLM inference directly on Apple Silicon. Swift applies deterministic physiological math before prompting for guaranteed safety.
Targeted Form Drills
Structured 4-field prescriptions: Purpose, Active Work, Effort, and Rest. Prescribes Cadence Pyramids, Rhythm Intervals, and Strides.
Take a closer look at the experience.
Intelligent Running Biometrics
Ditch static vanity metrics. Runalyst extracts live cadence deltas, vertical oscillation, and ground contact balance from Apple Watch and benchmarks them against your rolling 30-day relative baseline.
- Zero arbitrary targets: evaluated strictly against your personal physiology
- Sub-second cadence delta tracking against your 30-day average
- Working stats engine that automatically trims dead-stop traffic pauses
The science, math, and logic under the hood.
Most fitness apps compress hours of continuous running into flat scalar averages. Here is how Runalyst preserves physiological time series, isolates authentic biomechanics, and guarantees AI coaching safety.
Topological Time-Series Classification
Standard running apps use scalar Coefficient of Variation (CV = σ/μ) to classify runs. But scalar compression destroys temporal sequence: a continuous steady run with a natural warmup and late-run fade produces the exact same CV as an intentional 6-interval track session.
- 1. Zero-Crossing Oscillation Gate: Cadence time-series is smoothed with a 2-bucket (30-second) sliding window. We count transitions across the session mean to detect genuine surge-and-recovery alternations.
-
2. Multi-Signal Corroboration: A cycle is only validated if at least 2 of 3 physiological signals corroborate it:
Δ Cadence ≥ +8 SPM Δ Pace ≥ 15 s/km Δ HR ≥ +5 BPM (~15s lag comp)
-
3. Duration Guards & Regularity: Requires ≥15s surge and ≥20s recovery phases. If valid cycles ≥ 3, cycle variance separates structured Intervals from irregular Fartlek. If continuous, monotonic quintile acceleration (
CV < 0.09) detects true Progression runs.
Deterministic Guardrails over LLM Hallucinations
On-device generative models should never compute raw biomechanical arithmetic or balance competing physiological priorities. Small models hallucinate math and fail safety boundaries under prompt conflicts.
All baseline statistics, deltas, and fatigue states are computed locally in RunAnalyzerActor before Apple FoundationModels is ever invoked.
The LLM receives one unambiguous physiological directive. The model provides human empathy and conversational tone, but the prescriptive drill selection is 100% mathematically deterministic.
30-Day Time-Locked Longitudinal Baselines
Generic coaching claiming every runner must hold 180 SPM is physiologically incorrect. Optimal cadence and vertical oscillation depend on leg length, height, running economy, and aerobic history.
-
Strict Rolling Window: Every biometric split is benchmarked against the runner's personal rolling 30-day baseline time-locked to that workout (
$0.date < targetDate). - No Hindsight Leakage: Historical runs are evaluated strictly against what your baseline was on the day you ran—never retroactively biased by runs that happened weeks later.
- Coupled Form Diagnostics: Vertical oscillation (>10 cm) is mathematically linked to cadence to detect forward bounding vs. compact turnover.
Dead-Stop Trimming & Golden Fixtures
Waiting at traffic lights and street crossings corrupts pace and biomechanical averages with resting zeros. Synthetic metronomic unit tests mask real-world sensor noise.
-
Dead-Stop Bucket Trimming:
FramboiseEnginedetects and trims stationary pauses to calculate true Working Stats. When no stops occur, it strictly inherits raw HealthKit double-precision values to prevent floating-point drift. -
Authentic Golden Telemetry: Tested against authentic continuous recordings (
authenticIntervals31Min,hillySteadyAerobicRun,fadingCadenceFatigueRun). -
Adversarial Noise Injection: Validated with
TelemetryNoiseInjectorinjecting ±1.2 SPM cadence, ±1.5 BPM HR, and ±3.0 s/km pace jitter to guarantee zero-threshold fragility.
Architectural Comparison: Runalyst vs. Standard Fitness Apps
| Engineering Dimension | Standard Fitness Apps | Runalyst Native Engine |
|---|---|---|
| Run Classification | Scalar averages & Pace CV (collapses temporal sequence) | 3-Layer Topological Zero-Crossing with multi-signal corroboration (≥2/3 signals) |
| AI Coaching Logic | Prompting cloud LLMs to estimate physiological math | Local Deterministic Swift Guardrails • On-device FoundationModels |
| Baseline Standards | Static universal charts (e.g. 180 SPM for all body types) | Time-Locked 30-Day Moving Window ($0.date < targetDate) |
| Sensor Noise & Pauses | Traffic stops dilute stride length and cadence | FramboiseEngine Dead-Stop Trimming with raw double inheritance |
| Privacy & Data Pipeline | Health telemetry transmitted to external cloud servers | 100% On-Device • SwiftData SQLite • Apple Silicon local inference |
Ready to run smarter?
Experience private, on-device biometric telemetry and intelligent AI form coaching on your iPhone and Apple Watch.