Geometry. Signal processing. Machine learning. How Flying Start predicts when you'll cross the start line — and what makes it different from hardware costing 50× more.
Every Time-To-Line calculation starts with a geometry problem. Your GPS position is a point. Your course over ground defines a ray extending from that point. The start line is a segment between the PIN and RC boat. Where those two lines meet — and how far away that intersection is — determines how long until you cross.
Flying Start uses a parametric ray-segment intersection algorithm to find the exact point where your trajectory crosses the start line. Unlike a simple perpendicular distance calculation, this accounts for the angle of your approach. Heading at 45 degrees to the line? Your TTL is longer than the straight-line distance suggests. Heading parallel? TTL is undefined — you'll never cross.
This is the same fundamental geometry that dedicated instruments like the Velocitek ProStart and Vakaros Atlas 2 use. The maths is identical. What differs is how accurate the inputs are.
GPS gets a bad reputation from urban canyons. In a city, signals bounce off buildings and the receiver can't tell a direct signal from a reflected one. The result is multipath interference — position errors of 5–15 metres.
Sailing is the opposite of a city. You're on flat water with an unobstructed view of the entire sky. There's nothing to reflect off. Every satellite signal arrives clean.
Phones with dual-frequency GPS (L1+L5) achieve 1–2 metre accuracy in open water. The L5 band adds a second frequency that eliminates ionospheric errors and helps distinguish direct signals from reflections. On the water, this gets you close to the practical limits of consumer GNSS. Phones with single-frequency GPS (L1 only) still achieve 3–5 metres in open water — significantly better than in a city.
iPhone: iPhone 15, 16, and 17 (all models), and iPhone 14 Pro / Pro Max. Standard iPhone 14 and iPhone SE are L1 only.
Apple Watch: Ultra, Ultra 2, and Ultra 3. Standard Apple Watch (Series 9, 10, 11, SE) uses L1 only.
Android: Most flagships from 2020 onwards — including Pixel 5+, Samsung Galaxy S21+, and OnePlus 9+. Check your phone’s specs for “L5” or “dual-frequency” GNSS.
Flying Start works on all GPS-equipped devices. Dual-frequency gives you better accuracy, but the Kalman filter and ML model improve predictions regardless of which band your hardware supports.
Dedicated hardware like the Vakaros Atlas 2 does have an advantage: 25 Hz update rates and differential corrections that push accuracy to ~25 centimetres. That's genuinely better. But the gap is narrower than most people assume — and the remaining difference can be addressed in software.
The Vakaros Atlas 2 achieves ~25 cm accuracy with dual-band GNSS at 25 Hz. The Velocitek ProStart uses a 25 Hz multi-constellation receiver with WAAS augmentation. Both are purpose-built for this job, and for official, referee-grade OCS calling at championship level, that precision matters. Flying Start's approach is to close the accuracy gap through signal processing and machine learning rather than hardware.
Raw GPS gives you a position once per second. Between those updates, the world keeps moving. A boat doing 5 knots covers 2.5 metres per second — and the GPS position you received half a second ago is already stale. Worse, consecutive GPS fixes jump around randomly within their accuracy circle, making speed and heading readings jittery.
This is the single biggest source of TTL instability in a simple GPS instrument. You calculate TTL from speed and heading. If speed jitters between 2.3 and 2.7 m/s from one fix to the next, TTL jumps by several seconds every update. It's correct on average, but useless for timing a start.
Flying Start runs an Extended Kalman Filter (EKF) that solves both problems simultaneously.
The filter maintains a six-dimensional state: position, velocity, and acceleration in two axes. Four times per second, it predicts where the boat should be based on physics — constant acceleration extrapolation. Once per second, when a new GPS fix arrives, it blends the prediction with the measurement, weighing each by its uncertainty.
High-accuracy GPS fixes pull the state more. Poor fixes pull it less. Between fixes, the prediction fills in the gaps. The result: position and velocity that update smoothly at 4 Hz instead of jumping at 1 Hz.
What this means for TTL: the speed and heading feeding into the TTL calculation are filtered, stable values — not raw GPS noise. The filter also tracks acceleration, so if you're decelerating into a tack, the state estimate reflects that rather than assuming constant speed.
The EKF is the single biggest improvement to TTL accuracy. It eliminates the jitter that makes raw-GPS TTL calculations unreliable, smooths out speed and heading noise, and fills the 1 Hz gaps with physics-based prediction. It's fully deterministic — no training data needed, no cloud dependency, works identically on iPhone, Apple Watch, and Android. This one layer closes roughly 80% of the accuracy gap between a phone and a dedicated 25 Hz instrument.
The Kalman filter assumes constant acceleration. That's a good model for a boat sailing in a straight line, but it breaks down in the scenarios that matter most: the final 30 seconds of a start approach, when sailors are adjusting speed, bearing away, luffing, and tacking.
Flying Start uses a 1D convolutional neural network that runs entirely on your device to predict what the Kalman filter gets wrong. It's trained on tens of thousands of simulated start approaches with realistic sailing physics — tacks, speed changes, current effects, GPS noise — and learns the patterns that simple physics models miss.
How it works: the model looks at a rolling 30-second window of your approach — speed profile, heading changes, acceleration pattern, distance to line, closing speed, and countdown remaining. It's seen thousands of simulated scenarios where it knows the actual crossing time, and it's learned what patterns the Kalman filter misjudges.
When it helps most: the model adds the most value in the final 15–30 seconds before a start, when you're actively manoeuvring. A sailor luffing to kill speed, then bearing away to accelerate — the Kalman filter predicts based on current acceleration, but the model recognises the pattern and anticipates the coming speed change. In test scenarios with tacks and speed changes, the ML layer reduces TTL error by 1–3 seconds compared to the Kalman filter alone.
When it doesn't help much: on a clean, straight-line approach with steady speed, the Kalman filter is already very accurate. The ML model doesn't add much in those situations — and it's designed not to. The correction is clamped to ±30% of the Kalman TTL, so the model can refine the estimate but can never produce a wildly wrong result.
Neither the Velocitek ProStart nor the Vakaros Atlas 2 uses machine learning for TTL prediction. They rely on kinematics: distance divided by speed. That works well with 25 Hz GPS, but it has the same blind spot — it can't anticipate manoeuvres. The ML layer is something software can do that hardware can't, because it improves over time as it learns from more data.
The current model is trained on synthetic data — computer-simulated race starts with realistic sailing physics. It's good, but simulations can't capture everything. The way a 420 approaches a start line is different from a J/70 or a Laser. Tidal patterns in the Solent are different from San Francisco Bay. Light-air starts look nothing like 25-knot windward starts.
Flying Start captures anonymised timing metrics from real race starts — filtered speed, heading, distance to line, and the actual moment of crossing — and uses this data to retrain the model. Every start, across every class and venue, makes the predictions more accurate for everyone.
You don't need to do anything. If GPS track recording is enabled (it is by default), your starts contribute to the training data automatically. The data is anonymised before training — no names, no locations, just the shape of the approach and the timing of the crossing. You can opt out at any time in Settings.
Hardware instruments can't do this. A Velocitek ProStart has no network connection and no way to improve its algorithms after it leaves the factory. A Vakaros Atlas 2 has connectivity, but its TTL calculation is pure kinematics — there's no learning loop. Flying Start's model gets better every season.
The base model learns from all sailors. But your starts have patterns unique to you — how aggressively you accelerate in the final 10 seconds, how much speed you lose in a tack, how early you begin your final approach.
Apple's Core ML framework supports on-device model updates via MLUpdateTask on iPhone. This means Flying Start can fine-tune the base model to your specific patterns after 20–30 starts — without sending any data to a server. The personalisation happens entirely on your phone.
After enough starts, your TTL predictions won't just reflect how an average sailor approaches a start line. They'll reflect how you approach it — your boat, your style, your tendencies.
On-device personalisation requires an iPhone. Apple Watch can run the personalised model (synced from the paired iPhone) but can't perform the on-device training itself. Android support depends on equivalent on-device training frameworks. Personalisation needs enough data — the first 20–30 starts use the base model before personalisation kicks in.
We're engineers. We believe in honest comparisons. Here's what a $1,000 instrument gives you that software alone can't replicate.
The Vakaros Atlas 2 achieves ~25 cm accuracy with 25 Hz dual-band GNSS and real-time differential corrections. Even with our Kalman filter, a phone GPS has a noise floor of 1–2 metres. For official, referee-grade OCS calling at championship level — where centimetres determine whether you start or sit out — dedicated hardware has the edge.
25 Hz means a new position fix every 40 milliseconds. Phone GPS runs at 1 Hz (one fix per second). Our Kalman filter interpolates between fixes at 4 Hz, but it's prediction, not measurement. In the final 3–5 seconds of a start, when things change fastest, higher-rate raw GPS has an inherent advantage.
A sunlight-readable 4.4" screen with Gorilla Glass and 100-hour battery is purpose-built for the cockpit. A phone in a waterproof case is good, but it's still a phone. Apple Watch on the wrist is arguably better for glanceability, but the screen is small.
Download Flying Start free. Upgrade to Premium for TTL, DTL, OCS alerts, Start Rating, Apple Watch, and everything else.