Improving Mileage Tracking Accuracy from 50% to 95%

A founder details two mechanisms, physics-based filtering and dead reckoning, that reportedly improved a mileage tracking app's accuracy from 50% to 95% by challenging unreliable GPS data.

A mileage tracking app, initially reporting 50% trip accuracy, improved to 95% by teaching phones to 'disbelieve' faulty GPS signals. The founder, posting on dev.to, describes a system that filters out physically impossible readings and fills data gaps. This approach addresses common GPS failures in environments like tunnels and urban canyons, where signals are frequently distorted or lost.

GPS: A Witness with a Concussion

The core problem, as the founder frames it, is that GPS often provides confident but incorrect data. Three scenarios commonly degrade accuracy for mileage tracking:

• Tunnels: Loss of satellite view causes the receiver to report the last known position for extended periods, then 'snap' to a new location upon reacquisition. This creates an apparent, impossible jump in speed.
• Urban Canyons: Signals bounce off tall buildings (multipath interference), leading the phone to average reflections and report a position a street or two away, despite high confidence.
• Parking Garages: These environments typically result in intermittent or entirely absent signals, producing data gaps or noise.

These are not edge cases; the founder describes them as daily occurrences. The tactical shift was from attempting to acquire 'better' GPS data to actively rejecting 'bad' GPS data.

Fix 1: Physics-Based Filtering

The first mechanism involves a simple guardrail: any reported movement that implies physically impossible speed is discarded. The founder provides a Kotlin function isImpossible that calculates speed between consecutive GPS fixes and compares it against a maxPlausibleSpeed for the current travel mode (e.g., walking, cycling, driving).

This filter directly addresses errors like the user who was reported to be traveling at 400 kmph while stopped at a red light. By dropping data points that violate basic physics, the system eliminates the most egregious errors. The founder states this single guardrail significantly reduced embarrassing errors by refusing to accept implausible data.

Fix 2: Dead Reckoning for Data Gaps

Rejecting bad GPS points creates gaps in the trip data. The second mechanism, dead reckoning, aims to bridge these holes. While the source material cuts off before detailing the implementation, the principle involves estimating current position based on a previously determined position, combined with estimates of speed and direction. For mileage tracking, this means inferring movement through tunnels or areas of poor signal based on the last reliable data and the expected trajectory. This prevents a temporary loss of signal from creating a permanent hole in the trip log, maintaining continuity even when direct GPS input is unavailable.

What We'd Change

The physics-based filtering is a foundational step for any location-aware application, but its effectiveness relies on accurate maxPlausibleSpeed thresholds. For broader application, these thresholds would need to be dynamically adaptive or user-configurable, especially for niche use cases like high-performance vehicle tracking or specific industrial applications where speeds might genuinely exceed typical driving limits. The current implementation, tied to general 'mode' (walking, cycling, driving), assumes standard vehicle dynamics. An edge case, such as a user on a high-speed train, could see valid data points incorrectly filtered.

The source does not detail how the 'mode' (walking, cycling, driving) is detected. If this is a manual input, it introduces user friction and potential for error. If it's an automated heuristic, its accuracy becomes another critical dependency. For robust commercial deployment, automatic and reliable mode detection, potentially leveraging accelerometer or gyroscope data, would be necessary to ensure the maxPlausibleSpeed is always appropriate. Furthermore, while rejecting bad data improves accuracy, an overly aggressive filter could lead to incomplete trip logs, which might be acceptable for expense tracking but problematic for other applications requiring full data fidelity.

Landing

The founder's approach to improving mileage tracking accuracy highlights a critical lesson in data processing: not all inputs are equally reliable. By shifting focus from signal acquisition to signal validation, the app moved from 50% to 95% accuracy. This tactic underscores that for applications where data integrity is the product, a robust filtering layer that 'disbelieves' physically impossible readings is as crucial as the raw data collection itself. It prioritizes plausible continuity over raw, but often flawed, sensor output.

The investor read

The reported improvement in mileage tracking accuracy from 50% to 95% signals a critical value proposition in a niche but essential B2B/B2C SaaS category. For investors, this level of accuracy directly impacts user trust, retention, and the defensibility of a product in a market where precision for expense and tax purposes is paramount. The underlying technical approach—prioritizing data validation over raw sensor input—is a scalable strategy applicable to other location-dependent services. While the founder's claims are unverified, a product demonstrating this accuracy could command higher subscription tiers or integrate more effectively into enterprise expense management platforms. The challenge for investable products in this space lies in proving the claimed accuracy with verifiable, independent audits and demonstrating market share against established players like TripLog or Everlance.

Pull quote: “By shifting focus from signal acquisition to signal validation, the app moved from 50% to 95% accuracy.”