MTA Tests AI TrackInspect Pilot to Detect Defects

The Metropolitan Transportation Authority (MTA) announced a pilot program with Google Public Sector called TrackInspect that retrofits standard Google Pixel smartphones onto R46 subway cars to capture subtle vibrations and sound patterns, per an MTA press release. "By being able to detect early defects in the rails, it saves not just money but also time - for both crew members and riders," said Demetrius Crichlow, President, New York City Transit, in the MTA release. The release and Google public sector materials describe that sensor data is sent in real time to cloud systems where AI and machine learning generate predictive insights and surface locations for physical inspection.

StateScoop reported that an initial TrackInspect pilot, conducted on six cars of the A line from September 2024 to January 2025, collected 335 million sensor readings, 1 million GPS locations and 1,200 hours of audio, according to results presented by the Google team. The MTA press release and Google Public Sector blog state that the pilot pairs those sensor streams with the agency's existing track information and uses generative AI for natural language assistance so inspectors can query maintenance history and protocols conversationally. Track inspectors perform physical verifications of flagged locations and provide feedback intended to improve the model, per the MTA and Google materials.

Programs that combine low-cost, off-the-shelf sensors with cloud ML and human-in-the-loop verification follow an applied ML pattern now seen across transport networks. In production settings, that pattern raises familiar engineering challenges for practitioners: noisy sensor fusion across mobile platforms, GPS accuracy in underground environments, labeled-data sparsity for rare failure modes, and the need for robust on-device preprocessing and secure telemetry pipelines. Teams deploying similar systems typically invest heavily in data quality controls, event deduplication, and retraining pipelines that incorporate inspector feedback.

PYMNTS reports that a 2025 State Comptroller report found roughly 82% of New York City's 2.7 million trains ran on time and attributed nearly half of that year's delays to infrastructure and equipment problems. That public-statistics backdrop explains why transit agencies are piloting predictive approaches rather than relying solely on manual inspections. Industry deployments in other cities, such as Singapore's JARVIS platform described in coverage, illustrate a broader move toward consolidating long-run operational and engineering records to prioritize limited maintenance windows.

Indicators observers should follow include published pilot evaluation metrics (false positive and false negative rates on flagged defects), how the teams handle underground localization and timestamp alignment, whether the MTA expands beyond the initial A line and six cars, and how inspector feedback is operationalized into model updates. Reporting that includes model performance numbers or publicly released datasets would materially change practitioner assessment of TrackInspect's readiness for scale.

If your team is building similar predictive-maintenance pipelines, prioritize rigorous end-to-end telemetry, strategies for labeling rare events, and tooling to integrate human verification into continuous training loops. Observability on sensor health and simple baselines for anomaly detection will be critical before investing in more complex generative or deep-learning layers.