Uber's DragonCrawl: Generative AI for High-Quality Mobile Testing
Original post DragonCrawl: Generative AI for High-Quality Mobile Testing
Article Summary
Ali Zamani, Anam Hira, and Juan Marcano from Uber reveal how they built DragonCrawl, an AI that tests mobile apps like a human would—and it's already caught 10 high-priority bugs in three months.
Uber's Developer Platform team faced a massive testing challenge: 3,000+ simultaneous experiments, 50+ languages, and engineers spending 30-40% of their time maintaining brittle test scripts. Their solution? Train a language model to navigate apps with human-like intuition, adapting to UI changes without manual updates.
Key Takeaways
- DragonCrawl achieved 99%+ stability with zero maintenance across 85 cities
- Uses MPNet model with 110M parameters, 1000x smaller than GPT-3.5
- Blocked 10 high-priority bugs while saving thousands of developer hours
- Automatically retried going online for 5 minutes in Brisbane, mimicking real user persistence
- Successfully handles adversarial cases by restarting apps when payment methods fail
Uber built a goal-oriented AI testing system that works across 85 cities and 50+ languages without maintenance, fundamentally changing how they ensure app quality at scale.
Continue reading
Posts on X
Feb 26, 2026
More from Uber
Real-Time Analytics for Mobile App Crashes using Apache Pinot
Uber
Nov 2, 2023
Also on iOS
How the Xcode 27 MCP can lead to higher token costs when navigating the Simulator
SwiftLee
Oct 26, 2026
This page summarizes “DragonCrawl: Generative AI for High-Quality Mobile Testing” by Uber, published Apr 23, 2024. All credit for the original work goes to the authors.
