As I delved into the May 15, 2024 article, "3 Interesting Engineering Problems You Could Only Solve at Duolingo" by the Duolingo Team, I found myself captivated by the unique technical challenges that Duolingo’s engineering team faces. With over 500 million users worldwide, Duolingo is not just an app—it’s a platform committed to making world-class education universally accessible. Achieving this ambitious goal demands constant innovation and problem-solving at a scale few companies encounter. In this retrospective, I’ll explore three key engineering challenges that Duolingo has tackled, each one showcasing the intricate balance between technical ingenuity and the mission to educate.
Ensuring a High-Quality Experience for All Learners
One of the most pressing challenges Duolingo faces is ensuring that all learners, regardless of their device or internet quality, have a seamless and high-quality experience. This is particularly crucial in emerging markets, where users often rely on less performant devices and deal with unreliable internet connections. The team identified that the app's startup time, especially on Android, was a significant pain point.
To address this, Duolingo engineers utilized system traces—detailed records of the app's execution process—to pinpoint key performance bottlenecks. By breaking down the startup process into critical steps and identifying areas where delays occurred, they implemented optimizations like delaying non-essential processes until after the home screen loaded. This approach led to a 40% reduction in startup time, making the app much more accessible for users on older devices.
Scaling Personalization for Learners
Personalizing practice sessions for millions of learners is no small feat. Duolingo’s Birdbrain system, which tailors practice sessions based on each learner’s proficiency, is central to this effort. However, the process of personalizing learning requires extensive A/B testing of new machine learning models, which initially doubled storage costs and API latency.
To overcome this, Duolingo engineers optimized their system by writing data less frequently to the database, specifically DynamoDB. By buffering changes in-memory and enforcing a Least Recently Used (LRU) policy, they significantly reduced the number of database writes, thereby lowering storage costs. This optimization allowed Duolingo to run simultaneous A/B tests for multiple models efficiently, reducing the cost of testing by 50% and enabling the introduction of new features like personalized vocabulary practice.
Making English Certification Testing More Accessible
Duolingo’s mission to make education accessible extends to English certification through the Duolingo English Test (DET), a more affordable and accessible alternative to traditional tests like TOEFL. However, ensuring that the DET is both accurate and free from bias posed a significant challenge, especially given the potential for human error in online proctoring.
To address this, Duolingo developed a system of AI-assisted human proctoring. This system uses computer vision models to detect prohibited items, monitor eye gaze, and ensure the visibility of test-takers’ faces during the exam. These models alert human proctors to potential issues, allowing for a more efficient and consistent evaluation process. The AI system not only helps mitigate bias but also reduces the overall cost of administering the test, making it five times less expensive than its competitors.
Conclusion
Duolingo’s approach to solving these unique engineering problems highlights the intricate interplay between technical innovation and the company’s mission to democratize education. By improving app performance, scaling personalized learning, and ensuring fair and accessible English certification, Duolingo continues to push the boundaries of what is possible in educational technology. As the platform evolves, so too will the challenges it faces, offering new opportunities for engineers to innovate and contribute to a truly global impact.
Top comments (0)