Design an AI Speaking Partner for Duolingo: Roleplay Chats with Live Corrections

You are a product manager at Duolingo. How would you use generative AI to build conversational agents that let learners practice a language through live, back-and-forth conversation and receive feedback as they go? Explain the goal, the learners you would build for, how the experience would work, how you would make the AI reliable, and how you would measure success. ```hint Feedback is the product A chatbot that talks is easy to demo; the hard choices are what the agent corrects, when it corrects it, and how it stays at the learner's level. Settle those for a specific group of learners before discussing models. ``` ### Clarifying Questions - Which learners is this for: beginners, learners in the middle of a course, or advanced speakers looking for fluency? - Should practice be spoken, typed, or both? - Is the feature meant for the paid subscription, for free users, or for both? - Is success defined by learning outcomes, engagement, subscription revenue, or a combination? ### What a Strong Answer Covers - A goal tied to Duolingo's mission and business, and why generative AI is the right tool for it - A chosen learner segment with concrete pain points in speaking practice - An experience design that keeps conversations at the learner's level and gives feedback without breaking the flow - AI quality, safety, latency and cost, including the risk of wrong corrections and weaker performance in some languages - Metrics that measure learning, not only usage, with guardrails and a staged rollout ### Follow-up Questions - How would you show that learners actually speak better, not just that they use the feature? - The agent sometimes corrects sentences that were already right. How do you detect and limit this? - How would you keep inference cost viable if free users get access?