Language practice with an AI conversation partner. Speak naturally, review detailed feedback, and bring what you learn into your next attempt.
Watch it in actionPracticing a language takes more than finding the right words. It means listening, responding, and trying again. Oral Loop brings that cycle into one space, with an AI partner for practice and a separate review afterward.
A central part of my work was choosing the right technology for each task: a language model for CEFR-based evaluation and speech-to-text that captures the details of spoken language. I also shaped the conversational flow and interface.
A French practice session, from setting up the conversation to reviewing the feedback.
I chose a language model to evaluate performance against Common European Framework of Reference (CEFR) criteria and turn evidence from the session into specific feedback. Evaluation happens afterward, so the AI can stay in its role as a conversation partner during practice.
A key selection criterion was whether speech recognition could faithfully retain fillers, repetitions, and hesitations in the transcript. Those details provide evidence for assessing fluency and identifying what to practice next.
While studying French, I found that inaccurate speech recognition could distort feedback. My background in language teaching and the CEFR helped me identify transcription quality as a prerequisite for useful assessment.
I defined the product hypothesis, technology selection criteria, conversational flow, CEFR-based feedback and interface. I directed iterative implementation with AI coding agents.
Prioritize faithful transcription over the lowest API cost; choose a separate model for evaluation; keep feedback until the end so practice stays close to an exam experience.
AI product thinking: translate domain knowledge into requirements, evaluate technology against the task and prioritize the dependency that determines the usefulness of the result.
Next: compare transcripts against human-reviewed recordings and assess feedback with language teachers, then measure completion, cost per session and repeat use. This is an independent prototype; CEFR-based feedback is not an official certification. Before production, I would work with a development team to review the code, architecture, security and maintainability, and prioritize any required changes.
Speech-to-text APIs · LLM-based evaluation · CEFR criteria
I defined requirements, experience and model-selection criteria, and reviewed iterative results. AI coding agents generated the implementation code under my direction.
