Sarah Aranda

The problem

I tutor an adult English learner at a beginner-intermediate level. We practice free conversation, often in a busy café and sometimes over video calls. In a real conversation, mistakes go by too fast to write down, and afterward she has no record of what to work on.

There are speaking apps already, but none were personalized to her: her level, her interests, and the mistakes she actually makes. So I built a prototype that listens to her side of the conversation, keeps her exact words, mistakes included, and turns them into practice she can come back to. I built it around one learner first, but it adapts to any learner's level, interests and mistakes.

The learner's name is blurred in the recordings.

Features

Free Talk is a single web page that runs in the browser on a laptop, with nothing to install.

• Prompt deckOver 50 conversation prompts built around her interests: food, travel, Pilates, and differences between Korea and the US. They mix formats, like role-plays, would-you-rather questions and "teach me" prompts. Each comes with follow-up questions, useful words, and the grammar it practices. Shuffle through them with one button.
• Editable topicsWhen she mentions a new interest, I add it and Claude writes eight fresh prompts for it. Topics she isn't into can be hidden.
• Live transcriptHer headset mic streams to a transcription service, and her words appear as she talks. Words that were hard to hear are marked, and I can fix misheard words before asking for feedback.
• Feedback in simple EnglishWhat she did well, her top recurring patterns, each correction with a short explanation, more natural ways to say things, and practice questions with answers. Explanations also come in Korean, her first language.
• Prompts aimed at her weak spotsAfter each conversation, she gets new prompts that push her to use what she got wrong, like past tense if she keeps missing it.
• Feedback notebookEvery conversation is saved, with patterns that keep coming up across sessions and a searchable list of every correction.
• Review flashcardsHer own mistakes become flashcards. She says or types the correct version, then marks how well she knew it, and cards come back after 1, 3, 7, 14 and 30 days.
• Voice partnerAn AI voice asks a question, listens to her answer, and replies with a natural follow-up. I have controls to pause it, have it repeat itself, slow it down, or tell it she's finished talking.
• Email and cleanupI can send her the feedback by email, delete a conversation, or remove a single correction the mic got wrong.

How I used AI

I used Claude Code to build the app, and the Claude API inside it to write feedback, voice replies and new prompts. I set the requirements based on how she and I actually practice, tested each version myself, and decided what to build next.

Choosing the services

• Deepgram to listen. Many transcription tools, including popular note-taking apps like Otter, quietly tidy up grammar, which would hide the very mistakes the app needs to see. Deepgram can keep her words exactly as spoken, including "um" and "uh." It also scores how sure it is about each word, so the app can mark unclear words and avoid flagging café noise as mistakes.
• Cartesia to speak. I wanted the voice partner to sound natural, not robotic, and to answer quickly so the conversation doesn't stall. Cartesia focuses on expressive voices and can slow its speech down, which helps a beginner-intermediate learner follow along.
• Claude to think. Claude writes the feedback, the voice partner's replies, and new prompts, all aimed at her level.

What the AI got wrong, and what I caught

1. Its tests passed, but recording didn't work.The AI tested the app in a browser with no microphone, so its checks couldn't catch the real problem. When I tried it, nothing recorded and no error appeared. The cause was my browser's privacy settings blocking the mic. Now the app always shows a clear message when recording fails.
2. It kept transcribing after I pressed stop.I noticed this when I tested it myself. Clicking the button again while it was still starting up launched a second, hidden recording that the stop button couldn't reach. The AI's tests had never tried an impatient double-click, which is exactly what people do.
3. The level was too high for her.When I ran a mock session on my own, the prompts and replies felt above a beginner-intermediate learner. That's the first thing I'll change before she uses it.

Decisions I made

• Only her voice. She wears a headset with its own mic, so only her speech is analyzed and the café noise stays out.
• Feedback she can absorb. Explanations are in simple English, and the report leads with her top three patterns instead of every small error.
• Conversation first, corrections later. In fluency practice, good teaching keeps the conversation going and saves most corrections for afterward, so the learner isn't interrupted mid-thought. The voice partner follows that: it repeats her sentence back the right way, the way a good tutor does, and the full corrections come at the end.
• No false lessons. In a noisy place, the mic can mishear a word, and the AI then flags a "mistake" she never made. I can remove any single correction, and it disappears from her review cards and emails too.
• The teacher stays in the loop. The app never sends an email on its own. It opens a draft in my email app, and I read it and press send myself.

Skills this shows

• Turning teaching practice into product requirements
• Choosing the right AI service for each job
• Testing AI output myself
• Voice interaction design
• Learning design
• Plain-language writing

What's next

This is a prototype. I've tested it in mock sessions on my own, and she'll try it in a real session soon. Before then, I plan to:

• Lower the level, or let the app adjust it automatically based on how she's doing.
• Go further with personalization. Topics already follow her interests. Next is a personalization page for her level, interests and goals, so every prompt that comes up is relevant and interesting to her.
• Add a privacy notice and consent step, so she knows how her voice and transcripts are used and agrees before recording.

Keep an eye out for the next version. After her first real session, I'll share what worked, what didn't, and what I changed in a follow-up post.

The problem

I teach English to adult learners and build self-study learning apps they use on their own. I wanted those apps to feel more joyful to practice with, and I didn't want to reinvent the interaction every time I started a new one. So I built a repository of interactive components for myself: a tested, documented library I can browse to pick the right interaction for each new app. It streamlines how I build every learning app that follows.

What I built

Five reusable components, each a single, lightweight web page that works on a phone:

• Bubbles: a timed game. Learners drag a sentence (or tap) into the bubble holding the phrase that completes it. Right answers get absorbed and the bubble grows; wrong ones bounce back.
• Pronunciation Accuracy: press and hold, read a sentence aloud, and see which words came through clearly.
• Grammar Check: learners write their own sentence and get mistakes marked in place, with suggested fixes.
• Progressive Disclosure and Section Jumper: structure pieces that reveal a lesson one step at a time and show learners where they are.

How I used AI

I used Claude Code to build and document the components, and Perplexity to research and refine the learner-facing wording. I acted as the product owner and designer: I set the rules, tested the results myself, and decided what shipped.

What the AI got wrong, and what I caught

1. It misunderstood the game at first.The first prototype missed my idea. I described the real mechanic, one sentence at a time with five fixed bubbles, and that version worked far better on a phone, because there's nothing to scroll past while dragging.
2. The sounds were out of sync.By ear, I noticed the reward sounds playing before the animation, one by almost a second. The AI's automated tests had passed because they only checked that the sound played, not when. We made it a standing rule: sounds fire from the animation itself.
3. It assumed a teacher would be there.One message told learners to check with their teacher. My products are self-study, so there is no teacher. I made that a permanent design rule: every limitation must come with something learners can do on their own.
4. Buttons were too small for thumbs.When I restated that everything must be mobile-first, an audit found four buttons the AI had already called finished that were under the 44-pixel minimum for a fingertip.
5. The microphone tool failed silently.When I tested it, it said "Didn't catch anything." The cause turned out to be the AI's own decorative sound meter competing with the speech recognizer for the microphone.

Decisions I made

• Privacy first. I asked whether learners' recordings were being saved anywhere. They aren't, and I added a plain-language note telling learners where their audio goes.
• No paid AI where it isn't needed. For grammar checking I chose a free grammar service over a paid AI model. In a test with five common learner errors, it caught all five.
• Feel matters. I auditioned ten candidate sounds and picked the two that felt upbeat and short without being childish.

Skills this shows

• Testing AI output myself
• Writing rules an AI can follow
• Designing for adult learners on phones
• Instructional design
• UX
• Plain-language writing

What's next

Next, I'm giving Bubbles a friendlier, more playful look using Lovable, a second AI tool. I'll check that the new version still meets my standards for dragging, sound timing, and phone use. Then I'll test it with learners on real phones and compare it with the original: what they enjoy, where they get stuck, and what I'd change. I'll share what I find in a follow-up post.

I'll also keep adding to the repository as new learning apps call for new kinds of interaction. Each new component goes through the same rules and checks before it's ready to reuse.