Case study · AI product
DuSu: building an AI English-speaking coach
DuSu is a voice-first AI English-speaking coach I built for Indian, mobile-first learners who know some English but freeze when they have to speak it. It is in live beta and runs at $0 a month: speech stays in the browser, and replies come from a free LLM chain that fails over in order of measured latency.

At a glance
- What
- Voice-first AI English-speaking coach
- My role
- Designed and built it end to end — product, frontend, backend, AI layer, deployment
- Status
- Live beta · installable PWA + Android app
- Running cost
- $0 — free LLM tiers and free browser speech
- Stack
- Single-file web app (vanilla JS) · FastAPI + one WebSocket · Postgres · PWA + Android TWA (Kotlin)
- AI
- Groq → Gemini → OpenRouter (OpenAI-compatible), with GitHub Models as a dormant last fallback
The problem
Knowing English isn’t the same as speaking it
Many learners — freshers preparing for interviews and campus placements, and speakers outside the metros — know enough English to read and write it, then freeze the moment they have to say it out loud to another person. Grammar drills and flashcards don’t train that. DuSu trains the thing people actually freeze on: speaking under pressure.
It had to work for people on budget Android phones and mobile data, which ruled out anything expensive to run or heavy to download.
What it does
Four ways to practise
- Talk
- A warm, open-ended English conversation that follows your interests.
- Interview
- An adaptive mock HR interview that ends with a scored report — grammar, fluency, confidence, communication, vocabulary and professionalism — plus a rewritten “better answer”.
- Learn
- Say a Hindi or Hinglish sentence; get the natural spoken-English version to repeat.
- Daily Talk
- A Hindi-first “close friend” check-in on your day that nudges you back to practice.
Around the modes: an AI-generated level test on first login (CEFR A0–B2), a 7-level roadmap with lessons, XP and streaks, and a memory of each learner — nickname, dreams and key moments — so DuSu behaves like one companion across every mode, not a lesson app.
Architecture
How DuSu is built
- 01
Browser: ears and voice
Speech-to-text and text-to-speech run on the device through the Web Speech API. Speech never leaves the browser — the wire carries text only.
- 02
One WebSocket, four modes
A single socket, /ws/interview, drives Talk, Interview, Learn and Daily Talk.
- 03
FastAPI backend
Auth, the level test and roadmap, interview scoring, the learner’s memory, and bring-your-own AI keys.
- 04
Free LLM chain with failover
Providers tried in order of measured latency — Groq first, then Gemini, then OpenRouter — with per-provider cooldown when one fails.
- 05
Postgres
Progress, the 7-level roadmap and each learner’s memory persist between sessions.
Decisions
Decisions and trade-offs
Speech in the browser, not on a server
Running speech-to-text and text-to-speech on the device keeps the cost at zero, keeps voices private — speech never leaves the phone — and keeps latency low. The trade-off: it needs a browser with the Web Speech API. A plain Android WebView doesn’t have it, so the Android app ships as a Trusted Web Activity instead.
A free LLM chain ordered by measured latency
Groq answers first (about 0.7 s measured), then Gemini, then OpenRouter’s free models as a late fallback. Every model in the chain was benchmarked with a live call, because a dead model isn’t just useless — it costs a full round-trip on every turn before the chain reaches one that answers. Real examples: Groq retired older models that then returned 404 on every call; a busier Gemini model was so often overloaded that retries burned about 25 seconds per turn, so DuSu uses flash-lite; and OpenRouter’s free tier rate-limits heavily.
Bring your own key
Users can plug in their own free Groq, Gemini or OpenRouter keys. DuSu builds the same chain, in the same order, from whichever keys are valid — so heavy users, like a team using it at work, don’t drain the shared quota.
Honest scores
Interview and confidence scores are labelled in the product as AI-estimated, not an exact measurement. Feedback that overstates its precision would undermine the confidence DuSu is trying to build.
Lessons
What I learned
- Free AI tiers change without notice — verify every provider with a real call before relying on it.
- Order a fallback chain by measured latency, not by reputation.
- Free hosting disks can be wiped on every deploy, so SQLite didn’t survive; persistent data moved to Postgres.
- For speaking practice, warmth and memory matter as much as correction.
Written by
Contact
Want an AI product like this built?
Tell me what you want to build, who it's for, and your timeline — projects are delivered by the Rana Brothers team.