EchoWise: I shipped an AI English-speaking app in 2 days with Agentic Engineering

Mon, 08 Jun 2026 12:00:00 +0000

TL;DR

In two days, I shipped an open-source, cross-platform desktop AI English-speaking tutor: EchoWise.

Built end-to-end with Vibe Coding / Agentic Engineering.

Why EchoWise

Almost every English-learning app on the market teaches you to study English: flashcards, drills, scores, leaderboards. But what I actually needed when I was learning English wasn’t more learning — it was more using.

Stop studying English. Start using English.

EchoWise is built on the opposite philosophy:

Communication before perfection — Imperfect English that connects beats perfect grammar that stays silent
Relationship before knowledge — Polite stranger on Day 1, easy old friend by Day 30. Your AI companion’s tone evolves naturally with the day count
Gentle nudges, never interruptions — The AI never lectures mid-reply. Scores sit quietly in the corner of your bubble — one glance when you want them
Local-first — Every conversation, recording, and AI voice playback stays on your machine. Nothing is uploaded

What EchoWise actually does

The whole product revolves around one core idea: a companion that grows with you.

🗣️ A companion that grows

You give them a name, pick an avatar, choose a voice, write a persona. EchoWise tracks the Day count and maps the relationship into 5 tiers: Stranger → Acquaintance → Familiar → Friend → Old friend. The AI’s tone follows naturally — polite stranger on Day 1, the kind of old friend who teases you and remembers what you said by Day 30.

🎙️ Voice-first conversation

Hold the mic to record, release to transcribe; type when you can’t talk. The AI’s reply comes back as a voice-message-style bubble — tap the waveform to play. Don’t want to listen? Click “Show transcript” for the text. The whole flow feels like a normal messaging app, not a drill exercise.

💡 Coaching that never interrupts

The most annoying anti-pattern is an AI that breaks mid-conversation to lecture you about “you should have used the present perfect here.” EchoWise refuses to do that.

The AI never mixes grammar lectures into its replies
After every sentence you say, a 1–100 score chip appears in the corner of your bubble, in 5 color bands:
- 🔴 Try again (< 50)
- 🟠 Getting there (50–59)
- 🟡 Clear (60–69)
- 🟢 Natural (70–89)
- ⭐ Native-like (90+)
Tap the chip to see original / better / native-like + a short note
Don’t tap? Keep talking. It never interrupts you

📈 Growth that actually means something

Not “how many words you reviewed today,” but:

Confidence trend — rolling average of your last 5 sentences, updates as you talk
5-band distribution — where do your sentences cluster
Auto-picked “best so far” and “worth revisiting”

Open the Growth view and see at a glance that you’re getting better. That positive feedback loop beats streak-counting any day.

🎨 Full appearance customization

Theme (system / light / dark), 4 fonts, 4 sizes, 6 preset gradient backgrounds — or upload your own image. Each companion can have their own avatar and background, too.

🔌 Multi-provider

OpenAI / Azure OpenAI in one click. Defaults to gpt-5 + gpt-4o-transcribe + gpt-4o-mini-tts. TTS voice instructions adapt dynamically to the companion’s persona — a lively companion gets a more energetic voice, a calm one gets a quieter delivery.

💾 Fully local

SQLite database + filesystem. Every conversation, every recording you made, every AI voice clip — lives on your machine forever, never uploaded to any server.

About the “two days”

The entire build was a Vibe Coding / Agentic Engineering workflow: I focused on product decisions and architectural judgment, the AI handled implementation and grind work.

My job:

Decide what features ship, how the UX should flow, how many tiers in the relationship arc
Judge whether the architecture holds, whether an abstraction is worth it, where bugs actually live
Write the PRD, iterate on UX, review PRs

The AI’s job:

Write React components, call Tauri APIs, write unit tests
Tune Vite / Cargo / Tailwind config
Edit GitHub Actions workflows, fix CI failures
Propose improvements, catch my mistakes

After two days, this is what the product looked like:

Metric	Number
Application code (TS + TSX + Rust)	~4,571 lines
Unit test code	~3,405 lines
Unit tests	281
Line coverage	94.5%
Cross-platform CI/CD	macOS + Windows, tag-triggered releases
Auto-update	tauri-plugin-updater with signature verification

Try EchoWise

GitHub: https://github.com/hujiulin/EchoWise
Download: Releases page has macOS .dmg and Windows .msi
Setup: Built-in OpenAI and Azure OpenAI presets — drop in your own API key and you’re talking

macOS note: the first time you open it you’ll see “Apple could not verify…”. That’s because I haven’t paid for an Apple Developer ID ($99/year) for notarization. Open System Settings → Privacy & Security → scroll to the bottom → click Open Anyway once, and you’re done forever. Full steps are in the README.

If you enjoy it, a ⭐ on the EchoWise repo is very welcome — it helps more people find it. Feedback, bugs, ideas — open an Issue or PR on GitHub.

Agentic Engineering on Coinidea's Blog