LinguaCoach

AI

Fix only your weak points — while you speak.

Engage in fluid AI conversations while a dedicated side panel flags grammar and phrasing errors in real time.
Every mistake is logged to your personal database, feeding a diagnostic dashboard designed for targeted review.
(Spanish version includes DELE-compliant, LLM-driven grammar modules.)

Next.js 15TypeScriptSupabaseOpenAI API

Try demo GitHub ↗

01Problem → Solution

PROBLEMSOLUTION

PROBLEM 01

Talk to the AI and all you get back is bland, safe replies

The AI stays stuck as a "safe tutor," and no real depth ever emerges. The cause is neglecting prompt design and model selection.

SOLUTION

Manufacture immersion itself through model selection and persona strategy.

PROBLEM 02

Feedback that vanishes on the spot

Grammar mistakes and corrections disappear the moment they appear, so weak points never accumulate. The continuity of learning is severed.

SOLUTION

Accumulate weak-point data automatically, without the user ever noticing.

PROBLEM 03

Language schools cost a fortune, yet you can't move at your own pace

High tuition comes bundled with one-size-fits-all pacing, so you can't learn systematically at your own speed. The cost structure itself is a barrier to entry.

SOLUTION

Turn the DELE syllabus into a product and replace the language school.

Content breadth and vocabulary features were deliberately deferred. The core was fixed on three things: immersion, accumulation, and systematic learning.

02Key features

01

Real-time grammar feedback

Automatic grammar checkEvery message shows errors, suggested fixes, and a three-level rating
Multilingual explanationsSwitch between Japanese, English, and Spanish
History lookupPast feedback in one click

DESIGN RATIONALE

Recognizing errors without stopping the conversation proved the most effective way to make them stick. Leaving the explanation language up to the user prevents drop-off caused by limited reading ability.

02

Error insights and on-demand focused practice

Weakness visualizationTallies mistakes by type and ranks them
Weakness-focused chatStart practice from any category in one click

DESIGN RATIONALE

I decided the analysis shouldn't end at "just looking." Linking discovery straight to review in one click cuts out the time lost to aimless free chat.

03

DELE-aligned grammar learning mode

Structured DELE syllabusLearn B1 and B2 grammar systematically through AI dialogue
Progress trackingSupports resuming mid-session and self-scored results
Supported languageSpanish only for now

04

A dictionary and inline vocabulary book, all on one screen

One-tap translate and saveSelect an unknown word, get a translation pop-up, save instantly
Manage from the dashboardFilter by part of speech and sort by review order
Shares data with LinguaGymSaved words flow into GYM's repetition drills too

DESIGN RATIONALE

Switching to a dictionary or another app causes drop-off, so I made select-to-save the shortest possible step. On top of that, the vocabulary data is shared with LinguaGym, creating a path where words you meet mid-conversation get cemented through GYM's repetition drills.

03Tech stack

Next.js 15

App Router + Server Components

TypeScript

End-to-end type safety

Supabase

Postgres + Row Level Security

OpenAI API

GPT-4o — 会話生成・文法チェック・チューター

04Why I built this

Learning Spanish in Barcelona quickly exposed the limits of existing options: apps felt formulaic, and schools lacked flexibility.

I wanted a tool I'd actually stick with.

Language apps fall into repetitive AI response loops and offer throwaway feedback, making recurring mistakes hard to catch. Schools offer structure, but their fixed pace wastes time on concepts you've already mastered.

Experiencing these gaps firsthand shaped a single, unified solution:

Unscripted conversation
Passive feedback tracking
Self-paced, structured learning

— The Outcome —

Just chat naturally. Your error patterns and vocabulary log automatically in the background—no breaking your flow for corrections, and no sitting down for dedicated study sessions.

05Engineering decisions

01

Set Gemini as the default model and secured immersion through purpose-specific personas

PROBLEMExisting language AIs stay locked as a "safe tutor," and the conversation goes nowhere

I weighed the trade-offs of token cost, response speed, and context resolution. Early on I implemented automatic model switching based on how complex the context was, using Mistral and Llama in the low-cost tier. In practice, though, their response stability and cost efficiency fell short of expectations and didn't justify the complexity of switching, so I reworked it to fix Gemini as the default. Treating conversational immersion itself as the axis of differentiation, I built it to switch personas according to the user's purpose.

Dynamically optimizes tone and vocabulary level for each stage of learning

02

Ran error processing asynchronously from the conversation, designing a UX where weakness analysis piles up before you notice

PROBLEMBuilding feedback collection into the conversation flow directly erodes immersion

I made error processing run asynchronously from the main conversation flow, so detailed weakness data accumulates without the user ever being aware of it.

Never interrupts the conversation, translating directly into higher retention

03

Turned the DELE syllabus directly into features, replacing the language school with a product

PROBLEMThe structural problem of wanting to learn systematically while language schools offer poor value

I structured the international-standard DELE curriculum as features. Rather than stopping at delivering material, it integrates AI dialogue, progress tracking, and session resuming.

A scalable design that extends to other languages and certifications with the same structure (currently Spanish B1 and B2)

Try it out

No signup required — guest mode available

Try demo GitHub ↗

Next to validate: whether the weakness dashboard's at-a-glance overview works as a starting point for review

← LinguaGym ↑ Back to Work TubeRefine →