Valoa AI - Learn any languages

Inspiration

The inspiration for Valoa AI came from a major gap in the language learning market: the “Fake Fluency” trap.

Most traditional platforms rely on scripted, predictable scenarios where users can pass multiple-choice tests but still struggle during spontaneous real-world conversations. Many also rely on “Black Hat” gamification, such as streaks designed to prevent point loss, which often leads to sharp engagement drops after about two weeks.

We wanted to shift the focus from “teaching a language” to “enabling performance.”

Our goal was to create an Atomic Unit of Practice: a simulator that moves beyond rote memorization and helps adult learners build genuine communicative competence through active language production.

What It Does

Valoa AI is a language performance simulator that uses Task-Based Language Teaching (TBLT) to help users accomplish real communicative goals, such as:

• Negotiating a business deal
  
• Resolving a workplace conflict
  
• Participating in professional discussions
  

Key Features

Immersive Missions
Unscripted roleplays with live AI characters that respond with near-zero latency.

Lexical Harvesting
An AR-enabled mode that lets users use their camera to identify real-world objects and add them to a personalized vocabulary inventory.

The Supertutor
A smart interface using modular building blocks to help learners construct complex sentences without facing a blank page.

Mission Debrief
A detailed performance breakdown analyzing growth in Complexity, Accuracy, and Fluency (CAF).

How We Built It

Valoa AI is designed as an AI-native application powered by the Gemini ecosystem.

Using Gemini models and Gemini Live, the platform enables real-time conversational interactions with minimal latency. This allows learners to engage in dynamic spoken roleplays that feel natural and responsive.

Our system is built around a multi-agent architecture, where specialized AI agents collaborate to create immersive learning sessions:

• Persona Agent: plays the role of the conversation partner
  
• Critique Agent: analyzes the learner’s performance and provides detailed feedback
  

Gemini Live powers the real-time conversational layer, enabling natural voice interaction and continuous dialogue during missions.

Unlike platforms relying on extrinsic point-scoring, Valoa uses “White Hat” gamification, emphasizing:

• personal impact
  
• creativity
  
• the Epic Meaning of the learner’s journey

Challenges We Ran Into

Working with real-time conversational models introduced several technical challenges:

Live API
It was confusing because Google provides Gemini Developer API and Vertex AI Gemini API, and multiple SDKs, namely GenAI SDK, Firebase Logic which have slightly different capabilities. A simple prototype was quickly done with AI Studio. However, when we started implementing with Kotlin Multiplatform, we did try with both Gen AI SDK (with ephemeral token) and Firebase Logic (which needs wrappers to make it works across Android, iOS, Web).

Transcription Errors
Occasionally, voice input was transcribed into the wrong language, interrupting the flow of a session.

Feedback Depth
Standard AI models often produced shallow pronunciation feedback, requiring additional work to generate meaningful phonetic analysis.

Session Management
Proactively ending real-time sessions with Gemini Live proved tricky to handle smoothly.

Cost Efficiency
Although the models support Voice Activity Detection (VAD), we needed to implement our own handling to maintain sustainable operational costs.

Accomplishments We're Proud Of

We successfully moved beyond the laboratory phase and into real-world testing.

Currently, a group of early users is actively testing the platform and providing valuable feedback. The most encouraging outcome is that users report greater confidence speaking the language, validating our focus on active performance rather than passive study.

What We Learned

Developing with real-time conversational AI systems is uniquely challenging.

Balancing low latency with high accuracy requires constant optimization, especially when aiming for human-like conversational response times.

We also learned that security and privacy are critical. Because the platform handles personal learning data and aims to be a professional-grade tool, protecting users and system resources is a top priority.

Where we are now

We are working multiple dimensions to improve the app:

• Improve the real-time learning experience
• Currently support the “roleplay” mission type
• Plan to expand with additional alternative mission formats
• Improve stability across multiple platforms through testing
• Prepare and release updated versions to app stores

What's Next for Valoa AI

Our roadmap focuses on expanding the ecosystem and learning experience.

Personalized Learning Paths

Adaptive curricula tailored to each learner’s professional background, goals, and interests.

Expanded Scenarios

Beyond traditional roleplays, future missions will include:

• investigative challenges
  
• creative design tasks
  
• strategic logic puzzles
  

Community-Driven Content

We are building a community ecosystem where users can create and share their own:

• Missions
• Collections

These will be inspired by real-world situations users encounter in their daily lives.

Built With

• cloud-run
• cloud-sql
• firebase-logic
• go
• google-cloud
• google-genai-sdk
• kotlin
• kotlin-multiplatform
• silero-vad
• vertex