Home / Work / Brain Mosaic
Brain Mosaic · AI product development

Giving a Unity game a safer, editable AI conversation layer.

For Learning Chronicles' Brain Mosaic project, Kydos built an admin platform that lets the team author, test and serve AI-driven NPC scenarios for a children's social-skills app without embedding scenario logic directly inside the game.

The problem

Scenario design and AI evaluation needed a manageable control layer.

The product team needed to edit scenes, objectives, characters and scoring rules, test them against the real AI and make the resulting behaviour available to Unity through a stable API.

Kydos role

Product architecture, dashboard development, AI evaluation and API delivery.

Kydos handled the authoring/testing layer, Azure AI integration, production APIs, safety logic and deployment workflow.

2Unity-facing API endpoints live
6Built-in AI reply safety checks
81Automated tests documented in the delivered project
LiveDashboard and Azure AI evaluation in production
01 · Author

Scenario control outside the game build.

The dashboard lets authorised users create and edit the scene setup, objective, NPC behaviour and scoring configuration that define each conversation.

Create

Author the setup, objective, character and exchange rules that make up a scenario.

Test

Run a scenario against the real AI and see the score and NPC reply before it is used in the game.

Serve

Expose scenario data and evaluation through REST endpoints that Unity can call during gameplay.

02 · Runtime flow

Child response in, scored NPC reply out.

Speech-to-text and text-to-speech stay on the Unity side. The dashboard handles the scenario rules, AI evaluation, score and written NPC response.

Child speaks

The player responds aloud inside the game.

Unity sends

The transcribed response is sent to the dashboard evaluate endpoint with the scenario ID.

Dashboard scores

Scenario rules and Azure AI evaluate the response and create the NPC reply.

Unity voices

The scored reply returns to the game, where Unity handles text-to-speech.

03 · Safety

Every AI reply passes six checks before use.

A response only becomes shippable if it passes all configured checks. Otherwise the system uses a safe fallback.

Banned terms

Scenario-specific prohibited language is blocked.

Profanity

Swearing is rejected.

Length

Replies are kept suitable for a young child.

Reading level

Language is checked for age suitability.

Clinical claims

Medical or therapeutic claims are prevented.

On-topic

The reply must stay aligned with the scene and objective.

04 · Production state

Live and serving the game.

The delivered project documents the dashboard reading from the cloud database, two Unity endpoints verified from external machines, real Azure AI evaluation in production, a login gate for the authoring dashboard and in-app guidance across the editor.

Need an AI workflow that has to be editable, testable and safe?

Talk to Kydos