LIVR AI Avatars | Responsive VR Rehabilitation System for the Singapore Prison Service

What we delivered

• Hyperreal AI avatars with dynamic emotional response
• Fully adaptive, interaction-led VR scenarios
• On-device AI, ~1s response latency, full data isolation
• Built on top of an existing Unreal Engine programme

The challenge

The project addressed a well-documented limitation in immersive training design: scripted experiences deliver impact on first use but offer diminishing returns over repeated sessions. The SPS had an existing VR programme consisting of 11 fixed scenarios built in Unreal Engine, aimed at preparing incarcerated drug users to resist the social pressures that trigger relapse. The scenarios were effective; the problem was that they were finite.

A person who has been through a scenario knows what is coming. The resistance they are supposed to be building is, in part, being built against the programme itself rather than the real-world triggers it is designed to simulate. LIVR was brought in to solve that problem.

The solution

LIVR AVATARS is a fully immersive, interaction-led AI VR avatar platform that responds dynamically to user inputs, verbal, non-verbal and behavioural, generating scenarios that are genuinely unique every session. Rather than following a script, therapists define scenario frameworks and guardrails via prompt and the AI steers the immersive experience from within those parameters. The system is designed to desensitise users to real-world societal triggers through repeated, varied, high-fidelity stress testing in a controlled virtual environment.

What that means in practice is this: a person sits in a VR headset and is confronted by a hyperrealistic avatar who attempts to pressure them into relapsing. Whatever they say, the AI finds a new angle. It does not loop, it does not repeat itself, it reads the user and adapts with nuance. The session is different every time because the user is different every time, in their mood, their responses, their resistance levels. The system meets them there.

Key technical features

Hyperreal avatars that significantly increase the emotional depth and immersive quality of each session. The fidelity here is not cosmetic. The more convincingly human the avatar, the more genuinely the user's stress response is activated and the more meaningful the desensitisation process becomes. A cartoon character cannot replicate peer pressure the same way an emotionally intelligent photorealistic one can.

Fully adaptive interaction in which the central avatar responds in real time to whatever the user says or does, continuously finding new vectors to apply social pressure. No two sessions are identical. This directly addresses the core limitation of all current-generation training programmes: the experience retains its impact across repeated use because it never presents the same challenge scenario twice.

Localised AI running entirely on device, eliminating the need for network connectivity, reducing response latency to approximately one second, and ensuring complete data isolation. In a government (and specifically a custodial) setting, security and data sovereignty are non-negotiable.