AI · 04
MindShift uses AI as infrastructure, not as a marketing surface. Here is the exact stack, the exact role of each model, and the privacy posture we don't compromise on.
The Stack
Anthropic's frontier model writes the scene graph: who appears, what they want, how the conflict escalates, what the participant must decide. Operating on Claude Opus 4.6 with extended thinking when the scenario requires multi-step planning.
Stable Diffusion XL through ComfyUI generates the visual environments — boardrooms, factory floors, hostile press conferences, terminations — at the fidelity Quest 3 demands. Asset packs are cached for sub-second load times.
Generated environments and assets are wired together inside Unity. Blender handles geometry cleanup and avatar rigging. The whole pipeline ships a scenario from prompt to headset in under 90 minutes.
Inside a session, every NPC speaks via a local Gemma instance running on the on-premise box. Response latency under 500ms. No cloud round-trip. No audio leaving the room. Local-first is a compliance feature and a UX feature.
The Loop
Claude Opus ingests your PTI history, your stated blockers, and the operator's session goal. It outputs a scene brief — characters, beats, success criteria.
ComfyUI generates fresh environment assets. Unity assembles. The headset receives a session bundle calibrated specifically to you.
Local Gemma runs every NPC in real time. The behavior tree adapts to your biometric stress signals — when HRV drops, the room intensifies.
All telemetry rolls up to a composite delta. Claude Opus drafts the integration note. The operator reviews. The participant receives a written report.
Compliance Profile
Dialog Latency
Local Audio
AI Act Ready
GDPR Compliant
Senior leaders cannot — and should not — be asked to stream their unfiltered biometric responses through commercial cloud APIs. Local-first is the only architecture that survives a serious privacy review. It also happens to be the only architecture that achieves under-500ms dialog latency on a Quest 3.
Ethics
Your biometric data feeds your PTI delta. It is not used to train any third-party foundation model. The Atlas is anonymized at ingestion and used only to personalize future scenarios.
Claude Opus writes scenarios; it does not give you advice in-session. The reflection conversations are with humans — operators and consulting psychologists — because identity work requires a person on the other end.
The protocol is the IP. The models behind it are commodity. When a better local model than Gemma ships, we swap it. When Claude Opus 5 ships, we re-prompt. Lock-in is not the architecture.
Every participant receives a written description of which AI components touched their session, what data they touched, and where that data lives.
Why This Matters
A traditional simulation studio would charge six figures per bespoke scenario. With our pipeline, fresh personalized scenarios cost minutes of compute and a single operator review. That is what compresses the protocol price to something a head of L&D can put on a budget line — and that is what makes “personalized at scale” an actual operating reality rather than a slide deck claim.
Get Early Access
Join the waitlist to see the loop in action — including a sample scenario brief and PTI report from a live pilot.