How to Use AI Assistants in Spatial Computing with CLM-8B and Jev

Introduction: Why AI assistants are essential for spatial computing

Spatial computing is revolutionizing the way we interact with 3D data, digital twins, and immersive environments. AI assistants bring the power of natural language and structured reasoning directly into these contexts, enabling developers and users to create, modify, and query virtual spaces with simple prompts.

By 2026, the ecosystem is evolving rapidly: Contrastive-LM’s CLM-8B accelerates action evaluation, Jev offers typed decisions, and Gemini 3.8 Flash TTS provides real-time voice for fully spatial interfaces. This practical guide demonstrates how to integrate these tools to build responsive and space-aware AI assistants.

How to build a spatial AI assistant with CLM-8B

CLM-8B is an open-source System One model that evaluates candidate actions against a spatial state without generating text. It’s ideal for rapid feedback loops in AR/VR applications.

Quick installation

pip install contrastive-lm
from contrastive_lm import CLM8B
model = CLM8B()

Spatial evaluation example

Imagine moving a 3D object in Unity. Our AI assistant determines whether the proposed movement stays within bounds.

def check_movement(current_pos, proposed_pos, bounds):
    score = model.score(current_pos, proposed_pos, bounds)
    return score > 0.7  # action allowed

This snippet illustrates a compact decision loop: the model returns a confidence score, and our assistant reacts immediately, ensuring a smooth user experience.

Prompt engineering for spatial interfaces with Jev

Jev is a typed System One model that excels at making non-textual, structured decisions. It uses a type-safe syntax to express spatial constraints.

Building a prompt for Jev

import jev

prompt = """
# Goal: place a cube at (x=2, y=0, z=3) within a 10x10x10 grid.
# Constraints:
# - x ∈ [0,9]
# - y ∈ [0,4] (limited height)
# - z ∈ [0,9]
# Return a valid movement plan as an array of steps.
"""

plan = jev.run(prompt, schema=MovementPlan)
print(plan)

The result is a typed object that can be displayed directly in a game engine, ensuring every action meets user-defined constraints.

Enabling voice conversation with Gemini 3.8 Flash TTS

When an AI assistant operates in a three-dimensional space, voice adds a layer of naturalness. The new Gemini 3.8 Flash TTS models offer optimized synthetic voices for scripts directly within game engines.

Integration with Unity

import google.gemini

voice = google.gemini.GeminiTTS(model='gemini-3.8-flash-tts')

def speak_instruction(text):
    audio_clip = voice.generate(text)
    # Send the clip to Unity's audio system
    unity_audio_source.clip = audio_clip
    unity_audio_source.Play()

Now you can say, "Rotate the camera 45 degrees," and the assistant will play a voice alert and execute the corresponding action, all in real time.

Workflow example: Collaborative guidance in augmented reality

• Input:The user points the device at a wall and asks, "Show the floor plan of this level."
• Processing:The spatial model (CLM-8B) validates the camera position; Jev plans the floor plan rendering; Gemini TTS provides voice instructions.
• Output:The 2D floor plan appears overlaid on the real environment, and a voice says, "Floor plan displayed."

Best practices for optimal results

• Use Jev's typed constraintsto prevent actions outside the area of interest.
• Optimize CLM-8B scoreswith real-time feedback to improve responsiveness.
• Leverage Gemini's voice synthesisfor brief contextual cues to keep user engagement high.
• Create a centralized prompt hubfor reusing prompts across the spatial pipeline.

Conclusion: The future of spatial AI assistants

By 2026, the convergence of open-source models like CLM-8B and Jev with advanced voice services like Gemini 3.8 Flash TTS is transforming AI assistants into integrated spatial partners. Whether you're developing AR experiences, 3D editors, or interactive simulations, mastering prompts for these models enables you to create intuitive and responsive workflows.

Next steps

• Clone thecontrastive-lm/CLM-8Brepository and test action evaluation in your Unity application.
• Design a new schema for Jev that includes user preferences (e.g., "prioritize fast movement").
• Record a custom voice with Gemini TTS and create a set of voice instructions for your assistant.