Geolava, a San Francisco spatial AI company built around world models, aims to replace 2D inspection with systems that forecast how a scene evolves, CEO Hantz Fevry said on the Visions podcast hosted by Vision Systems Design.

"World models will become the intelligence layer for the physical environments," Fevry said, drawing the comparison to language models as the intelligence layer for text.

What is a world model for vision?

Spatial intelligence, as Fevry uses the term, is a multi-modal platform that fuses optical, LiDAR, thermal, and 3D scene data with contextual text, then tracks the scene across time. Output is a structured representation of the physical world, not an annotated image.

Asked why it matters now, Fevry pointed to causal reasoning. A forklift and a worker detected separately tell a 2D system little; the same two objects fed into a world model expose a collision path and the risk it carries.

"There is a bit more of the reasoning beyond just the perception or the object detection," Fevry said.

How does this beat 2D vision in measurable ways?

In an industrial scene, a world model can infer occluded regions, tolerate reflections, and compensate for inconsistent lighting by reconstructing what the camera does not directly see. Fevry framed sensor fusion as central: "Instead of relying on a single image, you can focus on fusing information across sensors, viewpoints, time to infer what basically cannot be observed."

The implication for spec sheets: performance figures tied to one camera, one lens, and one exposure no longer fully describe the system. Benchmark conditions now include viewpoint diversity, sensor mix, and the time horizon for prediction.

What is the benchmark engineers should care about?

Fevry urged a shift in KPIs from detection accuracy to generalization under change. "The real test is not whether it detects a defect today, but whether it predicts the defect before it becomes a failure," he said.

For validation, Geolava advises testing against ground-truth labels of defects across time, not static image sets. Fevry warned that vendors may report strong perception accuracy while masking hallucinations in forecasting — the failure mode that matters most in predictive maintenance.

Where does it fit on the factory floor?

Practical deployments include production-line co-pilots for analysis, predictive-maintenance routines, and physical AI for robotic automation. Fevry recommends phased rollouts tied to measurable ROI.

"Begin with a production line where reducing downtime by ten percent delivers an immediate and measurable return on investment," he said, "rather than just testing the spatial intelligence for the sake of having a new technology."

Legacy integration remains the largest friction point. Existing PLC, SCADA, and vision pipelines were not designed for a model that continuously ingests new sensor streams, and Fevry described the work to retrain those workflows as "not easy but also not impossible."

Who is building Geolava?

Fevry has worked in AI for ten years, including a stint on a Google speech team that became part of Google DeepMind. He built and sold a prior AI venture in 2024 and assembled Geolava's team from alumni of DeepMind, Apple, and Meta. Current products target world models for critical infrastructure and property.

What is left to prove?

The adoption question is whether a world model can be specified, validated, and audited against a manufacturing standard the way a 2D vision system can today. Fevry drew the line himself in the closing minute: "World models will revolutionize the AI and the physical world the same way LMs revolutionized AI with text." Whether that revolution arrives on a metrologist's timeline remains the open question.