FELIX ran for hours across the facility. In just the first 8 minutes of analyzed footage, the model produced a sustained risk score of 78/100: High Risk, immediate intervention recommended, with peak danger hitting 100/100. Across that window, the model fired roughly 10 risk signals per second. Every PPE gap, proximity violation, and crowd-density spike, on every frame.
I reviewed FELIX's output against my time running Amazon warehouses. The categories the model is flagging are what an experienced operator would catch on the floor. FELIX just does it at machine scale.
A human inspector sees this once a year, for an hour, on a scheduled visit. FELIX sees it continuously, on every frame. That is the underwriting signal we are building.
The score is built from 9 weighted factors. Not a qualitative checklist, but a quantified, persistent signal running on every frame.
V1's inference focuses on dynamic worker safety: PPE, crowding, proximity to equipment. Static facility hazards like unsafe stacking, fire egress, and capacity violations come into scope with V2's expanded sensor suite.
The risk score reflects our initial scoring methodology. We will continue refining the model and factor weights as we scan more facilities and gather more data.
10 risk signals per second, for 8 minutes straight.
V1 was never meant to ship as a finished product. It was a test of how AI-driven risk assessment holds up in a real warehouse. Two observations stood out, and both shape V2 directly.
The goal for June: V2 built, deployed, and scanning. That scan is the real test of what the full sensor suite produces when everything works together.
V1 revealed what matters. V2 is built around it.
Substance is running two GTM motions in parallel: brokerage and MGA. Here is what moved on each this month.
If you are reading this, you are already in our corner. Here is where we need the most help right now.