TECHNOLOGY · LRM

A reasoning model that respects the laws it was trained on.

Vermilion is built on Physics-Informed Neural Networks: neural architectures whose loss functions encode the conservation laws and constitutive relations that govern real machinery. The result: predictions that are accurate, explainable, and physically possible.

5
Auditable reasoning stages, every prediction
0.012
Expected Calibration Error on field deployments
312ms
Median inference latency · edge or cloud
LRM · 5 STAGES200 Hz
STAGE 03 · PHYSICS FIT
PHYSICS-INFORMED · LRMp = 0.93 · ECE 0.012 · 312 ms
HOW IT WORKS

From raw signal to grounded recommendation, in five steps.

Each stage is auditable. Each stage publishes its uncertainty. Each stage refuses to produce conclusions that violate the physics of the asset under observation.

  1. 01
    Ingestion
    Collects sensor data, maintenance records, and equipment specifications from any source.
  2. 02
    Conditioning
    Phase-aware denoising, operating-state inference, regime classification.
  3. 03
    Physics fit
    Constitutive parameters fitted per asset — bearings, gearboxes, pumps — under known physical priors.
  4. 04
    Reasoning
    PINN-grounded inference produces a failure-mode hypothesis with calibrated confidence.
  5. 05
    Action
    Time-bounded recommendation pushed to the work-order system with full reasoning trace.
KEY COMPONENTS

Three breakthrough components in one platform.

Our platform combines several breakthrough technologies to deliver unmatched predictive capability, held together by a single reasoning loop.

01 / 03

Physics-Informed Neural Networks

Neural networks that incorporate physical laws and constraints, allowing the model to respect fundamental principles like conservation of energy and momentum.

02 / 03

Equipment Models

First-principles digital twins — bearing dynamics, gear-mesh harmonics, fluid-film stability — that the LRM reasons against in real time.

03 / 03

Failure-Mode Analysis

A continuously growing catalog of failure mechanisms: spalling, fatigue, cavitation, resonance, thermal runaway. Anonymized and shared across the fleet, opt-in.

INSIDE THE PRODUCT

The interface, not just the claim.

Three views from the Vermilion operations console. Each is a real screen our customers see — not marketing illustration. The reasoning model's outputs are surfaced where they get acted on.

live screenshots from app.vermilion.ai
app.vermilion.ai · remaining useful life
FLEET · ALPHA SQUADRON · RUL FORECAST
3 urgent
Remaining useful life across 5 assets
Drone-A2376d left
Motor 3 bearing
CRITICAL
p=0.94
Drone-B11812d left
Battery pack cell
MONITOR
p=0.88
Drone-A10521d left
Propeller balance
MONITOR
p=0.81
Drone-C02228d left
Cell-pack imbalance
WATCH
p=0.83
Drone-B23130d left
Control surface
HEALTHY
p=0.96
FORECAST HORIZON · 30 DAYSVERMILION-R · 9DEB2BA
PLATFORM ARCHITECTURE

Cloud, edge, or air-gapped — same reasoning.

A scalable, deployment-flexible architecture designed for reliability and performance.

DATA SOURCES

Sensors, historians, maintenance logs, equipment specs. Over OPC-UA, MQTT, REST, gRPC, or batch CSV.

OPC-UA / MQTT
Historians (PI, Ignition, Wonderware)
REST / gRPC
Batch CSV / Parquet
Edge gateways (custom protocol adapters)
OPC-UAMQTTREST/gRPCHistorianCSV/ParquetCOLLECTOR00000 rec/s
STAGE 01 · DATA SOURCES
TECHNICAL DEPTH

The questions our customers' engineers ask.

Our networks include differentiable physics layers. The training loss penalizes outputs that violate the conservation laws and constitutive relations of the asset class, so the network is incentivized, throughout training, to converge on physically consistent solutions.
We use deep ensembles plus a temperature-scaled softmax head, validated quarterly against held-out field outcomes. Reliability diagrams are published in the customer portal.
Per-asset fitting of constitutive parameters is required (a few hours, automated). The reasoning core itself does not require retraining for new asset classes within a covered family.
Yes. We support fully on-prem deployments with periodic, optional, opt-in synchronization of anonymized failure-mode patterns.
READY WHEN YOU ARE

From principle
to practice.

Spend thirty minutes with the team building the reasoning core. We'll walk through how the science fits your assets, and where it doesn't.

Request a private demo How it works