Infrastructure · Integration · Intelligence

Industrial vision,
interpreted in real time.

From the camera mount to the model. Vantage designs the layout, integrates with your existing PLCs, and runs the AI that flags PPE breaches, defects and anomalies on the edge, in under 40 ms.

Inference
38ms p95
Streams / node
64concurrent
Detection F1
0.94avg.
Factory floor with vision monitoring active
REC CAM-04 · BAY-B
38ms · 30fps YOLOv9-T · INT8
PERSON · PPE OK · 0.97
NO HARD-HAT · 0.91
FORKLIFT · IN-LANE · 0.94
ZONE BREACH
14:22:08PPE violationWARN
14:21:47Zone breachP1
14:21:12Cycle completeOK
14:20:55Forklift detectedOK
14:20:31Spill clearedOK
4 / 4 Persons tracked
1 active PPE violations
0 P1 Open alerts · 24h
// Trusted on the floors of 80+ industrial sites worldwide
Halford Steel NORDFAB / 04 Meridian KOBALT. Stellar Foods ≡ axis_one
01 · The stack

Three things
one stack.

Vantage is not just a model. We design the camera layout that makes vision possible, integrate with the PLCs and MES that already run your plant, and ship the AI that runs on top.

Factory production line for camera infrastructure planning PILLAR 01
A
B
C
D
E
/ 01

Camera and belt
infrastructure design.

Line-of-sight, lux, lens, mount and cabling. Surveyed and specified by our field engineers before we push a single line of code.

  • Site survey, lux mapping and lens selection
  • IP67, IK10 and ATEX-rated cameras for hazardous zones
  • PoE+, fibre or private 5G backhaul
  • Vibration-damped mounts for press and conveyor lines
  • Retrofit of existing RTSP and ONVIF cameras
Electrical control panel and industrial cabling PILLAR 02
OPC-UA
MODBUS TCP
PROFINET
ETHERNET/IP
MQTT
REST / Webhook
/ 02

Industrial
Industrial.

Vantage speaks the protocols that already run your plant. No rip and replace, no parallel data silo, no proprietary stack.

  • OPC-UA, Modbus TCP, Profinet and EtherNet/IP
  • Siemens, Rockwell, Mitsubishi and Beckhoff PLCs
  • MES and SCADA: Ignition, Wonderware, AVEVA
  • Bidirectional: read cycle pulses, trigger interlocks
  • 21 CFR Part 11 and ISA-95 compliant audit trail
Industrial workers in safety gear at steel factory PILLAR 03
PERSON · PPE OK · 0.97
HARD-HAT · 0.95
/ 03

AI and computer
vision.

Edge-resident detectors, pre-trained on industrial footage and tuned on your floor. Eighty production models out of the box.

  • 80+ production detectors at F1 0.94 average
  • INT8 quantised inference at 38 ms p95
  • Active learning: 50 frames trains a new class
  • Unsupervised anomaly detection, no labels needed
  • Frames stay on premises unless you say otherwise
02 · In motion

What 38 ms
of inference looks like.

Every part on the line gets a track ID, a class label, a confidence score and a defect verdict, at 30 fps, in real time. The video below is an actual unedited line feed; the overlays are Vantage running live.

LIVE ·  CAM-12 · Bay-B · packaging line 04
Detector: part_qa_v6 · INT8
Stream: 1920×1080 @ 30 fps
Inference: 38 ms p95
REC · LIVE CAM-12 / BAY-B 38 ms · 30 fps
part_qa_v6 · INT8 F1 0.94
PART · OK · 0.97
PART · OK · 0.95
DEFECT · TEAR · 0.91
PART · OK · 0.96
// Shift · B-04 · 14:22
1,247+38
OK1,234
Reject · tear9
Reject · misalign4
Avg cycle1.42 s
4 tracked p50 32 ms  ·  p95 38 ms  ·  p99 51 ms GPU 47%  ·  CPU 22% Auto-clip armed  ·  S3 // eu-west-1
Throughput today
38,412parts
Counted automatically. No PLC integration required.
Reject rate
1.04%
−0.6 pp vs. last shift. Tear defects trending up at station 3.
Mean cycle
1.42s
Within ±50 ms of target. Auto-flagged station 2 (1.81 s).
False positives
0.3%
Active-learning loop re-trained the classifier 4 days ago.
03 · Capabilities

What it sees,
and what it does about it.

Eighty pre-trained detection models, plus a no-code rule builder for the events specific to your plant. Train a new class with as few as fifty labelled frames. No ML team required.

PERSON · PPE OK · 0.97
/ 01

PPE compliance

Hard-hats, hi-vis, gloves, glasses, masks, harnesses. Per-person, per-zone, with grace-window logic so transit doesn't trigger noise.

Modelppe_v6 F10.96 p9532 ms
ZONE BREACH · LOTO-3
/ 02

Zone & intrusion

Draw any polygon on a live feed. Define entry, exit, dwell-time and direction rules. Tied to interlocks via OPC-UA or Modbus.

Modelzone_v4 F10.95 p9536 ms
ANOMALY · 0.89
/ 03

Anomaly detection

Unsupervised models that learn your floor's baseline and flag what shouldn't be there: open panels, leaks, unexpected motion, line stops.

Modelanomaly_v3 F10.91 p9544 ms
CYCLE · ON-TARGET
/ 04

Cycle & throughput

Measure station cycle time, dwell, idle and changeover, automatically. Compare shift vs. shift, line vs. line, without a stopwatch.

Modelcycle_v3 F10.94 p9530 ms
SMOKE / FIRE · 0.91
/ 05

Spill, smoke & fire

Early detection of liquid spills, dust, smoke and open flame, even before standard smoke detectors trigger. UL-aligned classifiers.

Modelthermal_v2 F10.93 p9540 ms
DEFECT · TEAR · 0.92
/ 06

Defect & visual QA

Surface defects, fill-level, label placement, missing components. Pixel-precise even on glossy and reflective parts under variable light.

Modeldefect_v6 F10.91 p9538 ms
04 · The platform

One console.
Every floor, every camera.

From a single browser tab, supervisors triage live events; EHS leaders audit compliance; plant managers track throughput across sites. Role-based, on-prem or SOC 2 cloud.

vantage.app/ops · halford-steel · bay-b
eng@halford · admin
Cameras online
142+2
Events · 24h
3,184−6%
Open · P1
2+1
P95 latency
38ms
Live floor · Bay B // 6 of 18 cams
CAM-04 / BAY-BOK
30 fps38ms
money CAM-11 / PRESSWARN
30 fps41ms
CAM-07 / WELDOK
30 fps36ms
CAM-02 / DOCKP1
30 fps44ms
CAM-09 / LINE-3OK
30 fps39ms
CAM-15 / YARDOK
30 fps37ms
05 · Inside the console

The whole product,
not just the marketing page.

Vantage is built for a control-room operator at 03:00 · not a vendor showing slides at a trade show. Every screen is designed for triage speed, evidence chain, and audit-grade traceability.

// Screen 01 · Detector library

80+ detectors pre-trained for the floor.

Browse, preview and deploy production-grade models in two clicks. Each detector ships with benchmark F1, latency, and false-positive rates measured on real industrial footage, not on the COCO dataset.

  • Per-detector cost: GPU cycles, not seat licenses.
  • Confidence thresholds and grace windows per zone.
  • Active learning: re-train on your floor's footage in hours.
vantage.app/library
// 80 detectors
PPE
Hard-hat compliance
F1 0.9632 ms
PPE
Hi-vis vest
F1 0.9430 ms
SAFETY
Gloves & welding mask
F1 0.9234 ms
pedestrian forklift ZONE
Pedestrian–forklift
F1 0.9536 ms
QA
Surface defect
F1 0.9138 ms
SPILL
Liquid spill / smoke
F1 0.9340 ms
vantage.app/cams/cam-04/zones
// editing: walkway-A
WALKWAY · A
RESTRICTED · LOTO
COUNTING LINE · L1
// Screen 02 · Zone configuration

Draw a polygon.
Skip the integration SOW.

Configure restricted zones, walkways, counting lines and dwell rules on a paused video frame, no Python, no calibration matrix, no on-site engineer. Vantage maps pixel space to real-world coordinates automatically.

  • Five-second authoring loop: draw, set rule, deploy.
  • Per-zone schedules, severities and routing channels.
  • Version-controlled. Every zone edit is auditable.
// Screen 03 · Event investigation

From alert to root cause in under a minute.

Every event arrives with a 30-second video clip, the model's confidence breakdown, the zone & schedule it tripped, and the recommended next action. Triage, assign, or mark a false positive, and feed the correction back into training.

  • One-click evidence export for incident reports.
  • Auto-correlation with PLC events via OPC-UA timestamps.
  • 21 CFR Part 11 audit log. Every action signed and timestamped.
vantage.app/events/ev_0x9c44a1
// P1 · open · 00:01:14
CAM-02 · DOCK-3 · 14:21:47
PERSON · RESTRICTED ZONE · 0.97
14:20:32 14:21:47 · event in 14:22:14
● P1 · OPEN

Zone breach · restricted

14:21:47 IST · Sat 23 May · Bay B · CAM-02
Detector
zone_breach_v4polygon · LOTO-3
Confidence
0.97track id #4429
Dwell
14.2 sthreshold 6 s
Linked PLC
PRESS-04cycle 41 · running
Severity
P1auto-escalated · 60 s
Assignee
M. HollowayEHS lead · Bay B
06 · How it works

Live in four days,
not four quarters.

No new cameras. No new cabling. Vantage runs as a containerised gateway on a single edge node per site, ingesting any RTSP, ONVIF or USB stream you already have.

STEP 01

Connect any camera

RTSP, ONVIF, USB or IP. Auto-discovery on your subnet. No firmware updates required.

STEP 02

Draw the rules

Polygons, lines, classes, schedules. Configured by an EHS lead in a browser, not a developer.

STEP 03

Detect on the edge

INT8 quantised models run on a single GPU node. No frames leave the plant unless you say so.

STEP 04

Route the signal

Webhook, MQTT, PagerDuty, Slack, Teams, OPC-UA, Modbus. Auto-clip evidence to S3 or on-prem.

Inference latency
38ms p95
Edge-resident. Frames never leave the plant boundary.
Streams per node
64
On a single L4-class GPU at 30 fps, 720p, mixed detectors.
Detection F1
0.94avg.
Across 14 baseline classes, audited quarterly.
Time to first event
4days
From PO to first production-grade alert, on average.
Customer · Halford Steel · 6 sites · 412 cameras
“We spent eleven years promising the board we'd see every near-miss in real time. Vantage shipped that in seven weeks, across all six plants, without re-cabling a single line.”
M
Maren Holloway
VP, Operations & Safety · Halford Steel
Recordable-incident rate
68%

Year-over-year, after 90 days of deployment.

Unplanned downtime
23%

Anomaly detection on press lines and conveyors.

Audit prep time
4×

Faster. Automated evidence clips replace clipboards.

08 · Get started

The factory floor is talking.
Start listening in four days.

Send us a single camera feed and a use case. We'll send back a live detector running on your stream, at no cost, within a week.

SOC 2 Type II  ·  ISO 27001  ·  on-prem & air-gapped supported

Let’s Talk About Your Product

Get expert guidance on scope, architecture, timelines, and delivery approach so you can move forward with confidence.

What happens next?