The Problem
Walk into almost any factory, water treatment plant, or refinery built before 2010. You will find walls of dials. Pressure gauges, temperature dials, flow meters, and small windows with spinning mechanical counters cover the equipment. They are reliable, yet none of them can send their data to a computer. Plants employ staff to walk the site on a schedule, read hundreds of these instruments by eye, and write the numbers down.
My family runs a small manufacturing mill. I have seen the time and effort required to train employees for these daily manual reading rounds. This project addresses the gap between advanced computing and the analog reality of the physical world.
The Idea
Argus is a Metis-powered camera appliance that reads legacy analog and digital instruments — dial gauges, 7-segment panel meters, status lights, mechanical counters — live, entirely on-device, and turns them into a structured data feed the rest of the plant's systems can actually use. No rewiring the gauge. No cloud round-trip. No new hole in the wall.
That last part isn't a nice-to-have, it's the whole point. Retrofitting a single gauge with a "proper" wired IIoT sensor routinely costs anywhere from $450 to $8,500 per point, mostly labor and conduit, not hardware. And even where a site wants AI to do this instead, the control network is very often required by policy to stay completely air-gapped from the internet — which quietly rules out every cloud-connected "smart camera" gauge-reader on the market today, no matter how good the model behind it is.
Argus is designed to be the one that's allowed in the building.
The need for Metis
This isn't a computer-vision project that happens to run on Axelera hardware — it's a project that only works because of what Metis specifically is:
- On-prem-only isn't a design choice here, it's the compliance requirement. Metis does real inference with nothing going out over the network, which is the actual feature the customer is buying.
- One low-power card, many gauges at once. Metis pushes up to 214 TOPS (INT8) at ~15 TOPS/W in an 8–15W PCIe card — efficient enough to run a full multi-stage pipeline across a dozen-plus instruments in frame simultaneously, instead of the one-gauge-one-camera demos everything else in this space ships.
- The PCIe card has enough memory to keep a detector, a classifier, and a pose/keypoint model resident at once for genuine cascaded execution, which the Voyager SDK already treats as a first-class pattern.
How it works - The Wingman Pipeline
The Wingman Pipeline chains four stages together.
```
camera frame
|
v
[1] LOCATE finds every gauge, panel, and light in the frame.
|
v
[2] CLASSIFY identifies the instrument as a dial gauge, digital panel, status light, or counter.
|
+-- dial gauge --> [3a] NEEDLE-ANGLE POSE processes center and tip keypoints to calculate a calibrated value.
+-- digital panel --> [3b] DIGIT OCR extracts numbers from seven-segment or LCD screens.
+-- status light --> [3c] COLOR STATE identifies lights as red, amber, or green.
|
v
[4] BRIDGE pushes readings out via Modbus, MQTT, or OPC-UA to a live dashboard.
```
The project adapts existing specialized models instead of building from zero. The pipeline uses ETH Zürich's analog_gauge_reader for needle-angle math and a YOLOv8 and CRNN seven-segment reader documented at 96 percent real-world accuracy. An OpenCV Hough-transform gauge locator serves as a low-compute pre-filter ahead of the deep models. Wingman integrates these components into a working cascaded pipeline on Metis hardware.
MVP Scope
The demo is deliberately physical: a small panel of real gauges, a couple of digital displays, and a status light, sitting on a desk, wired to nothing but a power outlet — next to a screen showing the live dashboard. Turn a dial by hand. Watch the number update in real time. I’ll even disconnect from the internet to show you that it’s on prem!
Planned Tasks :
- Week 1 — Detection + type classification running on the real rig; a few hundred real photos captured across lighting and angles to fine-tune.
- Week 2 — Needle-angle reading calibrated against real gauge markings; digit-OCR fallback for the panels; status-light classification.
- Week 3 — MQTT bridge and the live dashboard; get the "unplug the cable, turn the dial" demo bulletproof and repeatable.
- Week 4 — Polish, demo video, full write-up of every prompt and iteration.
Planned Deliverables
- The final submission will include the complete Wingman prompt history for every pipeline stage.
- I will provide a demo video showing a dial turned by hand with the network cable disconnected and the dashboard updating.
- A public code repository will cover the pipeline, protocol bridge, and dashboard.
- A technical write-up will detail model accuracy and current limitations.
MVP Limits and Future Scope
This proof covers four instrument classes. It handles specific gauge styles rather than every legacy indicator. Testing on a real ICS network is out of scope for this initial phase. The objective is to validate a multi-stage cascaded pipeline reading analog instruments on-device using Metis within one month.
I will continue working on the remaining instrument types and network protocols after the challenge to prepare the system for deployment in our family mills.
Why This, and Why Now?
Every camera-based gauge-reading product that already exists commercially assumes a network connection somewhere in the loop. That's a completely reasonable assumption for a warehouse or an office building. It's a disqualifying one for the sites that actually have the most gauges and the most money on the line — refineries, water treatment plants, power generation — because those control networks are frequently required, by policy, to never talk to the internet at all. That's not a gap in the market that needs to be discovered; it's a gap that exists because nobody building gauge-reading AI has had a real reason to make it work with zero network dependency until now. Metis is that reason.
I hope to hear from you soon!
