Industrial monitoring

RF first.
Camera on trigger.

172.8 GB/day → 128 KB/day. The thesis is not that cameras go away — only that they should run less often. Wi-Fi senses continuously. Vision verifies selectively. Language explains the result.

Wi-Fi senses continuously. The camera wakes only when the radio says something interesting happened.

The radio runs flat-out at sub-2 watts. The camera sits at 5% duty cycle, asleep until a CSI anomaly elbows it. That split — boring continuous sensing, sparse expensive recording — is what gets the daily data volume down from 172.8 GB to 128 KB. Same coverage. Different architecture.

The four numbers that decide the architecture

The gap is not subtle. It is architectural.

1.59WContinuous drawvs 23–41W camera-first (Jetson Orin Nano)
128 KB/dayStructured eventsvs 172.8 GB/day continuous video at 4 Mbps
91.5%Lighting energy reduction20× 150W high-bays · 72 → 6.12 kWh/day
200–314hBattery autonomy on 500 Whvs 12–22 hours camera-first

All four come from the same architectural choice: sense with the radio you already have; record with the camera only when the radio says something matters.

3-layer architecture

Signal · World-model · Interface.

See the camera-on-trigger loop run on a synthetic site — Open the sandbox.

Layer 1

Signal — continuous + sparse

WiFi CSI runs 24/7 on $4 ESP32-S3 nodes. Cameras stay asleep ~95% of the time and only wake on an RF trigger. Cameras now have a 5% duty cycle.

Layer 2

World-model — latent, not raw

Events are state changes, not frames. The world model is a low-cardinality stream of structured events that fits in any tool, any prompt, any database.

Layer 3

Interface — chat with space

Operators query in natural language. The LLM operates against deterministic events with rollback-safe actions. Co-pilot, not black box.

Continuous video is a streaming problem. Structured events are a telemetry problem.
Three things camera-first cannot do

The moat is physics, not features.

Survive jamming

Jamming makes us sharper. Under WiFi jamming, ESP32 nodes retry more aggressively → more emissions → more CSI samples. When WiFi is fully blocked, we read the physical layer over USB serial straight to a laptop.

A jammer cannot touch copper.

Reason about what it sees

172.8 GB/day of video does not fit in an LLM context. 128 KB/day of structured events does. The world model is the prompt.

Camera-first edge AI can run a classifier. It cannot run a conversation.

Boring-reliable AND intelligent

Deterministic RPA rules cover the 95% case, auditable end-to-end. Agentic LLM activates only on the 5% that doesn't match a rule.

Floor is robust automation. Ceiling is autonomous reasoning. You never get one without the guardrails of the other.

Out-of-band delivery

LoRa is the second mouth, not a backup mesh.

CameraPlus carries a Semtech SX1262 sub-GHz radio. It's deliberately on a different frequency band from your WiFi, so a jammer or congestion event on 2.4/5 GHz cannot silence it.

We promise a second independent delivery channel. We do not promise guaranteed multi-hop mesh routing. For an oil rig that loses satellite, one LoRa path to anywhere is the difference between ten hours of silence and an immediate alert.

Even if everything else fails, CameraPlus still talks.
Side-by-side

A one-week side-by-side reads like a sentencing.

Metric (per node, 24/7, one week)Camera-first edge AICameraPlusΔ
Power draw (continuous)23–41W1.59W~93% lower
Data emitted per day172.8 GB128 KB~1,350,000× less
Battery autonomy on 500 Wh12–22 hours200–314 hours~14× longer
Survives WiFi jammingGoes blindGets sharperInverted
Out-of-band escapeNetwork onlyLoRa sub-GHz+1 channel
Fits in an LLM contextNoYesArchitecturally

Camera baseline: UniFi G5 Bullet, AWS surveillance reference, 4 Mbps. CameraPlus: 3× ESP32-S3 (0.29W each) + 1× gateway (0.37W) + Jetson Orin Nano at 5% duty cycle.

What we don't claim

CameraPlus is not for every workload. That's intentional.

CameraPlus is not the right architecture if you need…

  • Forensic-grade continuous video evidence
  • Identity-grade face or license-plate recognition
  • Compliance recording in regulated jurisdictions where video is the legal artifact
  • Sub-centimetre localisation

Use cameras for those. We don't fight cameras where they win.

CameraPlus is the right architecture if you want…

  • Continuous coverage with sane power and bandwidth
  • Independent monitoring that survives network outages
  • Conversational query against your physical site
  • Energy-orchestration economics that justify the deployment
  • A monitoring stack a jammer cannot silence