We're Awiros. We write software that turns existing CCTV and inspection cameras into something operators can actually use: real-time alerts, searchable evidence, and a bit less "stare at a wall of screens forever."

This is our first post on DEV. We'll keep the product pitch short and talk like engineers who have lived with RTSP drops, flaky discovery, and models that look great in a notebook and weird under sodium lighting.

The boring problem that still isn't solved

Cameras are already everywhere: plants, campuses, rail platforms, data centers, event venues. Most of that video is recorded, stored and never watched. People can't watch it at scale. So teams either over-alert, under-alert, or give up and treat video as a forensic archive after something goes wrong.

Our bet is simple. If you can run reliable computer-vision agents on the feeds you already have, every installed camera becomes a useful sensor instead of a liability.

What we build

  • Aura: connect cameras / NVRs / VMS, place Vision AI agents on channels, route detections with video proof to the people who need them. Edge, on-prem, or cloud.
  • AppStack: a catalog of ready CV apps you can try on your own feeds and deploy through Aura.
  • ATOM: annotation and custom-model path when off-the-shelf apps aren't enough.
  • aiscope: visual inspection in the field (handheld or fixed), evidence included.

Outcomes we care about in the wild: safety, inspection, and operations. Same platform, same feeds.

We don't ask sites to rip out their cameras. ONVIF, RTSP, and the VMS they already run are the onboarding UX that matters.

One constraint we take seriously

Shipping CV into a demo is easy. Shipping it onto hundreds of live channels is a reliability and ops problem:

  • Discovery and onboarding are messy (cameras lie; networks lie harder).
  • Video plumbing still involves real streaming stacks (yes, GStreamer still earns its keep under a modern React ops UI).
  • Edge boxes are multi-arch, bandwidth-constrained, and unforgiving after quantization.
  • A bad model promote can look fine on a dashboard while operators lose trust. Once a channel gets muted, winning it back is slow.

So a lot of our engineering energy goes into orchestration, honest pipeline state, human judgment before risky publishes, and making search/evidence useful after the alert fires.

What you'll see from us here

We'll keep posting about the messy parts of putting CV on real cameras: what broke, what we changed, and what still sucks.

Coming up, roughly:

  • Why we put a human in the loop before a model goes live
  • ONVIF discovery (the onboarding UX nobody romanticizes)
  • Using embeddings / similarity search as memory for past events
  • Multi-arch containers that actually run next to the cameras

Soft ask

If you run CV, video ops, or MLOps in production: how do you decide a model is safe to put on a live camera? Drop a comment if you've got an take. We're listening more than pitching :)

Site: awiros.com · Account: @awirostech