Camera placement is the easy part. Watching the footage is the part nobody gets right.

1. Camera placement changes as the site does

Most retail security cameras get installed once and stay there for a decade. A construction site is the opposite. The site changes weekly. The lay down yard moves as material gets used. The trailer office relocates when foundations get poured. The fence line shifts when the scope expands. Cameras that pointed at a high traffic zone in week three are pointing at empty dirt by week ten.

The good news is most modern jobsite cameras (battery solar trailers, conduit-mount IP cameras) are easy to physically reposition. The bad news is nobody actually does it. Once the cameras are up, they stay up unless somebody complains.

2. The DVR in the construction office trailer

Almost every site with cameras has a DVR (digital video recorder) tucked into the construction office trailer. The DVR has a hard drive, usually 2 to 8 TB, and records continuously to a rolling buffer. The trailer monitor shows a 4x4 or 5x5 multiview live feed. Nobody watches it.

The reality is the site super has a job that is not 'monitor the DVR.' The DVR exists for after the fact review when something goes wrong. By the time it gets reviewed, the perpetrator has been on and off site for hours.

This is the dead DVR problem and it is universal. The hardware is fine. The problem is the lack of any layer between the DVR and a person who can act on what it sees.

3. AI overlay as the missing layer

The fix is not more cameras or better cameras. It is a small AI device that watches the DVR's output in real time and pages on call when a configured rule fires. Trespass into the lay down yard after 1900. Vehicle in the loading area at 3 AM. Person near the genset shed during a non-work hour.

Cyrano is one product shaped this way: an edge AI device that plugs into the back of the DVR over HDMI, runs object detection on the multiview at native frame rate, applies per zone per time window rules, and sends WhatsApp or SMS clips when the rules trigger. There are a few other products in adjacent shapes; the category is small but real.

The hardware install is one HDMI cable and one power cord, around 2 minutes. No new cameras, no new cabling. The cost is roughly $450 hardware plus $200 a month per site for the software.

4. Retrofit vs greenfield on a working camera system

If the cameras are 5 to 10 years old but the DVR records cleanly and the night IR is workable, retrofit beats greenfield by a factor of 10 on cost and a factor of 50 on disruption time. Retrofit means the existing investment is preserved and the missing intelligence layer is added on top.

Greenfield (rip and replace) makes sense when the cameras are physically failing, the DVR is dropping recordings, or the IR is so bad that nighttime footage is unusable. In that case the AI overlay would be detecting nothing because the underlying feed is bad.

A field check: pull a clip from each camera at 1 AM with the work lights off. If you can identify a person in the frame, the cameras are good enough for AI overlay retrofit. If the frame is black or grainy past 10 feet, the camera itself needs replacement first.

5. Zone and time rules, the part that matters

An object detection model that says 'person at coordinate (x,y)' is only useful if there is a rule layer on top. The detection model alone produces tens of thousands of person detections per day on a busy site. None of that is alertable.

What is alertable: person in the genset enclosure outside scheduled fueling hours. Vehicle in the loading bay between 2200 and 0500. Person climbing the perimeter fence at any hour. The rule layer turns the raw detections into a small list of events that warrant a real time clip to on call.

The richer the rule grammar, the fewer false positives. A simple system gives you 'motion in zone' which is useless. A useful system gives you 'object class + zone polygon + time window + dwell duration' as a single composable rule.

6. The on call workflow when an alert fires

When a rule fires, what should happen on the human side? The pattern that works is a 10 to 15 second clip pushed to a dedicated WhatsApp or SMS group, not a separate dashboard nobody opens. The on call sees the clip on their phone in 5 to 10 seconds, decides if it is real, and either dispatches or dismisses.

The dispatch path varies by site. Some GCs use a roving private patrol service with the site keys; some call local non-emergency police; some call the site super directly if the incident looks small. The point is the alert lands on a phone someone is already carrying, not on a dashboard nobody is logged into.

7. Scaling across a multi site GC

A small GC with 8 active sites and one rolling deployment of edge AI overlay devices typically pays back in a quarter. One prevented break in attempt at one site covers a year of software for the entire portfolio.

The discipline is to keep the rule grammar uniform across sites: a person in the lay down yard after 1900 is alertable everywhere. Site specific overrides (this site has a tenant doing late electrical work, exempt them from 1900 to 2200) live in a per site config that can be edited from a phone.