The question
behind the question.

01 / The request

A real question asks for an investigation.

Someone calls in and says: "I think a car clipped me in the parking lot yesterday afternoon." That sentence is not a search query. It's a request for an investigation.

Most video systems, including most of what's sold as "AI-powered" today, were built to answer a narrower question: is there an object in this frame that matches a category I defined in advance? Person. Vehicle. Package. A red car, if you're willing to write that rule yourself and only for the one camera you thought to check. That's a real capability, and a lot of the industry has spent a decade making it fast and cheap.

It just isn't what the person on the phone is asking for. And once you notice that gap, you start seeing how much room is sitting on the other side of it.

02 / Unpack the sentence

What the request actually contains

"A car clipped me in the parking lot yesterday afternoon" is a compressed version of a dozen real questions:

1. 01

   Which of the dozens of vehicles that passed through the lot yesterday could plausibly have done this?

2. 02

   Where else did that vehicle appear: before, after, on other cameras, at other entrances?

3. 03

   Does the pattern look like an accident, or does it look like someone who circled back, slowed near the same car twice, or left in a hurry?

4. 04

   Does anything else, such as an access badge, a license plate read, or a prior visit, attach a name or a pattern of behavior to that vehicle?

5. 05

   What's the smallest set of clips that actually tells this story, in order, without making a human scrub six hours of four camera feeds to find out?

None of these is "find the red car." The useful answer is the synthesis of all of them: a short report and a cut of footage, not a grid of thumbnails that happen to match a detected class.

03 / Relationships

Object detection is the floor, not the ceiling

Detection answers "what's in this frame." That's a real, necessary layer, and it's also just the input to a much bigger question: given everything a site has ever recorded, what actually happened, to whom, and does it matter?

In the illustration, two vehicles become a relationship across time: a red SUV passes beside a parked car. The connection says more than either label on its own.

04 / The context around it

Identity. Time.
Location. Connected.

Once you have identity, time, and location connected across every camera and every system on a property, including badge reads, plate reads, and prior incidents, the interesting work stops being "did we spot the object" and starts being "what does this pattern mean." A system that can hold that much context can do things a per-camera detector was never positioned to do: follow one thing across a whole property instead of one feed, notice that a pattern repeats across days instead of treating each clip as its own event, and go get the handful of relevant seconds out of thousands of hours instead of asking a person to know where to look first.

The illustration distinguishes what this example contains from the additional evidence a broader investigation would need.

05 / Reason across the site

A question becomes a plan for evidence.

That's the opportunity we think the industry is sitting on: not a better detector, but a system that reasons across everything a site knows the way an experienced investigator would, and does it at the speed a real question deserves.

The original report is a hypothesis to investigate. It should not become a conclusion just because it appeared in the question.

06 / The answer has to hold up

Where this goes

Detection was never the hard part to sell. It was the easy thing to build a whole industry around. The bigger opportunity is what you can do once a system is willing to connect what it sees to who it is, where it's been, and what happened around it. That's the part worth building for, and it's the part we spend most of our time on.

The useful answer is the story, with the evidence still attached.