Your cameras are already recording everything.
Nobody's watching it.

Brelisk watches your security footage and tells you what happened and why, in plain language, so nobody has to sit and scrub through hours of video.

Where this stands today

An early-stage R&D project, not a shipped product.

Brelisk pairs deterministic computer-vision tracking with a reasoning layer to turn ordinary security-camera footage into structured, explainable operational insight for physical retail. We're building and testing this incrementally, starting with one well-defined use case, checkout queue and dwell-time analysis, before expanding to the full metrics set further down this page.

Right now it's a working prototype, tested on recorded retail footage. This isn't a live deployment. The waitlist is for people who want to follow progress or talk about an early technical pilot, not sign up for a live product.

Sales dropped.

Why?

No one knows.

Customers entered.

Where did they lose interest?

Unknown.

Checkout lines formed.

How much revenue was lost?

Unknown.

Store A outperforms Store B.

What operational difference explains it?

No evidence.

Traditional analytics measure activity. BRELISK explains it.

How it's built

Two layers, split by how fast each one needs to run.

Perception layer

Fast, continuous: runs on every frame

  • YOLO-family object detection finds people and relevant objects in each frame.
  • A tracker from the ByteTrack / BoT-SORT family assigns each person a persistent ID across frames.
  • Zones are defined per camera (checkout lanes, product displays, entrances) so dwell time, entry/exit, and occupancy can be measured directly from tracked positions.

Reasoning layer

Slow, selective: triggered only on flagged events

  • Doesn't run on every frame. That would be too slow and too costly.
  • The perception layer flags a moment worth explaining (an unusually long dwell time, a queue building up) and crops the image around it.
  • A vision-capable language model turns that crop (raw geometry, boxes, tracks, durations) into a plain-language explanation of what's likely happening.
This split (a fast reflexive layer and a slower deliberative one) is a common pattern in robotics and autonomous-vehicle perception stacks, where low-latency control loops are kept separate from higher-level scene reasoning. We're borrowing the architectural idea, not claiming any equivalence to safety-critical AV systems. Re-identifying someone who leaves and re-enters frame, distinguishing staff from customers, and cross-camera tracking are still being worked out.

How BRELISK Thinks

01

Observe

Detect people on every frame and track each one with a persistent ID.

02

Understand

Measure dwell time, queue length, and movement inside defined zones.

03

Explain

Flag unusual moments and describe, in plain language, what's likely happening.

04

Recommend

Turn that explanation into one specific, actionable next step.

OBSERVEDPerson #4 dwelling near Apparel for 3m 10s
OBSERVEDCheckout queue exceeded 5 people
OBSERVEDPerson #4 approached the queue, then paused
INFERENCEPattern consistent with queue hesitation
RECOMMENDATIONOpen a second checkout lane
Illustrative example, not a live result

Where we are

A build log, not a features list. Here's what's actually working today versus what's still ahead.

VALIDATEDPerson detection and multi-object tracking with persistent IDs, on recorded footage
VALIDATEDZone-based dwell time measurement
VALIDATEDHeatmap generation showing spatial movement density
VALIDATEDNatural-language explanations of standing/dwelling behavior, generated from cropped images
IN PROGRESSTracking reliability through occlusion and camera exit/re-entry
IN PROGRESSDistinguishing staff from customers
ROADMAPLive camera feed integration
ROADMAPPredictive alerts
ROADMAPPOS / revenue data integration
ROADMAPCalibrated confidence scoring
ROADMAPMulti-camera cross-site identity tracking

Sample output format: recorded test clip, illustrative only

people_tracked                : 42
walking                       : 31
standing                      : 11
zones.checkout.dwell_avg      : 47s
zones.checkout.dwell_max      : 3m 12s
zones.apparel.occupancy_peak  : 6

[reasoning] flagged: prolonged dwell, zone=checkout, track_id=17
[reasoning] explanation: "Person paused near the register for an
            extended period, consistent with queue hesitation."

What Brelisk is designed to measure

This is the long-term direction. Active work is focused entirely on the two categories below.

Zones

Dwell time

How long customers linger in a zone

Occupancy

How many people are in a zone right now

Entry rate

Rate customers enter a specific zone

Queue

Queue length

Number of people currently waiting

Wait time

How long customers wait in line

Abandonment rate

Share of customers who leave the queue

Every metric above describes what happened. Brelisk is being built to go further: a plain-language explanation and a specific recommendation.

The goal

Most tools tell you what happened. The goal here is to explain why.

Common questions

Do I need new cameras?

No, Brelisk is designed to run on your existing IP camera feeds. Right now we're validating detection, tracking, and reasoning on recorded footage; live camera integration is next on the roadmap.

How is this different from a heatmap tool we already have?

Most heatmap tools show where people walked. Brelisk is built to track individual behavior over time (hesitation, queuing, abandonment) and turn it into a plain-language explanation and a specific recommendation, not just a colored map.

What about customer privacy?

Brelisk tracks movement and behavior, not identity. Privacy is a core design principle we're building in from day one, not yet an audited or certified guarantee, but a constraint the architecture is designed around.

When can I actually use this?

We're not yet deploying to live sites. Brelisk is currently a working prototype, tested on recorded retail footage. If you'd like to be an early technical pilot partner, reach out through the waitlist form below.

BRELISK is building the intelligence layer for physical retail.
Starting with the cameras you already have.

For people following our progress or exploring an early technical pilot, not a signup for a live product.