Leverage existing cameras to unlock instant insights for Quick Service
Understand every second. Act before it costs you.
Your existing security cameras become an operations team that never blinks. No new hardware — just the sentences a manager would say if they were standing there.
The gap is not the model. It is the sampling rate — and everything that falls between the frames.
"One frame every 3–5 seconds. A guest can arrive, wait, give up and walk out entirely between two samples."
"Every person tracked at 30 fps, unbroken — using skeleton and motion, never a face."
"A real-time vision-language model reads the actual scene and describes it in 242 milliseconds."
"Lane 2 has stalled — the guest at the front has been waiting 4:10 and two behind them are leaving."
Every item is a distinct event type with its own threshold, alert route and audit trail. Turn on what matters to your format; leave the rest off.
FLW-01Queue length by laneFLW-02Wait time per guestFLW-03Abandonment riskFLW-04Walk-in vs walk-by rateFLW-05Greeting latencyFLW-06Party size at entryFLW-07Occupancy vs capacityFLW-08Menu-board dwell timeFLW-09Aisle & doorway blockageFLW-10Drive-thru stack depthSVC-01Table occupancy stateSVC-02Table turn timeSVC-03Seat-to-order intervalSVC-04Order-to-serve intervalSVC-05Empty glass refill cueSVC-06Hand-raise service requestSVC-07Uncleared table detectionSVC-08Plate-left-uneaten signalSVC-09Server zone coverageSVC-10Check-drop to payment gapPAY-01Item count vs order ticketPAY-02Skip-scan at self-checkoutPAY-03Handoff to wrong guestPAY-04Void & refund video matchPAY-05Drawer open, no salePAY-06Queue-to-register ratioPAY-07Bag assembly completenessPAY-08Side or condiment omissionPAY-09Payment dwell timePAY-10Intervention need forecastLIN-01Station cycle timeLIN-02Bottleneck stationLIN-03Cook-time overrunLIN-04Holding-time violationLIN-05Order assembly checkLIN-06Portion consistencyLIN-07Waste & dump eventsLIN-08Equipment door left openLIN-09Cold-chain door durationLIN-10Cleaning task verificationSAF-01Handwash frequency & durationSAF-02Gloves before food contactSAF-03Hairnet & headwear complianceSAF-04Raw-to-ready-to-eat crossoverSAF-05Uncovered food on the lineSAF-06Dropped-item reuseSAF-07Floor spill detectionSAF-08Surface sanitising cadenceSAF-09Pest movement anomalySAF-10Allergen station separationPPL-01Zone staffing coveragePPL-02Shift changeover delayPPL-03SOP step completionPPL-04Training-moment capturePPL-05Slip & fall detectionPPL-06Distress or no-motion alertPPL-07Lone worker after hoursPPL-08Loitering & erratic movementPPL-09Altercation (motion-based)PPL-10Concealment gestureRunning a language model on every frame of every camera is not affordable, and anyone claiming otherwise has not done the arithmetic. So we split the work.
A vision-language model on an edge device answers roughly four times a second. Twenty cameras running continuously would need forty. So layer one — cheap, fast, always on — decides when layer two is worth waking up.
Cost drops by an order of magnitude, latency stays under a quarter second, and no video ever crosses your firewall. What syncs to head office is a line of text and a timestamp.
MISE analyses facial expression and body motion to score service quality. It does not match faces against any database, does not build a persistent profile, and stores no face images. Identity is a temporary track ID that expires the moment the person leaves the frame.
Expression analysis is regulated differently by market, so it is enabled per region and per role. Under the EU AI Act, emotion recognition in the workplace is a prohibited practice — staff-facing analysis is therefore disabled in EU deployments and cannot be switched on. Guest-facing analysis runs under GDPR with signage and a DPIA.
Published industry ranges for what operators recover after deploying AI video analytics. We would rather show you someone else's audited numbers than invent our own.
22% less identifiable theft after AI video analytics deployment
Industry benchmarkFrom a 20% cut in drive-thru service time — the top of the reported 15–20% range
Industry benchmarkFrom traffic-based scheduling — the top of the reported 5–15% range
Industry benchmark25% fewer food safety violations across monitored sites
Industry benchmarkSource: published 2026 industry benchmarks for AI video analytics in restaurant chains, which also report a typical payback period of 3–6 months. These are sector-wide figures, not MISE customer results — we will publish our own the moment we have audited ones.
No new cameras, no rewiring, no ceiling work. Priced per location, per month — never per camera.
Send your camera list or an NVR screenshot. We tell you exactly which detections your current angles support — and which they do not.
An edge appliance goes next to your recorder and pulls the RTSP streams you already have. Nothing else in the building changes.
Managers start getting plain-sentence alerts on their phones, and a shift report that reads like someone was actually there.
No. If a camera exposes an RTSP or RTSPS stream — which covers ONVIF-compliant IP cameras and almost every NVR sold in the last decade — MISE can read it. Analogue cameras already feeding a recorder work too, via the recorder.
Camera placement matters more than camera specification. A 2 MP feed at 10 fps pointed at the right angle beats a 4K feed pointed at a wall.
Not by default, and not at all if you do not want it to. Inference runs on an edge appliance inside your building. What leaves the site is event text and timestamps — "Lane 2 stalled, 4:10 wait, 14:22:07" — not video.
If you want clip review at head office, that is an option you switch on per site, not an architectural requirement.
No, and the system is not built to. MISE tracks people as skeletons and motion paths with a temporary ID that expires when they leave frame. We do not match, index or store faces, and we do not infer emotion or intent about any individual.
This is a deliberate constraint. It keeps us clear of the EU AI Act's prohibited practices and makes the legal review at your end considerably shorter.
Two different numbers, and we will not blur them. Continuous tracking of every person runs at 30 fps on all streams. Language-level scene understanding runs at 242 ms per frame on a Jetson Orin class device, gated to the moments the tracking layer flags.
Anyone telling you they run a full vision-language model at 30 fps on twenty cameras from one edge box is describing hardware that does not exist yet.
Per location, per month — not per camera. A restaurant with 18 cameras pays the same as one with 8, because charging per camera punishes exactly the operators who have the best coverage.
Multi-site groups get a single rate across the estate. No hardware markup on the edge appliance.
MISE reports on process, not on people. Alerts are addressed to a station and a shift, not to an individual — "the assembly station has been the bottleneck for 40 minutes", not "Maria is slow".
The staff-facing features operators actually keep switched on are the safety ones: slip and fall detection, distress alerts, lone-worker monitoring after close. Those are the ones crews ask for.
Fifteen minutes, your camera list, and an honest answer about what your current setup can and cannot do.