PRODUCTION
Multiple UK venues
Per-venue models and configuration
HOSPITALITY · CUSTOM SOFTWARE · COMPUTER VISION
Turning restaurant CCTV into live operational intelligence.
Eyezilla combines edge computer vision with sales, labour, weather and inventory data to show restaurant operators where demand is building, where service is slowing and how each venue is performing.
- Product
- Eyezilla
- Sector
- Hospitality operations
- Engagement
- Founder-built B2B SaaS
- Status
- Live in production
COMPUTER VISION
Edge processing
Detection, tracking and queue analytics
OPERATIONS
One decision layer
CCTV, POS, labour, weather and inventory
PROJECT SNAPSHOT
The engagement in three parts.
01
Situation
Restaurant operators could see sales and labour after the event, but the physical customer journey remained largely invisible. CCTV contained useful operational evidence, yet watching footage manually could not support live, multi-venue decisions.
02
What we delivered
We built a full B2B SaaS platform spanning on-site GPU processing, cross-camera tracking, data integrations, forecasting, operational evaluation, dashboards and remote fleet management.
03
Our responsibility
We owned product definition, computer vision, data integrations, application engineering, deployment and continued production operation.
What we now operate
A live operational platform spanning computer vision, integrations, forecasting and decision support across multiple UK restaurant venues.
- Per-venue camera geometry, models, thresholds and dashboards allow one product to operate across different layouts and service formats.
- Edge processing, fleet monitoring and data-freshness checks make the computer-vision system an operated service rather than an offline model demonstration.
Eyezilla is a real game-changer for us. Knowing exactly the behavior of our customers and the areas we have to focus on week after week is priceless.
Cristiano S.
Area Sensei & New Site Opening Manager
Eyezilla operator feedback
EDGE-TO-OPERATIONS ARCHITECTURE
From venue cameras to an operational decision layer.
Processing stays close to each venue while configuration, joined operational data and operator-facing outputs remain part of one managed product.
01 · System layer
Venue camera / NVR
02 · System layer
On-site GPU processing
03 · System layer
Event stream and cloud services
04 · System layer
POS, labour, weather and inventory joins
05 · System layer
Operator dashboard and alerts
SELECTED SYSTEM EVIDENCE
The layers that made live venue intelligence operational.
01
A real-time computer-vision engine for live venues
Each venue runs a producer-and-consumer pipeline on an on-site GPU device. Camera frames are captured from the existing NVR, masked to the useful regions and processed in batches. Fine-tuned person detection, single-camera tracking and appearance embeddings create usable track fragments without sending continuous raw video to a central cloud service.
Why it mattered
Processing close to the cameras reduced network dependency and made the system practical for several live feeds per venue.
02
Cross-camera journeys rather than isolated detections
Camera coordinates are projected onto a shared two-dimensional floor plan. The system combines appearance similarity with path plausibility, including walls, walking distance, time gaps and realistic movement speed. That allows fragments from different cameras to be joined into a more stable customer journey.
Why it mattered
Queue, station and end-to-end service analysis depend on understanding the same visit across several views.
03
A decision layer combining physical and business data
The platform joins computer-vision output with POS transactions, labour costs, weather and inventory on a consistent half-hourly timeline. Forecasting and evaluation models estimate demand, compare similar operating periods and classify the likely constraint behind each result. The dashboard presents these findings in business language rather than exposing raw model outputs.
Why it mattered
Restaurant teams needed to know what to investigate, not simply receive another set of charts.
ENGINEERING DECISIONS
Making multi-camera intelligence reliable across real venues.
Combine appearance with physical layout
Appearance embeddings alone become unreliable when staff uniforms, lighting and camera angles are similar. Eyezilla also evaluates whether two track fragments could represent a physically plausible journey through the mapped venue. This makes the identity decision more specific to the real environment.
Computer vision
Keep heavy inference at the edge
Continuous multi-camera video is processed on venue GPU devices. Cloud services receive operational results and manage configuration rather than acting as a mandatory relay for every frame. Remote access, deployments and monitoring run over a controlled mesh without requiring customer port forwarding.
Production architecture
Reconcile operational data instead of trusting one successful request
POS and labour integrations compare source totals with stored data and reload only a day that has drifted. The analytics mart waits for completeness across its required sources before building a period. This prevents a technically successful job from presenting a partial operating picture.
Data reliability
RESPONSIBILITY
What we owned end to end
Tested Works
- Product and UX
- Vision and data systems
- Edge and cloud infrastructure
- Deployment and operation
Operators contributed
- Venue context
- Operational priorities
- Site access and constraints
- Feedback on useful decisions
Founder-built product · ongoing operation and improvement
SERVICES USED
SELECTED TECHNOLOGY
PythonOpenCVPyTorchTensorRTPostgreSQLTimescaleDBAngularAWSDockerRabbitMQ
RELEVANT EXPERIENCE
Relevant if your situation includes…
You need to turn cameras or sensors into an operational product rather than an isolated detection model.
You need to combine physical-world evidence with existing sales, staffing, ERP or operational data.
You need software that can be configured and operated across several sites without rebuilding the product each time.
RELATED WORK
Two adjacent examples.
AURRENCY
Adapting computer vision for gallery security monitoring.
Client-specific anti-theft logic delivered through shared camera, job, scheduling and evidence infrastructure.
View case study
GKM TECH
Combining computer vision and expert review for goalkeeper analysis.
Customer platform, video-processing backend and specialist annotation workflow.
View case studySTART A CONVERSATION
Do you have valuable operational evidence trapped in cameras or disconnected systems?
We can help define the useful decision, test the difficult technical assumptions and build the complete path from input to operated product.