SAM 3
Computer vision & video analytics
Detect, count, classify and track selected objects or activities in images and video.
- People counting
- Queue analysis
- Object tracking
- Occupancy
- Zone activity
Home / Applied AI solutions
Applied AI solutions
Practical AI. Built around your business.
Turn existing cameras, images, video and operational data into useful intelligence. We combine proven AI models, custom processing and lightweight integrations to solve specific challenges without unnecessary complexity.
Camera frame
Zone activity · this frameModel output
78people detected
15tables detected
Example ruleFront tables: 10 people. A rule such as “notify staff above 8” would raise an event here.
The opportunity
You may already have the data you need.
Cameras, video, images and business systems often hold information that is hard to measure by hand. We help identify what AI can extract and turn it into reports, alerts or workflows. Instead of starting with a complex platform, we start with the outcome you want.
- 01
Existing camera or data
A frame from a camera you already have, or files and records.
- 02
AI processing
Models find what matters: here, people and carts, inside a drawn checkout zone.
- 03{ "source": "camera_07", "zone": "checkout_lane", "people": 5, "carts": 6, "people_in_frame": 36 }
Structured information
Counts and events as records a system can store, query and chart.
- 04Example ruleCheckout lane: 5 waitingAbove the threshold you set: suggest opening another till.
Business action
An alert, a report or an update in the tool your team already uses.
Whether a camera or data source can be used, and how well, depends on its quality, access and the specific requirement. We check this first, on your samples.
What we build
From visual data to useful decisions.
We combine existing models, specialised processing and custom development into AI workflows for specific operational challenges.
Drone footage
VLM description · sampled frameAerial view of a riverside day-use area: a packed gravel car park on the far bank, a suspension footbridge across the river, small groups at the waterline.
AI-powered video understanding
Vision-language models and other techniques produce structured descriptions, find selected events and make video collections searchable.
- Event summaries
- Video classification
- Activity descriptions
- Searchable metadata
Crack segmentation · 201 mm measured
Custom AI pipelines
Preprocessing, inference, task-specific rules and output validation, combined into one workflow designed around your requirements.
- Image enhancement
- Multi-model processing
- Defect detection
- Specialised classification
- Data extraction
{
"source": "camera_07",
"zone": "checkout_lane",
"people": 5,
"carts": 6,
"people_in_frame": 36
}
- Dashboard
- Scheduled report
- REST API
- Alert
Integrations & automation
Structured AI outputs connected to supported business tools, dashboards, reports or notification systems.
- APIs
- Event reports
- Scheduled analytics
- Dashboard integrations
- Automated alerts
Practical applications
Real problems. Purpose-built AI solutions.
From existing CCTV systems to industrial images, the right combination of models and processing turns raw data into operational information.
Understand what’s happening in your space.
Compatible camera feeds become operational analytics on customer traffic, occupancy and queues.
- Processing
- DetectionTrackingZone rulesAggregation
- Intermediate output
- 36 people and 6 carts in frame; 5 people inside the checkout-lane zone.
- Example business result
- 5waiting in the lane. Over a day, per-minute counts become queue length and peak-hour reports.
- Typical outputs
- Customer counts, occupancy trends, queue estimates, peak-hour reports, zone activity.
Real model output
Detect issues before they become bigger problems.
Custom image processing and suitable models find selected defects, anomalies and inconsistencies.
- Processing
- Image processingSegmentationMeasurementTraining when required
- Intermediate output
- SAM 3 proposed 32 candidate regions. A task rule keeps the 27 inside the marked test area and drops 5, including the plate seam. Kept cracks total 201 mm, measured against the ruler in the photo.
- Example business result
- ReviewPart routed to an inspector with masks and measurements attached, and logged in the quality report.
- Typical outputs
- Defect detection, segmentation masks, inspection results, review queues, quality reports.
Real model outputRobotics & industrial AI →
Turn operational video into measurable activity.
Selected objects, zones and movements become operational statistics from compatible visual sources.
- Processing
- DetectionTrackingSpatial rulesEvent processing
- Intermediate output
- 4 forklifts, 14 pallets and 5 people detected in one frame.
- Example business result
- TimelineOn video, the same detections become movement events and zone activity by shift.
- Typical outputs
- Pallet or object counts, zone activity, equipment movement events, activity timelines, reports.
Real model output
Make video easier to understand and use.
Video processing plus AI models identify relevant events and produce searchable, structured summaries.
- Processing
- SamplingDetectionVLM on selected framesTemporal rulesValidation
- Intermediate output
- VLM: Aerial view of a riverside day-use area: a packed gravel car park on the far bank, a suspension footbridge across the river, small groups at the waterline.
- VLM: Same view, closer: cars parked in rows, people unloading near the boat ramp, small craft at the bank.
- Peak: 58 vehicles in frame
- Peak: 23 people in frame
- Scene change: car park leaves frame
- VLM: Camera tilts down to the footbridge over the river: two rafts with people at the near bank, no vehicles in view.
- Example business result
- 3 of 243frames sent to the VLM. Counting ran on every frame; descriptions only where needed.
- Typical outputs
- Event logs, scene descriptions, object and activity metadata, timestamps, human-reviewed summaries.
Real model output
How we build
The right AI for the right task.
Different challenges need different approaches. We choose models and processing based on accuracy requirements, deployment constraints and operating cost.
- Understand the problemThe information you need, the data sources, performance requirements and the business outcome.
- Select the methodsExisting models, image processing, and whether custom training is needed at all.
- Build the pipelinePreprocessing, inference, task-specific logic and structured output.
- Validate performanceTest on representative real-world data and find the failures and edge cases.
- Integrate & improveConnect approved outputs to your systems and refine as requirements change.
Not every pipeline needs a vision-language model.
“How many people are waiting at checkout?”
- Video frames
- Person detector
- Tracker
- Zone rule
- Queue count per minute
“What happened at the loading dock overnight?”
- Video frames
- Motion & object filter
- VLM on selected frames only
- Checks & review
- Timestamped event summary
Related: how we combine models, processing and review in Intelligent data workflows.
Designed for efficiency
Powerful AI without unnecessary complexity.
The most advanced model isn’t always the right choice. We design workflows around the task, use efficient processing first and heavier models only where they add something. The aim is a balance of accuracy, processing cost, infrastructure and maintainability.
Broad processingThe largest model runs continuously on every frame.
Task-optimisedefficient base processing relevant events advanced model when needed → structured output
Use proven models
Start with existing technology instead of training everything from scratch.
Process only what’s needed
Sampling, event triggers and efficient pipelines avoid unnecessary computation.
Train when it adds value
Adapt models on representative data when general-purpose ones can’t meet the requirement.
Build for the real environment
Camera quality, lighting, hardware, latency, integration and running cost, from day one.
Work with what you have
AI that connects to your existing systems.
Where technically compatible, we work with your existing cameras, video feeds, files and software, so AI outputs are usable in your current operations.
Data sources
IP cameras & RTSP streams
Image & video files
Existing databases
Processing
Preprocessing & samplingDetection, tracking, segmentationVLMs where justified
Cloud, on-premises or hybrid
Events & analytics
Task rules & zonesEvents & aggregatesReview queue for uncertain cases
Outputs
Custom dashboardsScheduled reportsREST APIsEvent notifications
We confirm support for specific camera models, platforms and tools during scoping, rather than listing integrations we haven’t verified for your setup.
From idea to application
Start small. Prove value. Scale when ready.
Each stage ends with something you can judge before committing to the next. Timelines are agreed per project.
- 01
Discover
The operational problem, existing data, required outputs and constraints.
Deliverable: scoped problem & success criteria - 02
Validate
A focused feasibility assessment or pilot on representative samples, with measurable criteria.
Deliverable: feasibility results on your data - 03
Build & integrate
The processing workflow, connected to the agreed output or business system.
Deliverable: working pipeline & integration - 04
Refine & support
Real-world performance reviewed, recurring failures fixed, optional ongoing improvement.
Deliverable: performance review & support plan
Questions before you start.
The practical things we check early, so a pilot answers the right question. For labelled training data rather than a working system, see Managed data operations.
Often, yes, depending on camera compatibility, access to the video streams, recording quality and the intended application. We evaluate these requirements during initial scoping.
Not always. Many problems can be approached with pretrained models and task-specific processing. Custom training or fine-tuning is considered when it is needed to meet the performance requirements.
Depending on the models, hardware, security requirements and workload, local, cloud or hybrid processing may be appropriate.
We define project-specific acceptance criteria and evaluate performance on representative data, including relevant difficult cases and operating conditions.
Cost depends on the task, data sources, integration requirements, processing infrastructure, accuracy expectations and ongoing support. A limited pilot can establish technical feasibility and realistic operating costs.
We can scope ongoing monitoring, model updates, quality review and additional integrations according to the project requirements. Outputs and review history can be followed in Point AI Studio.
Start a conversation
What could AI solve for your business?
Have existing cameras, video, images or a repetitive process you want to improve? Tell us about the challenge and we’ll help explore a practical AI approach.
- What you want to measure, detect or automate
- The data or cameras you already have
- Where the result needs to go: a report, a dashboard, an alert or another system