Home / Technology
Intelligent workflowsThe right intelligence for every data challenge.
Custom image processing, multiple AI models, intelligent evaluation and selective human review, designed around your data. Our value is not running inference. It is designing and managing the workflow that turns a model’s output into data you can rely on.
Different data, different methods.
A detector trained on street scenes misses vehicles seen from 300 metres up, and a general segmentation model has no class for a lung field. So we start from the task, not from a favourite model.
One location, four altitudes, four methods: segmentation from satellite, oriented boxes and tracking from the air, cuboids at street level, drivable area from a vehicle camera.
Custom processing and specialised models, combined per task.
Detection, segmentation, tracking and geometry each have models that do one thing well. We choose and chain them for the task, and compare their outputs instead of trusting any single one.
- Preparation first: tiling, enhancement, de-duplication and filtering before any model runs
- Detectors and oriented-box models for objects, at the right scale for the sensor
- Segmentation models for regions and surfaces
- Trackers that keep identity across frames
- Geometry from camera models, for example cuboids and distances
- Image-processing steps where a model is the wrong tool
Which models we use depends on the project and on your data rights and environment. We name them in the project documentation, not in a generic stack.
Candidate labels first. Then evidence.
Models generate candidate labels or structured outputs at scale. Every candidate is then evaluated, so each one carries evidence about how far it can be trusted.
Model agreement is not ground truth. Two models can share the same blind spot. Agreement is a signal we calibrate against reference samples, not a substitute for them.
Generate candidates
Boxes, masks, tracks or keypoints from the chosen models.
Check and score
Automatic checks on format, geometry and consistency; confidence and disagreement between methods for everything else.
Calibrate, then route
Scores are compared with reference samples labelled by people, and only then used to decide what is accepted and what a person checks.
Selective human review.
People spend their time where judgment changes the outcome. What goes to review is decided by the project’s quality requirements, not by habit.
Predefined quality
Acceptance criteria and error tolerances are agreed with you before production.
Low confidence
Items the evidence does not support go to a trained reviewer.
Methods that differ
When models or methods disagree, a person decides and the case is recorded.
Rare or high-stakes
Rare classes, safety-relevant objects and new conditions are reviewed regardless of confidence.
Every correction is information. Corrections are recorded with their reason, recurring errors are traced to the class, condition or model that causes them, and guidelines and thresholds are updated, so the next batch starts from what the last one taught us.
Automated acceptance is earned per project. It is used only after the routing rule has been checked against reference labels for that data, it stays under sampled audit, and projects that need every item reviewed get every item reviewed.
The workflow, step by step.
One photo: a pretrained detection model proposes boxes, and a simple rule decides what a person should check.
1 image · 1600 × 1067115 candidates94 accept · 21 reviewrouted items onlyDOTA · GeoJSON · COCOVisible while it runs. Clean when it lands.
In Point AI Studio you follow batches through review, look at labels in the browser, see quality by batch, and export in the format your pipeline expects.
Have a challenging data problem?
Tell us what you’re building and what data you have. We’ll suggest a practical way to process, check and deliver it.