Our Vision AI Services

For teams and businesses.

Start an enquiry

AI Services

person vehicle · mask

Annotation

Labelling to your schema, with the review pass that decides whether the labels are worth training on.

  • Bounding boxes
  • Image segmentation masks
  • QA pass and agreement scoring
the archive the keepers

Bulk image selection

You send the archive, we send back the keepers. The culling our app does for a camera roll, run at the scale of an event, a studio or a back catalogue.

  • Sharpness, duplicates, faces and scenes
  • Your selection rules, not ours
  • Delivered as folders or a manifest
evaluated on yours epochs loss ↓score ↑

Custom model training

Detection and segmentation models trained on your data, evaluated on your cases, and handed over in a form you can actually deploy.

  • Object detection and instance segmentation
  • Evaluation against your own benchmark
  • Weights and inference code delivered to you
Invoice # Line items Total

Custom OCR

Text recognition built for your documents and scenes, not a generic one — reading the fields you need out of the layouts you actually have.

  • Documents, labels, forms, signage and handwriting
  • Layout-aware: fields and tables, not just text
  • Delivered as a pipeline you run, with accuracy on your samples
REC low light, rain, odd angles

Data capture for training

Sourcing and capturing the images and video a model needs to learn from — including the awkward cases a scraped dataset never contains.

  • Image and video capture
  • Scenario and edge-case coverage
  • Consent and provenance recorded
raw dataset worth labelling

Data filtering

Raw datasets are mostly redundancy. We cut yours down to the frames worth labelling, so your annotation budget goes on data that teaches the model something.

  • Near-duplicate and low-quality removal
  • Coverage and class-balance analysis
  • A report on what was dropped and why
class balance

Dataset audit & model evaluation

Find out what is wrong with a dataset or a model before it costs you — label errors, gaps in coverage, and how a model really performs on your data.

  • Label-error and duplicate detection
  • Coverage, balance and bias report
  • Benchmarking an existing model on your cases
pre infer post latency ↓ CPU GPU NPU on the device

Edge device inference optimisation

Make a model run fast on the device it ships on — the whole pipeline, not just the weights: the right accelerator, the right runtime, and no time wasted around the model.

  • GPU, NPU and DSP acceleration (DirectML, NNAPI, Core ML, TensorRT)
  • Pre/post-processing, threading and memory tuned
  • Latency, battery and heat profiled on real hardware
FP32 → INT8 on device

Edge model compression

Make a model small and fast enough to run where the images are — phones, embedded boards or plain CPUs — without losing what makes it accurate.

  • Quantisation, pruning and distillation
  • Export to ONNX, TensorRT, TFLite or Core ML
  • Accuracy and latency measured on your target device
keypoints, tracked

Pose & keypoint estimation

Keypoints on people, hands or objects, for movement analysis in sport, ergonomics, fitness or robotics.

  • Body, hand and custom keypoint sets
  • Per frame, or tracked across video
  • Angles and movement measures you can use
redacted

Privacy redaction

Blur or mask faces, number plates and other personal details automatically, before images or video are shared, stored or published.

  • Faces, plates, screens and documents
  • Images and video, in batch or streaming
  • A redaction report for every file
seed generate

Synthetic data generation

Generated images for the rare classes and edge cases real data never has enough of — matched to your domain and labelled from the start.

  • Rendered or generated scenes for rare cases
  • Labels come with every image
  • Mixed with real data and validated against it
#1 #2 count line

Video analytics

Turn video into events and numbers: what moved where, how many, and when.

  • Detection, tracking and re-identification
  • Counting, zones and dwell time
  • Activity and event recognition
What’s damaged? dent · left door small VLM, tuned on your images

Vision-language model fine-tuning

Small vision-language models (sVLM) taught to look at your images and answer your questions about them — small enough to run where your data lives.

  • Small VLMs (sVLM) adapted to your images and questions
  • Captioning, visual Q&A and structured extraction
  • Sized to run on your hardware, on-prem or edge
defect

Visual inspection & defect detection

Automated visual quality control that spots scratches, cracks and misassemblies — trained even when defect examples are rare.

  • Anomaly detection from mostly good samples
  • Defects located, classified and measured
  • Runs on the line, on your cameras
1 Capture 2 Filter 3 Label 4 Train start at any step, stop at any step

Something in between

Most work is a mix — capture, then filter, then label, then train. Tell us where you are and what you are trying to reach.

  • End-to-end dataset pipelines
  • One-off audits of an existing dataset

How an engagement runs

No self-serve tier, no instant quote. The first step is a conversation about the data.

1

You tell us the problem

What the data is, roughly how much of it there is, and what the model or the delivery has to do.

2

We scope it

A written scope with a deliverable, a timeline and a price. Usually after a look at a sample.

3

We do the work

Capture, filtering, labelling or training — with checkpoints you see rather than a black box.

4

You get the artefacts

Datasets, manifests, weights and inference code. Yours, in a form you can deploy without us.

Tell us what you need

There is no self-serve tier for this. A person reads every enquiry.

Talk to us