Six stages between a camera and a system of record
The model is one stage of six. The others are what make the output trustworthy enough to post an invoice, reject a part or open a door.
The vision pipeline. Validation catches what the model got plausibly wrong; the gate decides who sees it next; review corrections make next month's model better.
Who this is for
This is for operations and finance teams whose documents arrive as images, for quality and site managers who need every unit or location checked rather than a sample, for product owners who want the camera in their app to do something useful, and for event organisers who need thousands of people through a door in an hour. In each case the test is the same: the machine must read the image reliably enough that a person only sees the exceptions.
You have, or can collect, real images from the conditions the system will face
A wrong answer has a known cost, so a confidence threshold can be set
Someone will own the review queue and the corrections it produces
Defect detection on a line or at goods-in, with a trained model for speed and a review queue for the borderline. Common in manufacturing.
Retail shelf and inventory
Planogram compliance, stock-outs and price-label checks from shelf photos taken on ordinary phones. See retail and e-commerce.
Badges, QR codes and access at events
Fast, offline-tolerant scanning on kiosks and phones. Built into our own Eventify platform; see check-in and badge printing.
Safety and compliance imagery
PPE checks, site conditions and signage audits from site photos, with every flag reviewed by a person before it becomes a record.
Scene and media understanding
Tagging, moderation and description of images and video for search and accessibility, using multimodal models with human spot checks.
How to decide
Choosing between a classic vision model and a multimodal LLM
Concern
Trained classic model
Multimodal LLM
Training data needed
Hundreds to thousands of labelled examples
Often none; a prompt and a few examples
Time to first result
Weeks, including labelling
Days
Latency and cost at volume
Milliseconds, near-zero marginal cost, runs on device
Seconds and per-call cost; cloud only
Flexibility
Narrow; new classes need retraining
Handles new layouts and questions in language
Explainability
Bounding boxes, scores, deterministic behaviour
Free-text reasoning; can be confidently wrong
Best for
High-volume, stable, narrow tasks; edge devices
Documents, varied layouts, prototypes, anything needing language output
A common path: prototype and label with a multimodal model, then train a small model for the high-volume stage and keep the LLM for the exceptions.
Process
From sample images to a gated pipeline
Collect and label
Real images from real conditions, including bad ones. An eval set with ground truth, labelled with your experts.
Prototype
Multimodal model first for speed. Measure against the eval set. Decide whether a trained model is needed.
Build the pipeline
Pre-processing, validation rules, confidence gate, review queue, integration, monitoring. Data flows designed with our AI data engineering team.
Run and improve
Staged rollout; review corrections feed the eval set; retrain or re-prompt on evidence.
What goes wrong
Example: an extraction model is tested on clean scans and deployed to phone photos taken in a warehouse. Accuracy collapses and nobody knows because there is no confidence gate and no review queue. We collect the ugly images first, gate everything and measure continuously, so the failure is visible before it is expensive.
Classic computer vision model or a multimodal LLM?
A multimodal LLM is the fastest way to get started on documents, scenes and anything that needs language in the output. A trained classic model wins when the task is narrow, volume is high, latency is tight or the image must stay on a device. Many pipelines use both: an LLM to prototype and label, a small model to run on the full volume. See multimodal.
How accurate will it be?
We do not quote a number before we have seen your images. Accuracy depends on image quality, variety and how the output is validated. We build an eval set from your real samples in week one and report measured results against it, then improve from there.
Can it run on a phone or on an edge device?
Yes, for narrow tasks: document capture, barcode and badge reading, basic classification and detection. Heavy scene understanding usually runs in the cloud. Our AI for mobile apps page explains how we route between the two.
How is a human kept in the loop?
Through a confidence gate. Outputs above the threshold flow straight to your systems; those below it go to a review queue where a person corrects them. Corrections feed back into the eval set and, where applicable, into retraining. The threshold is tuned to the cost of a wrong answer.
How long does a computer vision project take?
A document extraction pipeline on a known layout family typically takes four to eight weeks to production. Inspection or detection tasks that need a trained model and real-world image collection usually take eight to sixteen weeks.
Talk to us about your AI project
Tell us what you are trying to do, which systems are involved and what is in the way. We reply within one business day with an opinion and a suggested next step.