Modern computer vision is one of the few disciplines where the engineer is not the main cost. The labelled data is. A team can appoint a strong candidate, hand them a well-specified problem, and still ship nothing for six months because nobody agreed who would draw ten thousand boxes, or to what standard.

So run this hire backwards from the annotation plan. If you cannot say how many labelled examples you hold, who produced them and how labeller disagreements were settled, you are not ready to interview.

How many labelled frames, and who is drawing the boxes?

A serious candidate answers with a number and a justification. Fine-tuning a pre-trained detector on a narrow problem — one viewpoint, three classes, a plain background — can work on a few hundred examples per class. Twenty classes with occlusion and small objects will want thousands each, and segmentation masks cost several times what boxes do.

Then settle who draws them. Three routes, three failure modes:

  • Your own domain experts. Highest quality, slowest, and underestimated: the expert already has a job, so budget their hours in writing.
  • An annotation vendor. Quick, priced per object, ignorant of your domain — which makes the guidelines document the deliverable that decides quality.
  • Model-assisted labelling. A weak first model proposes, a human corrects. Excellent value, though it bakes that model’s blind spots into the data unless somebody audits the corrections.

An engineer who has done this will write the guidelines themselves and ask for a few hundred images labelled twice, so agreement can be measured.

Label noise is a ceiling, and a bigger backbone will not lift it

If two experienced annotators agree only four times in five about what counts as a defect, no model trained on that data can be judged reliably above that line, and every figure afterwards partly measures the disagreement.

The tell is what they do when accuracy disappoints. The weak answer is a larger architecture or a longer schedule. The strong answer is to pull the worst-scoring examples and count how many are simply mislabelled. A quarter of apparent failures being label errors is unremarkable, and fixing the guidelines then buys more than any change of model.

If a candidate never mentions inspecting images by eye, don’t hire them for this. Vision is the one field where you can see your own mistake, and metrics alone are a blindfold.

Detection, segmentation, or one label per image?

Choose the model family from what you do with the output, not from a leaderboard. If the decision is present or absent, classification on a crop is cheaper to train and far cheaper to label. If you need to count or locate things, detection. If you need pixel-level area or shape — how much corrosion — segmentation, and the labelling cost climbs with it.

The good sign is a candidate who shrinks the problem. Turning a segmentation request into detection, or detection into classification on a fixed region, can cut the annotation bill several-fold. Somebody proposing the heaviest formulation available, citing a benchmark score, is optimising something that is not your project.

Where the requirement is really images and text together, you want a multimodal AI engineer instead.

Milliseconds per frame, and what you trade for them

Accuracy and throughput are one dial, so say in the brief where you want it set. Inspecting a part every four seconds needs a quarter of a frame per second, which affords a far larger model than a live tracker on a battery-powered camera could.

Candidates who have shipped speak in specifics: input resolution as the largest lever, a smaller backbone against a pruned large one, half precision, timing the whole path rather than the model alone. Anyone quoting only a published inference figure has not deployed one.

Name the target device before you write the advert

This line decides who is suitable, more than degree or years. A model served from a container with a current GPU behind it is broadly a training and evaluation job. The same model on an embedded module, a handset, or an industrial PC with no accelerator is an optimisation job, and people excellent at one are frequently ordinary at the other.

If the target is constrained, screen for it: quantisation and what it costs in accuracy, operators that refuse to convert, the gap between framework numbers and runtime numbers on the board. Our ONNX engineer guide covers that half, which on a tight target can exceed the modelling. If the camera is fixed and the scene controlled, ask whether you need a trained model at all — the OpenCV engineer guide makes that case.

Salary, day rates, and the money that should go to labelling

Indicatively, £55,000 to £90,000 permanent for somebody who has trained and deployed on a real problem, rising above £95,000 where the work combines vision with edge optimisation, robotics or a safety case. Day rates indicatively £450 to £700, higher where regulatory evidence is expected. Any range here is soft: the openly advertised sample is tiny.

Across the 1.1 million UK careers pages we crawl, nine employers put “computer vision” in a job title. The work hides inside machine learning, perception and robotics titles, so title-matching will find you almost nobody.

Set the annotation budget in the same conversation as the salary. Funding the engineer and not the data buys a capable person nothing to work with.

“It scored 0.94 in validation and failed on site.”

Everyone has a version of this story, and the story is the interview. Listen for whether they found the cause or patched the symptom:

  • How was the split made? Random frames from the same recordings is leakage: near-identical images land on both sides and the score inflates. Splitting by site or camera is honest and looks worse.
  • What differed about the deployed images? A replaced camera, a second site, winter light, a compression change in firmware. Each shifts the input distribution, and none appear in a training log.
  • Where did the confidence threshold come from? Tuned once on validation and never revisited is the usual answer. Ask which error costs you more; nobody else will decide it.
  • What measures it now? Something should sample live frames, get them labelled and compare, so drift arrives as a number rather than a complaint from the floor.

An engineer who treats launch as where measurement begins will keep your system working. One who treats it as delivery has handed you a model that was accurate once.