“Deep learning engineer” is the vaguest title in this corner of the market, and the vagueness is the whole hiring problem. Across the 1.1 million UK careers pages we crawl, no employer uses this exact title. Not a thin market — no market. Most roles covered here have at least some adverts behind them. This one has none, because it is not a job title — it is a description of a method, in the way “regression engineer” would be.

The consequence is that the phrase carries no shared meaning, so everybody who applies has supplied their own. One person on your shortlist will have run a multi-week distributed training job; another will have fine-tuned a released checkpoint in a hosted notebook over a weekend. Both will describe themselves in the same words, and neither of them will be lying.

A description masquerading as a job title

Strip the phrase back and it means “uses neural networks”. That covers language, images, audio, video, sensor streams and ordinary tabular data — problems with almost nothing in common beyond the underlying maths — and it stretches from inventing an architecture to wiring an existing model into a web service.

So the first piece of advice is the least welcome one: in most cases, advertise something else. If the work is getting models into production and keeping them alive there, you want a machine learning engineer, and saying so will improve your applications immediately. If it is genuinely new architecture work with publication attached, you want an AI research engineer. The method-shaped title earns its place only when your modality is unusual enough that no standard label fits, and even then the body of the advert has to do the work the title cannot.

The advert has to name the modality and the verb

Two sentences repair most of the damage. The first names the modality, because the craft transfers far less freely than the vocabulary implies: somebody excellent at speech models has spent years on problems that never arise in document classification. The second names the verb, and only three are worth distinguishing:

  • Train. Building a model for your data from a random start, or close to one. Rare, expensive, and correct for a small number of teams.
  • Adapt. Taking a released checkpoint and fine-tuning, distilling or quantising it until it does your job. This is what the overwhelming majority of paid work in Britain actually is.
  • Integrate. Serving a model somebody else trained, with the evaluation, monitoring and latency engineering that entails. Valuable, routinely mislabelled, and a different hire.

Decide which verb you mean before anybody writes the advert. If you cannot, you are not ready to post it, and no amount of interviewing recovers the ambiguity later.

Pre-training is almost never your problem

This is the distinction that quietly decides your shortlist. Training a large model from scratch is a capital decision rather than an engineering preference. It requires a corpus you own or can license, a budget that survives a failed run, and somebody who has watched a long job diverge on day nine and knew what to do next. Very few British teams hold all three, and those that do already know it.

Everyone else should be adapting. So put the question directly: under what circumstances would you pre-train rather than fine-tune? A strong answer opens with “almost never” and then supplies the real conditions — a domain whose vocabulary is absent from public corpora, a tokeniser that mangles your inputs, a licence that rules out the obvious checkpoint, an input shape no released model accepts. Someone who reaches for a from-scratch build on your forty thousand labelled examples has just told you how they intend to spend your first two quarters.

The loss curve that goes flat, and the story behind it

If you screen for one thing, screen for what the candidate does when a model refuses to converge. It is the job, it is unglamorous, and it separates people more reliably than any architecture question: everybody can sketch a transformer, and only some have sat with a run that would not learn.

Ask for a specific instance. Not a methodology — an instance. What did the curve look like, what did you check, in what order, and what turned out to be wrong? Good answers are boring and ordered. They overfit a single batch first, to establish that the training loop can learn anything at all. They confirm that inputs and targets are still aligned after the last change to the loader. They look at the learning rate and the warm-up, at normalisation, at whether mixed precision is overflowing, at whether the shuffle is really happening, at augmentation aggressive enough to have destroyed the signal. And they will tell you the culprit is usually the data rather than the model.

Weak answers never rise above “I tuned the hyperparameters” or “I swapped the optimiser”. That is guessing with extra steps, and on rented accelerators you pay for it by the hour.

Paying for a role the market does not price

Indicatively, somebody doing adaptation work well in Britain sits around £70,000 to £110,000, reaching perhaps £140,000 where they genuinely lead training at scale. On contract, indicative day rates run from around £450 to £750, and higher for short specialist engagements.

Treat every figure as soft, and understand why: with no employer advertising this title, there is no sample of adverts from which a benchmark for it could be built. Anyone quoting you a median is quoting some other job’s median and relabelling it. Price instead against the real title you would otherwise have posted — a PyTorch engineer, most commonly — and let the verb you settled on move the number.

Why the pile arrives sorted by nothing at all

Advertise the phrase unfiltered and your applications will span something like a decade of ability, described in identical language, because vocabulary is free and experience is not. A student’s coursework project and a production training pipeline both get written up as “built and deployed deep learning models”.

The cheap fix is to demand numbers at application stage. One box on the form: describe a model you took to production, including the dataset size, what you trained on, how long the run took, and the metric before and after. People who have done the work fill it in four lines without effort. People who have not will write about frameworks. Make it a rule that no numbers means no interview, and the decade of variance collapses into something you can read in an afternoon.