The word does a great deal of work in a job advert and very little in a codebase. Across the 1.1 million UK careers pages we crawl, exactly one employer puts “multimodal” in a job title. Plenty of teams combine images, audio, video and text; almost none advertise it by that name, and some who do have not worked out what they are asking for.

Work that out before the brief, because one constraint decides the rest and it is not the model.

The paired-data audit that should precede the job advert

Text models learn from text. Multimodal models learn from alignment — an image beside a caption that genuinely describes it, a video beside a transcript that matches what was said. The pairing is the training signal, and hardly anybody holds it in useful quantity.

So the first exercise is inventory, not recruitment. Most teams find one of three things: plenty of both modalities and no reliable link between them; a link produced by something nobody trusts, such as filenames or timestamp matching; or genuinely aligned pairs in the low thousands, enough to evaluate against and nowhere near enough to train on.

That changes the hire. Hundreds of thousands of clean pairs call for somebody who can run a training job; four thousand calls for somebody good at driving an existing model and measuring where it fails. Advertising for the first while holding the second is the commonest way this appointment goes wrong.

Two models and a join, and why that is not a failure

Most shipped “multimodal” products are not jointly multimodal. They run a vision model that describes an image, or a transcription model that turns audio into text, then hand that text to a language model. Two models and a join.

Frequently the right architecture, and nobody should apologise for it. It is cheaper, each stage can be tested alone, and when the output is wrong you can see which stage was wrong. Debugging a fused model is closer to guesswork.

Joint understanding earns its cost in a narrower set of cases: where the meaning lives in the relationship between modalities rather than in either one. Tone contradicting the words. A gesture that changes what a sentence refers to. Spatial detail no caption would have thought to mention. If you cannot name such a case in your product, you want a strong LLM engineer plus a competent vision or speech component, and you will ship sooner.

Confident descriptions of things that are not in the image

This work has a failure mode of its own. A multimodal model will describe, in fluent and specific prose, an object that is not present. Ask what the sign says and it will tell you, whether or not the sign is legible. Ask which of two items sits on the left and it will pick one.

The mechanism matters: the language half can produce a plausible sentence from priors alone, so where the visual evidence is weak the text fills the gap instead of reporting it. Grammatical, assured and wrong is far more dangerous than obviously broken.

Ask how they caught it. Good answers demand evidence rather than conclusions — making the model point at the region a claim rests on, putting the same question to a blank or cropped image and checking it declines, comparing two crops and treating disagreement as a signal. Weak answers lowered the temperature. Counting and ordering belong to the same family, and anybody who has shipped volunteers that unasked.

The composite CV is the realistic target here

Very few people anywhere have trained a multimodal model end to end. Those who have mostly did it inside a handful of large labs, they are not meaningfully on the market, and their work is not the work you need done.

Accept a composite instead: real depth in one modality, credible working experience in the second. Usually a computer vision engineer who has spent recent years on language models, or a speech person who travelled the same way — the voice AI guide covers that side if audio is your second modality.

Depth in one with fluency in the other beats shallow familiarity with both. Somebody who has fought with resolution, calibration and real-world capture will spot your data problem; somebody who has only called a hosted endpoint will not. Accept adapter work in place of pre-training, and projects where only the alignment layer was trained while the encoders stayed frozen. That is nearly all applied work.

Questions that expose how the pairs were made

  • “Where did your pairs come from, and who checked them?” Scraped alt text, paid annotation, an operational by-product, one model captioning for another — each fails differently, and they should describe theirs unprompted.
  • “What proportion were wrong?” Anybody who has done this knows a number and it is usually uncomfortable. Whoever says the data was clean has not inspected it.
  • “How did you handle a missing modality at inference?” One side is often absent in production: no audio track, a corrupt image, an empty attachment. Happy-path systems collapse here.
  • “Which modality was the model actually using?” The good sign is having tested it — blank the image and see whether the answers move. Often they barely do, and the thing has quietly become a text model with an expensive habit.

Money, and why scarcity here is global rather than British

Indicatively £85,000 to £130,000 permanent for somebody genuinely able to do this, with day rates from around £650 and into the £800s where the work is training rather than integration. The top of that range competes with research salaries.

The thinness is not a British problem. No city anywhere holds a deep bench of multimodal engineers: the specialism is young, the training runs that produce the experience happen in very few places, and everyone else is chasing the same composite profiles.

Write the advert around the modality pair, not the word itself

“Multimodal AI engineer” attracts people who have attached an image to an API call. “Engineer to build video-and-transcript search across our archive” attracts people who have done that, and it signals what data exists — the part serious candidates want to know first.

Name the two modalities, say roughly how much aligned material you hold, and state whether the job is training or integration. Fewer applications, a better shortlist, and no interview where a strong engineer works out halfway through that the paired data they assumed was waiting does not exist.