Machine Learning Engineer

We are building a model that predicts which content creators will actually sell a given product, trained on real affiliate conversion data rather than follower counts. You would own it. Not the wrapper around it, the model, the features, the evaluation, the decision about whether it works. This is a grant-funded R&D role. Innovate UK is backing the research through its Next Wave programme, which means twelve months of protected time to answer a real research question, with a live product and paying customers on the other side of it.

Brands have shifted from paying a handful of large influencers to working with hundreds or thousands of smaller creators. It converts better and costs less. It is also operationally chaotic, and the tooling is bad in a specific and solvable way: every platform on the market ranks creators by follower count, category tag and engagement rate. None of them learns from whether the creator actually sold anything. We have the thing that makes this tractable, first-party outcome data from real affiliate campaigns running on our platform. Conversion rates, revenue per view, campaign-level results, linked to the creators and the content that produced them. That is the training signal nobody else in this market has, and it is what turns creator matching from a metadata lookup into a genuine learning-to-rank problem.

The second half is more unusual. We are also building a multi-modal model over creator video, visual features, speech transcripts, audio, to work out which hooks, formats and calls to action actually drive sales, and feed that back to creators as concrete guidance. The two models connect: content embeddings become features in the ranking model, so matching is informed by what a creator actually makes rather than how they describe themselves. We are not asking you to build a GPT wrapper. If you have spent the last two years writing prompts and want to go back to building models where evaluation is the hard part, this is that job.

What you would own:

• You lead WP2, the creator-to-product matching model, from month one to month twelve.
• Frame and build the ranking model. Our starting position is gradient-boosted ranking (LambdaMART) benchmarked against a two-tower retrieval comparator, but you own that decision and we expect you to argue with it.
• Build the feature set from campaign signals, audience data and content embeddings produced by the computer vision model.
• Define and run the evaluation. Our existing heuristic ranker is the control condition, and the project succeeds or fails on whether you beat it by a pre-registered margin on held-out campaigns.
• Temporal and group-wise splits, NDCG and precision at k, plus the business metric that actually matters: realised revenue per view of the creators you rank highest.
• Own the ablations that establish where the value comes from — structured features alone, content embeddings alone, and the fused model.
• Report honestly at four decision gates. If the model does not beat the baseline, we need to know at month six, not month twelve, and saying so is part of the job rather than a failure of it.
• Ship it. You are allocated to the integration work package alongside our platform architect, because we would rather each model went live with the person who built it than be handed over.

What we are looking for:

• Essential: Supervised learning on tabular and mixed-type data, in production rather than in notebooks.
• Ranking, recommender or search relevance experience — you have built something where the output was an ordered list and the ordering mattered commercially.
• Fluency with gradient-boosted methods, and the judgement to know when they are the right answer for structured data and when they are not.
• Rigorous offline evaluation: you think about leakage, split design and baselines before you think about architecture, and you can explain a time you killed your own model because the evaluation did not hold up.
• Python, and comfort working in a codebase rather than alongside one.

• Useful, not required: Small-data regimes and fine-tuning pre-trained models. Our dataset is measured in tens of thousands of examples, not millions, and the approach is built around that.
• E-commerce, marketplace, adtech or creator-economy domain exposure.
• Experiment tracking and MLOps practice, you will have a dedicated AWS environment and tooling budget.

We are not looking for a PhD, a specific number of years, or a particular pedigree. We are looking for someone who has beaten a baseline and can prove it.

What you would be joining:

Lumo is a small team building a B2B platform for affiliate and creator commerce. The founders previously ran Pear Growth, a TikTok Shop Partner agency named a top five partner agency for FMCG in March 2026, so the campaign data you would be modelling comes from people who ran those campaigns themselves. That matters more than it sounds: label quality is the hardest part of this problem, and we know what a good campaign outcome actually looks like. You would be the second and third technical hires alongside a computer vision engineer and a data engineer, working with our AI engineer who owns platform architecture.

Terms: Permanent, not fixed-term. The grant funds twelve months of research; the role is intended to continue beyond it, and the commercial case for keeping the capability in-house is the reason the project exists.

£70,000 – £85,000 depending on experience London-based, hybrid.

A note on funding: this role is funded by an Innovate UK grant awarded through the Next Wave: Breakthrough Wave 1 competition. Funding confirmation is expected on 1 October 2026, and offers are conditional on it. We would rather say that plainly than have you find out at contract stage.