Applied ML
or AI research
— two doors, both open.

Reference-letter template (industry / Tech Nation): 'I worked with [Engineer] at [Project / Company] from [Year]-[Year]. They led the design of the [specific model / system / pipeline] that now serves [N inference requests / serves N users at named companies]. They authored [paper at NeurIPS / blog post] explaining the work, which influenced [downstream adoption]. Among ML engineers I've collaborated with on [sub-domain], they're in the top ~5.'

Reference-letter template (academic / Royal Society / RAEng): 'I have known [Researcher] since [Year] when they joined [my lab / our collaboration on Project X]. They are first / last author on [N papers] at [NeurIPS / ICML / ACL] including [single paper highlight, citation count]. Their independent research direction in [sub-area] has shaped my thinking and the work of [N other researchers]. I consider them among the strongest [career stage] researchers I have worked with.'

Innovation-criterion narrative example: 'Released [open-source model / library] under [license], trained on [scale]. As of [date], [N] downloads on Hugging Face and used in production at [named users]. The technique introduced [specific novel contribution, e.g. scaled attention variant], measurably improving [benchmark / production metric] over [prior SOTA].'

Recognition narrative example: 'Invited NeurIPS 2024 workshop talk on [topic]; Area Chair at ICML 2024 (handled [N] submissions); 4 published first-author papers at NeurIPS / ICML / ICLR with citation counts of [N total / [N] in 2024]; advisory at [N] funded AI startups; sustained AI-safety contribution including [AISI engagement / DSIT input].'

Most professional roles in the Global Talent visa map cleanly to one endorsing body — software engineers go to Tech Nation, postdocs go to Royal Society or British Academy, architects go to Arts Council via RIBA. ML / AI engineers are unusual: the field straddles industry and research, so the same person can plausibly fit either path.

The decision principle: pick the body whose panel will recognise your strongest evidence as exceptional. If you have first-author NeurIPS / ICML / ICLR / ACL papers + citations + peer review duties + collaborator letters, the Royal Society or RAEng panel is full of academics who weight that evidence accurately. If you have productionised models + engineering recognition + open-source maintainership + named-conference industry talks, the Tech Nation panel is full of practitioners who weight that evidence accurately.

The mistake is picking the route you think is 'easier'. Both routes have rigorous criteria. A Tech Nation Talent application with thin industry evidence will fail; an academic application with thin publication record will fail. Honest evidence-route fit is the strongest predictor of success.

Hybrid careers — academic-trained ML researchers now in industry, or industry ML engineers with research outputs — should pick the route matching the dominant 60% of evidence and use the rest as context. A research-trained Tech Nation applicant can lead with productionised work and use papers as supporting innovation evidence; an industry-rooted academic-route applicant can lead with publications and use deployment as 'significant contribution' evidence.

Tech Nation evaluates ML engineers on the same four criteria as software engineers but the evidence catalogue tilts toward AI-specific signals. Strong industry evidence includes: top-N maintainer of a major ML library (PyTorch, TensorFlow, JAX, scikit-learn, HuggingFace Transformers, LangChain, vLLM, Ray, Lightning) with verifiable adoption; HuggingFace model checkpoints with substantial downloads + named-user adoption; productionised models at named-brand companies with measurable business impact; advisory roles at funded AI startups; named industry-conference invited talks (NeurIPS workshop keynotes, ICML Industry Day, KubeCon AI / ML Day, Re:Invent AI / ML keynotes, Strange Loop AI tracks).

Public technical writing is well-rewarded — high-readership blog posts on Distill / The Gradient / lilianweng.github.io / Hugging Face blog / OpenAI / Anthropic / DeepMind blog with verifiable audience numbers. Lectures at named industry events (Stanford CS25, MIT 6.S191 invited speaker, OpenAI / DeepMind / Anthropic invited talks).

What doesn't clear the bar on its own: internal promotions, benchmark numbers without context, conference attendance (only invited speaking), Kaggle medals (used to corroborate, not lead). The recurring rejection-pattern is 'I led the ML team at Company X' — internal scope without external recognition.

Royal Society and RAEng peer-review routes weight publication record, citation impact, peer-review duties, and senior referee letters. Strong research evidence includes: 3+ first-author or last-author papers at top venues (NeurIPS, ICML, ICLR, ACL, NAACL, EMNLP, CVPR, AAAI, IJCAI, KDD); citation count corresponding to senior-IC level in your sub-field; sustained peer-review or Area Chair duties; workshop / special track organisation; PI or co-PI on competitive grants (UKRI, ERC, NSF, NSERC equivalents).

Senior referee letters matter materially. Three letters from Royal Society Fellows, RAEng Fellows, full professors at named universities (UK or international), or senior figures at leading AI labs (DeepMind, OpenAI, Anthropic, Meta FAIR, Microsoft Research, Google Brain). Letters should attest to specific work, not just status.

What doesn't clear the bar: industry titles without research output, Kaggle / leaderboard results without published work, attendance at workshops without speaking, GitHub maintainership without published research. The academic panel weights peer-reviewed publication far higher than industry adoption.

The UK has invested heavily in AI safety research and policy infrastructure post-2023: the AI Safety Institute (AISI) within DSIT, the AI Safety Summit hosted at Bletchley Park, follow-on summits, and growing public investment in alignment / interpretability / evaluations research. For ML engineers and AI researchers, this creates an unusually strong UK-side narrative for the mandatory 'significant contribution to UK digital economy / sciences' criterion.

Researchers with published contributions to alignment, interpretability, evaluations, or red-teaming have a natural route fit. Industry engineers building safety-relevant tooling (model cards, evaluation harnesses, monitoring systems) have a Tech Nation industry narrative tied to a UK strategic priority.

Practical: cite AISI engagement, DSIT-related advisory work, or contributions to UK-side research initiatives in the personal statement. The panel reads contemporary UK industrial-strategy context; aligning your narrative with current UK priorities is legitimately weighted.