What this playbook covers
A frontier AI lab is one of the handful of organisations training models at the outer edge of what is possible — Anthropic, OpenAI, Google DeepMind and their close peers. The Research Engineer, or RE, is the person who makes the research real: writing the training and evaluation code, keeping the infrastructure honest, and turning a scientist's idea into an experiment that runs at scale. It is one of the most competitive roles in technology, and one of the best paid. DeepMind's acceptance rate sits below one per cent.
This guide is written for AI engineers in India and the UK who want a serious, unromantic view of how to get in. It covers which role is actually you, the rubric these labs reward, the interview loop lab by lab, how to build proof-of-work that clears the bar, the two-tier compensation market, and the practical routes in — remote, relocation and visas — from both Bengaluru and London. No motivational filler; just what the loop tests and what you can do about it over the next few months.
- Research Engineer is an engineering role — ML engineering, infrastructure and experiment implementation, not novel research with publications.
- The rubric is narrow and deep — PyTorch fluency, ML intuition on infra-critical topics, strong general software engineering, and one compelling past project you can defend.
- Rubrics differ by lab — preparing for Anthropic does not prepare you for DeepMind. The loop moves fast, roughly four to six weeks.
- Pay has bifurcated — frontier roles run roughly 200k to 800k US dollars; enterprise ML sits around 170k to 245k. Same title, very different cheque.
- India and UK both have routes in — DeepMind in London, GCCs and Sarvam or Krutrim in India, plus remote and relocation. Plan the visa early.
Research Engineer versus Research Scientist — which role is you?
Before you touch an application, get the role right. The single most common reason strong candidates get rejected is applying for the wrong track and being measured against the wrong rubric. Frontier labs split their research staff into two families, and they are genuinely different jobs.
A Research Scientist drives novel research. The role is judged on original ideas, publications and the ability to open a new line of enquiry that others follow. If your instinct on reading a new result is to ask what the next paper should be, and you have a track record of producing those papers, this is your lane.
A Research Engineer makes research work. You implement experiments, own the training and evaluation infrastructure, debug the distributed job that dies at hour nine, and squeeze the throughput that lets a scientist run ten ideas instead of three. The role is judged on engineering quality, ML intuition and delivery, not on a publication list. For most applied AI engineers, the RE track is both the better fit and the more numerous role — there are simply more RE seats than RS seats at every frontier lab.
| Dimension | Research Engineer (RE) | Research Scientist (RS) |
|---|---|---|
| Core output | Working experiments, training and eval infra, tooling | Novel methods, research direction, publications |
| What is measured | Engineering quality, ML intuition, delivery at scale | Originality, publication record, research taste |
| Typical background | Strong SWE plus deep-learning fluency | PhD or equivalent research output |
| Papers required? | Helpful, not mandatory | Effectively yes |
| Best fit for | Most applied AI engineers | Career researchers |
If you find yourself writing "I have no papers, but…" in a Research Scientist application, stop and apply for Research Engineer instead. You are not talking your way past the RS rubric — you are being measured against the wrong one. The RE loop is built to reward exactly the engineering strengths you are apologising for.
What the rubric actually rewards
Frontier labs are refreshingly consistent about what they test, even when the packaging differs. Across the RE loop, four capabilities come up again and again, and every strong candidate can name them.
Deep-learning library fluency
You are expected to be fluent in PyTorch the way a backend engineer is fluent in their web framework — not able to recite the docs, but able to reach for the right pattern under time pressure. Writing a custom training loop, implementing a loss from a paper, wiring a data loader that does not bottleneck the GPUs, reasoning about tensor shapes and device placement without a debugger. Interviewers watch how naturally you move, not whether you memorised an API.
ML intuition on infra-critical topics
This is where RE candidates separate. You need working intuition about the things that break at scale: mixed-precision numerics, gradient accumulation, how a learning-rate schedule interacts with batch size, why a distributed job stalls, where memory actually goes. You do not need to have trained a hundred-billion-parameter model, but you must reason about what would change if you did.
Strong general software engineering
REs write production code, so the loop still tests data structures, object-oriented design and system design. Do not let the glamour of the research fool you into skipping fundamentals — a candidate who is brilliant on transformers but cannot design a clean, testable module gets filtered. Our companion pieces on the question clusters that recur in AI-engineer interviews and the AI-flavoured system-design interview map this ground in detail.
One compelling ML project you can defend
Every loop includes a conversation where you describe a past ML effort in convincing depth. The bar is not "I used a model" — it is being able to explain the problem, the approach, the failure modes you hit, the trade-offs you chose and what you would do differently. This single conversation carries enormous weight, and it is the one thing you can prepare fully in advance.
Do not spread thin across ten half-finished projects. Interviewers can smell a shallow portfolio in ninety seconds. One project you have genuinely wrung out — with real numbers, honest failures and a clear account of the trade-offs — beats a dozen tutorials-with-a-twist every single time.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The loop, lab by lab
Here is the point most candidates learn too late: preparing for one lab does not prepare you for another. The loops move fast — roughly four to six weeks from application to decision — but they weight different things, and a rubric that wins at Anthropic can underperform at DeepMind. Tailor per lab.
Anthropic
Anthropic screens with a timed online coding assessment early in the process, so your general software engineering has to be sharp under a clock — this is not the round to be rusty on data structures. Beyond raw coding, Anthropic probes your understanding of Constitutional AI and safety-driven design. You are expected to engage seriously with why safety shapes the way models are built and evaluated, not to treat it as a box to tick. Candidates who can connect an engineering decision to a safety consequence stand out.
OpenAI
OpenAI's RE loop leans on strong general engineering plus practical ML depth, with an emphasis on shipping and scale. Expect the coding to be demanding and the ML questions to probe how you would make a real training or evaluation system faster and more reliable. The through-line is pragmatism: can you take an ambiguous research goal and turn it into infrastructure that runs.
Google DeepMind
DeepMind weighs publication history, systems-level thinking and the ability to lead large projects at scale. Even on the RE track, evidence that you can reason about a whole system — and that you have shipped or contributed to something substantial — carries real weight. Its acceptance rate is below one per cent, so the bar on every axis is high. DeepMind is headquartered in London, which makes it a local employer for UK candidates and a relocation target for everyone else.
| Lab | Screens hardest on | Signature round | What tips the decision |
|---|---|---|---|
| Anthropic | Coding under time pressure; safety reasoning | Timed online coding assessment | Engaging seriously with Constitutional AI and safety-driven design |
| OpenAI | General SWE plus practical ML at scale | Demanding coding plus systems-oriented ML | Turning an ambiguous goal into reliable infrastructure |
| Google DeepMind | Publications, systems thinking, project leadership | Depth interviews on past large-scale work | Evidence you can lead substantial projects; acceptance below 1% |
Building proof-of-work that gets you in
The RE loop rewards evidence over assertion, and the good news is that evidence is something you can manufacture deliberately over a few months. The aim is a small body of work that demonstrates exactly the four capabilities the rubric tests.
- Reproduce a known result from scratch. Reimplementing a well-regarded paper's training loop in clean PyTorch, and matching its reported numbers, signals fluency and rigour more convincingly than any certificate.
- Fix something real in the open. A merged pull request against a serious ML library — a training-loop bug, a numerics edge case, a data-loader speed-up — is a public, verifiable signal that you operate at the level the job demands.
- Own one project end to end. Pick a problem, take it to a genuine result, and write up the trade-offs honestly. This becomes the "one compelling ML effort" you defend in the loop.
- Show the throughput mindset. Even a small benchmark — "I cut training step time by 30% by fixing the input pipeline" — demonstrates the infra intuition REs are hired for.
Then make that work findable. Our guides on building a proof-of-work portfolio and choosing a specialisation across agents, evals, infra and the forward-deployed track go deeper on how to shape this into a coherent story rather than a scatter of repositories. For a frontier-lab RE application, weight the story towards training and evaluation infrastructure — that is the work you will actually be paid to do.
Put your proof-of-work where a hiring manager can find it without asking. A Verified Builder profile on AI Tech Connect, linking your reproduction repo, your merged PRs and a one-paragraph account of your flagship project, does for you exactly what the RE loop is trying to establish — that the work is real, and it is yours.
Compensation and the two-tier market
The most important thing to understand about pay in 2026 is that the market has split in two. A frontier-lab Research Engineer earns roughly 200,000 to 800,000 US dollars in total compensation, base plus equity, and a small cohort exceeds one million. An enterprise ML engineer doing serious, respectable work earns roughly 170,000 to 245,000 total. Same broad skill set, same nominal title in many cases — three to four times the cheque, depending purely on which tier of employer you join.
Specialisation compounds the effect. LLM specialists command roughly 220,000 to 280,000 US dollars, with demand for them reported to be rising sharply year on year. The highest-paid niches sit at the infrastructure edge, where a genuinely scarce skill meets an urgent need.
| Role or niche | Typical total compensation (USD) | Note |
|---|---|---|
| Frontier-lab Research Engineer | $200K – $800K+ | Base plus equity; a small cohort exceeds $1M |
| Enterprise ML engineer | $170K – $245K | The lower tier of the bifurcated market |
| LLM specialist | $220K – $280K | Demand rising sharply year on year |
| CUDA / GPU optimisation | $300K – $500K+ | Highest-paid niche; genuinely scarce skill |
| AI safety / alignment | $250K – $450K | Central to frontier-lab hiring |
| Distributed-training infra | $280K – $420K | The work REs are most often hired to do |
These are US-benchmark figures, and they anchor the global market because frontier labs compete internationally for the same people. When you translate them to London or Bengaluru, the equity component and cost-of-living differences move the number, but the two-tier structure holds everywhere. The practical lesson: which tier you target matters more to your earnings than another year of incremental skill-building. When an offer does land, negotiate it deliberately — our guide to salary negotiation in the two-tier market walks through how to read and move a frontier-tier offer without leaving money on the table.
India and UK routes in — remote, relocation and visas
Geography is no longer the barrier it was, but the routes differ by region and you should choose yours consciously.
From the United Kingdom
The UK is unusually well placed. Google DeepMind is headquartered in London, and both Anthropic and OpenAI run London offices, so a UK-based candidate can apply for genuinely local frontier roles rather than chasing a relocation. That means no visa hurdle for those already with the right to work, a shorter loop, and the ability to interview on-site. The competition is correspondingly fierce — London draws frontier talent from across Europe — so the rubric is where you win, not your postcode.
From India
India's route is more varied and moving quickly. The nearest frontier-adjacent employers are the global capability centres, or GCCs, that many labs and large technology firms now run in Bengaluru, Hyderabad and Pune, plus a growing set of domestic labs. Sarvam is building models across 22 Indian languages and has secured an IndiaAI H100 allocation; Krutrim is training its own foundation models. These are not identical to a frontier lab, but they are the closest domestic proving grounds, and strong work there is exactly the proof-of-work that opens a relocation conversation.
For the frontier labs themselves, Indian candidates typically enter through remote roles or relocation with visa sponsorship. Both are increasingly common for strong RE candidates, but both need planning — a UK Global Talent or Skilled Worker route, or a US route, takes months and should be started in parallel with, not after, your applications. Our dedicated guide to visas and relocation for AI engineers across India and the UK lays out the specific routes and timelines.
Do not treat the visa as an afterthought you sort out once you have an offer. Relocation timelines can stretch past the point where an offer holds, and a lab will not always wait. Scope your route — Global Talent, Skilled Worker, or a remote-first arrangement — before you interview, so that when the offer comes you can say yes with a date attached.
Common mistakes that sink strong candidates
Most rejections are not talent failures — they are avoidable process failures. Watch for these.
- Applying to RS with an RE profile. The single biggest self-inflicted wound. Match the track to your strengths.
- Preparing generically. One preparation plan for all labs guarantees you are under-prepared for each. Anthropic's timed coding and safety focus is not DeepMind's systems-and-publications lens.
- Neglecting fundamentals. Deep transformer knowledge does not save you if you cannot design a clean module or clear a timed data-structures round.
- A shallow, sprawling portfolio. Ten tutorials read as noise. One deep project reads as signal.
- Ignoring the visa clock. A yes you cannot act on for six months is a no in practice.
Where to go from here
The RE role rewards depth, honesty about trade-offs and demonstrable engineering — all of which you can build on a deliberate timeline. Pick the right track, aim your proof-of-work at training and evaluation infrastructure, prepare per lab rather than in general, target the tier that pays what you are worth, and start the visa conversation early. None of it is luck; all of it is preparation. And when the work is real, the last step is simply making sure the people hiring can see it.