Artificial Intelligence
Research Engineer / Scientist / Member of Technical Staff
The role everyone pictures. A tenth of the postings, and the hardest to enter laterally.
What the job actually is
Making empirical progress on model capability under a compute budget. In practice that is far more engineering than the word "research" suggests — data pipelines, training runs that fail at hour nine, evaluation harnesses, and a great deal of careful measurement. Member of Technical Staff is worth understanding as a convention rather than a level: several labs use it deliberately flat, so the title carries no seniority signal and you must ask about scope directly.
Titles this hides behind
Who is hiring it
Anthropic, OpenAI, Cohere, Mistral, and — the notable Indian entry — Sarvam, which is posting genuine foundation-model roles: ML Researcher for foundational models, ML Engineer for training infrastructure, and ML Engineer for data. That is new. Until recently, frontier research roles were essentially not available in India.
How to recognise it in a JD
Look for publications, NeurIPS, ICML,
pre-training, RLHF, ablation,
scaling laws. Only 6% of the corpus mentions publications at all, so when
it appears it is a real filter rather than boilerplate.
The loop
Prepare like this
- Reproduce one paper end to end and be able to defend every hyperparameter. A single deeply-owned reproduction beats five shallow ones.
- Interrogate your own past work to three levels before the deep dive. Why that architecture, why that baseline, why you believed the result.
- Write a training loop from scratch in PyTorch or JAX, no trainer abstraction, until it is muscle memory.
- Be honest about negative results. "We tried it, it did not work, here is what we learned" is a strong answer here and a weak one almost nowhere else.
Be ready for
- "You have 100 GPU-hours and a hypothesis. Design the experiment." Testing whether you can scope research to a budget — the actual daily constraint.
- "Your eval improved by 2 points. Convince me that is real." Variance, seeds, contamination, and whether the eval measures what you claim.
- "What is the strongest argument against your own last result?"
Read this before you commit a month to it
This is the highest-effort, lowest-yield archetype for a lateral candidate. It is roughly a tenth of the market, screens partly on credentials you either have or do not, and competes against people with full-time research track records. If you have publications or a research-adjacent history, pursue it. If you do not, the same month invested in archetypes 1, 2 or 6 has a materially better expected outcome — and those roles pay comparably.
Compensation
A day that is actually representative
| Time | What you are actually doing |
|---|---|
| 09:00 | A run died at hour nine. Before anything else you work out whether it was infrastructure or your code. |
| 10:30 | Read. Two papers, properly — not abstracts. This is real work and it is what the job is for. |
| 13:00 | Implement an ablation. The point is not the result; it is isolating which change caused it. |
| 15:00 | Look at outputs by hand. Aggregate metrics hide the interesting failures and reading a hundred samples is how you find them. |
| 16:30 | Write up a negative result honestly, because the alternative is that someone repeats it next quarter. |
A well-documented pattern across the field: published results that do not reproduce because of unreported hyperparameters, seed variance, or baselines that were not tuned to the same effort as the proposed method.
Use it to say: “It is why I report variance across seeds rather than the best run, and why I tune the baseline as hard as my own method.”
Public benchmarks leaking into pre-training corpora has repeatedly inflated reported capability.
Use it to say: “I check whether the eval predates the training cutoff, and I keep a private held-out set.”
Week 1 — reproduce
Days 1–5: Reproduce one paper end to end. Not skim — get the number, and note every place the paper was ambiguous.
Week 2 — defend it
Days 1–3: Write the ablation: which single change produces the result, and does removing it give the gain back.
Days 4–5: Interrogate your own past work to three levels of why. This is the deep-dive round.
Week 3 — rehearse
Days 1–2: The 100-GPU-hour experiment design, out loud.
Days 3–4: The strongest-argument-against-your-result answer, with a real one.
Day 5: Write a training loop from scratch, no trainer abstraction, until it is muscle memory.
The book for this field
LLM Systems Engineering
Seven chapters on the engineering underneath applied AI: the roofline and KV-cache arithmetic that explain every serving decision, quantisation with the evidence, serving stacks and the prompt-cache traps that cost real money, retrieval internals and whether long context replaced RAG, agents at the token level, evaluation with actual statistics, and what won in post-training.
Cross-cutting
Skills every archetype tests
These are shared across every field on this site — the same question asked in different vocabulary — so preparation here compounds rather than being spent once. The book above is specific to your field.
Practice the 22 AI questions → · Back to Artificial Intelligence →