Artificial Intelligence
ML Engineer — Ranking, RecSys, Risk
The classical role. Still large, especially in India, and the one most often rebranded rather than changed.
What the job actually is
Models that serve live traffic: what to show, what to rank, what to block. Feed ordering, search relevance, ads, fraud and credit decisioning. The LLM content in these roles is frequently a paragraph bolted onto a job that is otherwise unchanged since 2022 — which is not a criticism, because this is where a great deal of measurable business value still sits, but it does change how you should prepare.
Titles this hides behind
Note the "Applied Scientist" convention — an Amazon lineage, and in India it very often means ranking and recommendations specifically rather than research.
Who is hiring it
India is the stronghold: Meesho, InMobi, Glance, PhonePe, Flipkart, Swiggy, Navi, Slice. In the US: Reddit, Pinterest, DoorDash, Instacart, Robinhood, Coinbase. Recommendations and ranking appear in 38% of Indian AI JDs against a far smaller share of US ones — the clearest geographic difference in the entire dataset.
How to recognise it in a JD
Look for ranking, personalisation, relevance,
CTR, A/B, feature store, Spark,
underwriting. If Spark and SQL appear near the top, this is archetype 3
whatever the title promises.
The loop
Prepare like this
- Own one end-to-end story with real numbers — baseline, change, measured lift, and how you knew it was the change rather than seasonality. Vague numbers read as borrowed credit.
- Be sharp on the offline-online gap, because it is the round most candidates lose. Training-serving skew, feedback loops, position bias in logged data.
- Separate model metrics from business metrics out loud. NDCG went up and revenue did not — be able to explain how that happens and what you do next.
- Refresh the classical material. This is the one archetype where gradient boosting, calibration and feature engineering still carry an interview.
Be ready for
- "Offline AUC improved, the A/B was flat. What are your hypotheses, in order?" The signature question of this archetype.
- "Your training data comes from what your current model showed users. What is the problem?" Position and exposure bias — and whether you have thought about counterfactual logging.
- "How often do you retrain, and how would you know that is right?"
Red flags in the posting
An LLM paragraph with no LLM substance. If "GenAI" appears once in the responsibilities and never again in the requirements, expect the actual work to be gradient boosting on tabular data. Fine if you want that — costly if you took the role to move into LLM work.
No mention of experimentation. A ranking team without A/B infrastructure cannot tell you whether your work mattered, and neither can your promotion packet.
Compensation
A day that is actually representative
| Time | What you are actually doing |
|---|---|
| 09:00 | Check yesterday's A/B. The offline gain has not appeared online, which is the defining event of this job. |
| 10:30 | Feature work. Adding a signal, and half the effort is making sure it is computed identically offline and at serving time. |
| 13:00 | Training run review. Loss curves, and whether the improvement is real or the eval split leaked. |
| 15:00 | Serving. p99 on the model call is creeping up and the ranking service has a latency budget you cannot exceed. |
| 16:30 | A conversation with product about why the metric they want optimised is not the metric that will move revenue. |
A long-running public lesson from recommender systems: offline metric improvements routinely fail to reproduce in live tests, because logged data reflects what the previous model chose to show.
Use it to say: “It is why I check exposure bias before anything else when an A/B goes flat.”
An algorithmic home-buying business was wound down after the pricing model's errors, compounded by scale and by acting on its own predictions, produced very large losses.
Use it to say: “My takeaway is that a model acting on its own predictions changes the distribution it was trained on. That feedback loop needs monitoring as a first-class risk.”
Week 1 — the end-to-end story
Days 1–3: Build a ranking or fraud model end to end on real data, with a proper temporal split rather than a random one.
Days 4–5: Deliberately introduce training-serving skew, then find it. Doing that once makes the interview answer permanent.
Week 2 — measurement
Days 1–2: Design the A/B: metric, minimum detectable effect, sample size, duration.
Days 3–5: Write the offline-online gap analysis you would run when it goes flat.
Week 3 — rehearse
Days 1–2: The ordered hypothesis list for the flat A/B, out loud.
Days 3–4: Refresh the classical material — calibration, class imbalance, feature leakage.
Day 5: Five stories, each with a number and a baseline.
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 →