What Is an ML Engineer?
A letting agency has five flats on its books, and an analyst has built a notebook that predicts their rents surprisingly well. Now the agency wants every new listing in the city priced automatically, every night, for years. The notebook cannot do that on its own. Closing that gap is the ML engineer's job.
This whole course uses the agency's five flats:
| Flat | Size (m²) | Rooms | Metro (km) | Rent (₹ thousand/month) |
|---|---|---|---|---|
| A | 40 | 1 | 2 | 14 |
| B | 60 | 2 | 1 | 22 |
| C | 80 | 3 | 3 | 26 |
| D | 50 | 2 | 4 | 16 |
| E | 70 | 2 | 1 | 22 |
The rule
An ML engineer turns a working model into a dependable daily system.
"Working model" means it gave good answers once, on data someone prepared by hand. "Dependable daily system" means it keeps giving good answers on data nobody has checked yet. Between the two sit five stages:
- Data: new listings arrive, get checked for missing or impossible values (a 0 m² flat), and are stored.
- Training: a script, not a notebook, rebuilds the model from versioned data and code, so any past model can be recreated.
- Evaluation: a new model is compared with the live one on the same held-out flats before anyone swaps them.
- Serving: the website asks for a price and gets an answer in milliseconds.
- Monitoring: someone watches the live errors, because the world changes under the model.
listing = {"size": 60, "rooms": 2, "metro_km": 1}
price = model.predict(listing) # must answer fast, every time, for yearsThat single line hides all five stages. Researchers invent new methods; data scientists answer questions with data; the ML engineer makes the answer arrive reliably.
The shortcut
For any task, ask "Which stage of the loop is this?" The stages form a loop: monitoring spots trouble, which sends you back to data and retraining. If a task fits no stage, it is probably not ML engineering.
Data
A new flat like B arrives: 60 m², 2 rooms, 1 km from the metro. The pipeline checks each field is present and sensible, then stores it.
The common mistake is treating "the model is trained" as the finish line. In a real system, training is one stage out of five.
Drills
Answer fast — trust the rule.
Which task is most clearly ML engineering rather than research?
City rents rise 10% over a year and the model keeps predicting last year's prices. Which stage should catch this?
Why version the training data as well as the code?
Before replacing the live model, an engineer scores old and new models on the same held-out flats. Which stage is that?
Remember
- An ML engineer turns a working model into a dependable daily system.
- Shortcut: place every task on the loop of data, training, evaluation, serving and monitoring.
- Trap: thinking the job ends when training ends.