Think of a translation app that turns English into Hindi and back. A general open model, such as one from the LLaMA family, can already translate a little, but it stumbles on everyday Hindi and gets names and numbers wrong. An AI model engineer gathered pairs of matching sentences, cleaned out the broken and duplicate ones, and fine-tuned the model on them in PyTorch, the Python library for building and training neural networks. Fine-tuning means training an existing general model a little more on narrower data, so that it fits one task. The engineer then measured the result on sentences the model had never seen, compared it with the old version, and asked native speakers to judge a sample. The tuned model was too slow for a phone app, so the engineer shrank it with quantization, storing its numbers with less precision so it runs faster and uses less memory, and set it up on GPU servers, machines built to do many small calculations at once, where it answers in a blink. When users later reported odd translations of cricket terms, the engineer went back to the data.

The work involves making a model better at one job and then making it usable. More of it is data work than the title suggests. There are datasets to collect and clean, training runs to start and watch, and charts to read for signs that a model has stopped learning. A run can take hours or days, so engineers keep several experiments going and note what each one changed. Evaluation, testing the model on examples it has never seen, decides whether a change is kept. The engineering that moves a model from a notebook to a service that answers real users quickly and cheaply is part of the core work too, and libraries such as Hugging Face Transformers turn up throughout.

Around the experiments sits a set of duties that every AI model engineer shares.

Experiments have to be repeatable. Engineers record each run's data, settings and results in a tracking tool such as MLflow or Weights and Biases, so that anyone on the team can see what was tried and rebuild the best model later. Results are written up for the team, and code is read by a teammate in a code review like any other software.

Training needs a lot of expensive hardware. Engineers book time on GPU clusters, run jobs on cloud platforms such as SageMaker or on the company's own machines, and keep an eye on the bill, because a careless run can waste days of GPU time. Once a model is serving users, they watch how fast it answers, what each answer costs and whether its quality slips.

Engineers also keep up with research. New methods appear in papers all the time, and part of the job is reading them, judging which ones matter and trying the promising ones.

A GenAI engineer takes a model that already exists and builds a product around it, while an AI model engineer is the person who made or reshaped that model. A machine learning engineer builds a model to answer one business question, such as the risk on a loan, and the model is a means to that answer. The AI model engineer works on general models, often language models, where the model itself is the thing being built. The servers and GPU clusters underneath belong to the cloud and DevOps engineer.

The ideal candidate enjoys maths and experiments, is comfortable reading research papers, and does not mind that many attempts fail. A strong grasp of linear algebra, probability and Python matters more than any one framework. People often come in through a master's degree, a research internship, or a few years as a machine learning engineer.

Over time the work can lead to owning a company's model training end to end, to research roles that try new methods, or to the specialised craft of making models run fast on particular hardware.