Tech hiring Data & AI
Hire machine learning engineers who keep models accurate.
A machine learning engineer trains models on your own data and runs them as dependable software: fraud scores, demand forecasts, recommendations, vision and speech models. The work covers the whole life of a model, from the training pipeline to serving, monitoring and retraining once the world has changed.
01 The role
What a machine learning engineer does, and what they ship.
A machine learning engineer turns historical data into a model that makes a prediction, then makes that model a reliable part of your product. They build the features, choose and train the algorithm, set up an evaluation that reflects the real decision, package the model behind an API or a batch job, and keep watching it after release.
Training is the smaller part of the job. Models fail quietly: a feature is computed one way in training and another in production, a label leaks the answer, customer behaviour shifts and accuracy slides for months before anyone looks. The engineering lies in reproducible pipelines, honest validation and monitoring that catches decay early. A notebook with a good score is a hypothesis, not a system.
This is not the same job as an AI engineer. AI engineers build features on foundation models that someone else trained, using prompts, retrieval and tool calling. Machine learning engineers own models trained or fine-tuned on your data, where you need your own labels, your own metrics and your own serving infrastructure. Tabular prediction, ranking and custom vision still belong here.
What they ship
- Scoring models: fraud, credit risk, churn, lead quality
- Demand, price and capacity forecasts that feed operational systems
- Recommendation, search ranking and personalisation models
- Custom vision, speech or document models, trained or fine-tuned in house
- Training pipelines with versioned data, code and models
- Model serving: real-time endpoints and scheduled batch scoring
- Monitoring for drift, data quality and accuracy, with retraining triggers
02 Skills and stack
What machine learning engineers work with, and what is optional.
Most business problems on tabular data are still won by gradient-boosted trees, not by deep learning. Tell us what kind of data you have (tables, text, images, audio, time series), because that decides which experience matters.
Core
The role cannot be done without these.
- Python, written as production software
- Machine learning fundamentals and validation design
- Feature engineering and data pipelines
- Model training: gradient-boosted trees and deep learning
- Offline and online evaluation, tied to a business metric
- Model serving, in batch and in real time
- MLOps: experiment tracking, versioning, CI/CD for models
- SQL and working with large datasets
Tooling
Varies from team to team.
- PyTorch; TensorFlow in older codebases
- scikit-learn, XGBoost or LightGBM
- Hugging Face Transformers and Datasets
- MLflow or Weights & Biases
- Airflow, Dagster or Kubeflow Pipelines
- Amazon SageMaker, Vertex AI or Azure Machine Learning
- Feature stores: Feast or Tecton
- Serving: KServe, BentoML, Triton or ONNX Runtime
- Docker and Kubernetes
- Spark or Ray for distributed training and scoring
Adjacent
Useful, not required.
- LLM application engineering: prompting, retrieval, agents
- Fine-tuning of open-weight language models (LoRA)
- Model compression: quantisation, distillation, pruning
- Data engineering and warehouse modelling
- Statistics and experiment design
- GPU provisioning and cost control
- Fairness, explainability and model documentation
03 When to hire one
Hire a machine learning engineer for the right problem, or hire someone else.
The wrong hire for the problem is an expensive way to learn what the problem was. Check your situation against both columns.
Hire a machine learning engineer when
-
A data scientist’s model works in a notebook and nowhere else
The analysis is sound, but there is no pipeline, no serving and no monitoring. Getting from a promising result to a production service is the centre of this role.
-
You have labelled history and a decision made thousands of times
Approve or decline, which item to show, how much stock to order. Repeated decisions with recorded outcomes are where a trained model pays for itself.
-
A model in production is getting worse and nobody knows why
It was trained two years ago by someone who has left. There is no retraining job and no record of the data it saw.
-
Hand-written rules have become unmanageable
The fraud or routing logic is hundreds of conditions that contradict each other. A model with a proper evaluation can replace most of them.
-
A general-purpose model is too slow, costly or inaccurate
For a narrow, high-volume task, a small model trained or fine-tuned on your own examples can be cheaper and better than a large hosted one.
Look elsewhere when
-
You want an assistant, a chatbot or an agent
Features built on hosted language models need prompting, retrieval and evaluation skills, not model training. That is the AI engineer’s job.
Hire AI engineers -
Nobody knows yet whether the data can predict anything
Exploring the data, framing the question and testing whether there is any signal comes first. Hire the engineer once there is something worth productionising.
Hire data scientists -
The data is not collected, joined or trustworthy
A model cannot be better than its inputs. If events are missing or tables disagree, pipelines and a warehouse are the first investment.
Hire data engineers -
A report or a simple rule would settle it
If three thresholds in a SQL query solve the problem, a model adds cost and maintenance for little gain. Start with analysis.
Hire data analysts -
You want to test one idea before committing
A scoped proof of concept with an agreed success measure is cheaper than a permanent hire made on a hunch.
Post a freelance project
04 How we assess
Four checks, before a name reaches you.
A CV says what someone claims. We test what they can do, then send you the evidence with the shortlist.
-
Stage 1
Profile review
We look for models that reached production and stayed there: what was predicted, on what data, how it was evaluated and served, and what the candidate personally built. Competition results and course projects count for less than one maintained system.
-
Stage 2
Technical test
A timed test covering Python and software fundamentals, core machine learning concepts such as validation, leakage and metric choice, and a practical task that involves training or evaluating a model and reasoning about its results.
-
Stage 3
Conversation with our team
Our engineers ask the candidate to walk through a model’s whole life: how the problem was framed, how the validation was designed, how it was deployed, and what happened to accuracy afterwards. We listen for scepticism about their own results.
-
Stage 4
Review and levelling
Reviewers score accuracy, speed, problem-solving, communication and teamwork, then assign a level using our AI framework. The notes distinguish people who have trained models from people who have also deployed and maintained them.
-
Then
Your shortlist
Assessed candidates, with notes and scores.
The engine
Evalia
Assessments run on Evalia, the assessment platform we built ourselves. Across the ProDevs network, more than 100,000 assessments have been completed.
The levels we assign are public. Read the AI framework, or see how a hire runs from brief to offer.
05 Seniority
Three levels. What each one can own.
Levels differ in scope more than in speed. Match the level to how much direction you can give.
Junior
Owns a step in the pipeline
- Implements a feature transformation, a training script or an evaluation report to a brief
- Trains standard models with established libraries and explains the metrics
- Needs review to catch leakage, poor validation splits and misleading scores
- Works best in a team that already has pipelines and a senior ML engineer
Mid-level
Owns a model in production
- Takes a model from framed problem to deployed service, with tests and monitoring
- Designs the validation scheme and chooses a metric that matches the decision
- Builds retraining pipelines and investigates drops in accuracy
- Explains to product owners what the model will get wrong, and how often
Senior
Owns the ML platform and roadmap
- Designs shared infrastructure: feature pipelines, model registry, serving, monitoring
- Decides whether a problem needs a model at all, and whether to build or buy
- Sets standards for reproducibility, review and release of every model
- Balances accuracy against latency, cost and regulatory limits, and mentors the team
Indicative salary
Published band for data and ML engineering. Annual, in US dollars, for remote roles with US, UK and EU companies.
- Mid-level
- $35–65k
- Senior
- $65–100k
- Typical US equivalent
- $140–220k
Indicative market ranges as of August 2026, based on ProDevs placement experience and published market data. The exact figure depends on stack, seniority and time-zone requirements. See the full cost breakdown.
06 Interview questions
Questions for machine learning engineers, and what to listen for.
Borrow these for your own interview loop. Beside each is what to listen for.
-
Tell me about a model you deployed. What happened to its accuracy over the following year?
A good answer shows: A real account of life after launch: what was monitored, when performance moved and what they did about it. Candidates who have only trained models end the story at the test-set score.
-
Your model scores 0.98 AUC offline and performs poorly in production. What do you check?
A good answer shows: Suspicion of the offline number first: label leakage, a random split where a time-based split was needed, or duplicates across train and test. Then training–serving skew in how features are computed. Reaching for a bigger model misses the point.
-
How do you choose the evaluation metric for a fraud model where one case in a thousand is fraud?
A good answer shows: Accuracy is useless here. They talk about precision and recall at an operating threshold, the cost of each kind of error, and calibration if the score drives a decision. The best answers ask who reviews the flagged cases and how many they can handle.
-
When would you use gradient-boosted trees instead of a neural network?
A good answer shows: For most structured, tabular problems: faster to train, easier to debug, strong on modest data. Neural networks earn their place with images, text, audio or very large datasets. Defaulting to deep learning for everything suggests limited production experience.
-
How do you make sure a feature is calculated the same way in training and at serving time?
A good answer shows: One shared definition of each feature, through a feature store or common code, plus point-in-time correct joins so training never sees the future. They can describe a bug this caused and how they detected it.
-
How would you decide between real-time and batch scoring for a new model?
A good answer shows: They start from when the decision is made and how fresh the inputs must be. Batch is simpler and cheaper and is right more often than people expect. Real time brings latency budgets, fallbacks and on-call, and should be chosen deliberately.
-
When would you fine-tune an open model instead of calling a hosted one?
A good answer shows: A reasoned trade-off: volume, latency, data residency, a narrow task with good labelled examples. They are frank about the cost of owning training and serving, and can say when a hosted model with better prompts is the sensible choice.
07 Ways to engage
Permanent, team, payroll or project. Four routes to a machine learning engineer.
A permanent hire is the usual route, not the only one. Choose by how long the work lasts and who should employ the person.
- Placement fee, of first-year salary. One-time.
- 10–15%
- Replacement guarantee on direct hires.
- 2 weeks
- To a first shortlist, for most roles.
- 48h
-
Direct hire
A permanent employee on your contract and your payroll. One-time placement fee of 10–15% of first-year salary.
Tech hiring -
Dedicated team
A machine learning engineer inside a team we assemble and run for you: product, design and engineering on your roadmap, for a monthly retainer.
Dedicated teams -
Outsourcing & payroll
A full-time person in a country where you have no legal entity. ProDevs employs and pays them: 20% of salary if we find the person for you, or 5% of salary, capped at $500 a month, for payroll and compliance on someone you have already chosen.
Outsourcing -
Freelance project
A defined piece of work with a clear end. You pay the agreed price per milestone through escrow, plus a 5% client fee. Free to post.
Freelancers
08 Around the role
A machine learning engineer is one seat. These are the others.
A machine learning engineer sits between data and product engineering. They depend on data engineers for inputs and on backend teams to consume predictions, and the role works best when a product owner is clear about which decision the model supports and what an error costs.
- Data scientist
- Takes a validated approach and rebuilds it as a tested, repeatable pipeline.
- Data engineer
- Agrees which tables and events feed the features, and how fresh they must be.
- Backend developers
- Defines the prediction API, its latency budget and what happens when it is unavailable.
- DevOps engineer
- Shares compute, GPUs, deployment pipelines and cost monitoring.
- Product owner or domain expert
- Sets the acceptable error rate and reviews cases the model gets wrong.
09 Client testimonials
Hear it first-hand, from companies that hired.
Founders, operators and executives describe working with ProDevs, in their own words.
Play video: Client testimonial: Olufemi Seyi, COO, Casafina development company
Play video: Client testimonial: Aproko Doctor, Building AwaDoc
Play video: Client testimonial: JT Liddell, Promenade CEO
10 Questions
Before you hire a machine learning engineer: the usual questions.
Something else? Book a call (opens in a new tab) and ask us directly.
-
What is the difference between a machine learning engineer and an AI engineer?
A machine learning engineer trains, evaluates, deploys and maintains models built on your own data, such as a fraud score or a demand forecast. An AI engineer builds product features on foundation models that already exist, using prompting, retrieval and tool calling. The first owns a training pipeline; the second usually never trains a model.
-
What is the difference between a machine learning engineer and a data scientist?
A data scientist works out whether a question can be answered from the data and which approach is valid, through analysis, experiments and prototype models. A machine learning engineer turns a working approach into software that runs every day and keeps running. In small teams one person does both; the skills separate as soon as models reach production.
-
Do we need machine learning at all now that large language models exist?
For many problems, yes. Language models are strong on text and reasoning but are a poor and costly fit for scoring millions of transactions, forecasting demand or ranking a catalogue. Those tasks rely on models trained on your own structured data. Many products end up using both, for different jobs.
-
How much data do we need before hiring a machine learning engineer?
There is no fixed threshold. What matters is having recorded outcomes for the decision you want to predict, enough examples of the rare case, and data that reflects how things work now. If you are unsure, a short analysis by a data scientist will tell you whether a model is realistic before you commit to a hire.
-
What does a machine learning engineer in Africa cost?
For remote data and ML engineering roles with US, UK and EU companies, indicative annual salary ranges are $35,000–$65,000 at mid-level and $65,000–$100,000 at senior level. A direct hire through ProDevs also carries a one-time placement fee of 10–15% of first-year salary, with a two-week replacement guarantee.
-
How do you assess machine learning engineers?
In four stages. We review models the candidate has put into production, set a timed technical test covering software and machine learning fundamentals, hold a conversation with our engineering team about a model’s whole life, and then review the results together to assign a level against our public AI framework.
Talk to ProDevs
Tell us about the role.
Tell us the level and the stack. We will tell you how quickly we can put assessed machine learning engineers in front of you.
Project work
Need it for a project instead?
Not every job needs a full-time hire. Describe the work, get proposals from vetted freelance machine learning engineers, and pay per milestone through escrow. Free to post.