Machine Learning Engineer / AI Developer
Puts models into production and keeps them working — the serving, the evaluation, the monitoring, and knowing when the model is quietly wrong.
Who to hire
The shape of person who does this job well — and whether you should be hiring it at all yet.
The distinction that matters at hiring time: a researcher improves a model, an ML engineer ships one and keeps it honest. Most companies need the second and advertise for the first. With foundation models, a growing share of this work is integration and evaluation rather than training.
Education
Quantitative or engineering background. For applied work, software engineering discipline matters more than publications.
Experience
4+ years with something in production that people depended on. Ask how they knew the model was degrading, and what they did about it.
Vertical / domain experience
Decisive where a wrong answer is a regulated event — lending, hiring, clinical, insurance. In those settings, the evaluation and audit trail are most of the job.
What it pays, roughly
$106,080 median, Canada — for Mathematicians, statisticians, actuaries and data scientists, the closest occupation Statistics Canada actually measures, measured from $95,992 to $106,080 over 27,545 people. That occupation is broader than this role, so read it as an anchor, not a salary band — and note that a startup usually pays below it in cash and above it in equity.
Credentials and certifications
What exists, who issues it, and whether it means anything for this seat. A credential is evidence of a floor, never of judgment — and in a few of these roles it is a legal requirement rather than a nice-to-have.
| Credential | Issued by | When it matters | Skiploma |
|---|---|---|---|
| A note on the word “engineer” in Canada | Engineers Canada and the twelve provincial and territorial regulators | In a joint statement of July 2022 the regulators restated that “software engineer”, “computer engineer” and related IT titles prefixed with engineer may not be used by anyone not licensed as an engineer — in job titles, postings, résumés, websites or social media. APEGA obtained a court injunction against unlicensed use in 2019 and later pursued a software company over its postings. Alberta tabled Bill 7 in November 2023 proposing an exemption for the title; whether it is in force is not established here, so check your own province before writing it on a posting. This is why Canadian job ads so often say Developer. | — |
| NIST AI Risk Management Framework | US National Institute of Standards and Technology | Free and practical; what a cautious enterprise buyer will ask whether you follow. | — |
| Cloud ML certifications | Google Cloud, AWS, Microsoft | Evidence of platform breadth. Shipped systems outrank it. | — |
What this role does at each stage
The same title means three different jobs depending on how much money is in the bank. Read the stage you are at, then read the next one — that is what you are hiring toward.
Nobody. Use an API and find out whether the feature matters.
Usually still an API and a backend engineer. A dedicated ML hire before product-market fit is an expensive way to build a demo.
A seat where the model is the product or materially changes its cost base.
Can one person hold this and something else?
The whole point of the many-hats problem: which of these jobs merge safely into one person, and which merges quietly cost you money or control.
Merges well with
- Chief AI Officer (Founders & fractional leadership)
- Data Engineer / Data Developer
- Backend Engineer / Back-End Developer
- Director of Data / Head of Data
ML engineering merges with backend in most companies and with data engineering in the rest. A separate ML team under ten engineers is usually a mistake.
Keep separate from
No conflict, but whoever builds the model should not be the only person evaluating it.
Reading a candidate
What good looks like in an hour, and what should stop the process.
Good signals
- Starts with the evaluation, not the model
- Knows the failure modes of what they ship
- Can say when not to use a model
Red flags
- Benchmarks with no production evidence
- No monitoring after deployment
- Cannot explain a wrong answer
The same role in three kinds of company
This job changes shape depending on who is paying for it. An agency sells the hours, a contract shop delivers into someone else's estate, and a product company is building something it keeps — and the same title means something different in each.
In an agency
Agencies sell AI work heavily right now. Scope evaluation into the contract or you will be arguing about whether it works.
On a client contract
On contract, agree in writing what "good enough" means, measured how, before the build. This is the single biggest source of AI project disputes.
In a product company
In a product company, the monitoring matters more than the model: a model that silently degrades is worse than no model.
What this role runs on
The software categories this job lives in. Each links to our buying brief and to the research behind it.
| Category | Who the default is | What leaving costs | |
|---|---|---|---|
| Workflow Automation — BPA, iPaaS & RPA | Microsoft Power Automate across the board; by segment, UiPath (RPA), ServiceNow (enterprise workflow), Workato / Boomi / MuleSoft (iPaaS) and Zapier… | Hundreds of live automations wired into production systems, each with its own credentials, connectors and edge-case logic | The research → |
| Knowledge Management & Enterprise Search | Glean and Notion | — | The research → |
| GRC & Compliance Automation | OneTrust, with Vanta and Drata owning the automation tier | The evidence history | The research → |