ML Engineer
An ML Engineer builds and ships production ML systems — training pipelines, feature engineering, model integration, deployment, and monitoring. The center of gravity is making ML work reliably in production, not inventing new algorithms or tuning novel architectures. The most common misconception going in: this is mostly deep learning and hyperparameter tuning. Reality, straight from practitioner accounts: ML-specific code is "just a few percents" of the codebase — the rest is data pipelines, APIs, monitoring, and production hardening, which is why one well-known practitioner calls the role "a full stack developer of the data science." It's also not research: companies explicitly prefer strong engineers who can pick up ML over ML theorists trying to pick up production engineering discipline. One thing that's shifted fast: LLM and agentic-workflow work isn't a nice-to-have anymore — it shows up in the majority of current postings, and roughly a third of listings now use the title "AI Engineer" instead of "Machine Learning Engineer" for what is functionally the same job, so don't read the title too literally when you're job-hunting.
What matters most for this role
Added as a real domain-fluency gate after testing showed this archetype could rank #1 purely on generic senior-engineer temperament (ambiguity tolerance, systems design, coding intensity, outcome accountability) with no check against actual ML/data background — none of those dimensions are ML-specific. Mirrors physical_constraint_engineering (Embedded) and adversarial_threat_modeling (Security), both of which already function as real domain gates.
Seniority inferred from 'the ability to translate vague business problems into measurable ML objectives' — defining the problem and its success metric is itself a senior-level skill.
ML Platform/Infra variant requires designing feature stores and serving infra for many ML teams (Uber's Michelangelo: ~400 projects/5,000+ models) — a variant within the archetype, not universal.
Success is 'A/B test outcomes and business KPIs moved, not offline model metrics' — a model 'isn't successful because it achieves high AUC, but because it survives the real world.'
'The majority of your code is not tied to machine learning' but is still production code (pipelines, APIs, monitoring) that ships.
A day in this role
Expect feature-engineering and data-pipeline work feeding both training and inference, training and evaluating models against a specific business metric (fraud scoring, churn prediction, personalization, underwriting), and building the production system around the model — API integration, monitoring, and CI/CD. LLM fine-tuning, RAG, and agentic-workflow integration are now baseline expectations across current postings, not a trend confined to AI-native startups — fintechs like Affirm and Brex, a DevSecOps company like GitLab, and consumer/health names like Oura and Whoop all list this work in otherwise-generalist MLE postings. One practitioner's blunt summary: the job is "less about sophisticated transformer architectures and more about building resilient systems that can survive contact with reality," with most time spent getting a training set that faithfully represents the real-world distribution — once that's solved, a well-understood classical model is often good enough. At smaller companies, expect to build the platform and use it simultaneously, which one practitioner directly warns "is a quick way to burn out an individual or a team" if it isn't managed deliberately. At senior/staff/principal levels the job shifts toward setting technical direction and mentoring — driving roadmap for an ML domain, not just executing tickets. Success is judged on shipped, reliable, business-impacting systems, not model sophistication or research novelty.
Comp structure
Typical: $220K
Comp mirrors general software engineering at the same company — this is typically a title/focus-tag on the generalist SWE ladder, not a separately-compensated track, so equity weighting grows sharply with seniority the same way it does for SWE (roughly 15-17% of total comp at entry level rising to 50-55% at senior/staff at Meta and Google). Cross-company aggregate median total comp is around $272,500, running meaningfully above general SWE and data-engineer bands at the same companies — a current market premium on ML talent, particularly at senior/staff levels. Current postings back this up with real base-salary bands: Whoop lists $150K-$230K for senior/staff ML roles, Affirm runs $128K-$188K for MLE II up to $265K-$325K at senior staff, Sentry posts $240K-$300K for staff ML, and Faire lists $308K-$423K for principal — these are base-only figures without the RSU component, so they read lower than blended total-comp numbers like the $272,500 median above. Entry-level real-world example: Affirm's MLE I posts $130K-$170K depending on location, notably lower than big-tech totals for the same reason.
▸ Full compensation breakdown by level and company tier▾ Full compensation breakdown by level and company tier
Compensation by Company Tier
Total compensation (base + bonus + annualized equity) across five company tiers, at each career level. The same role pays very differently depending on where you take it.
ml-engineer · total comp (base + bonus + annualized equity) · P25–P75 band, P50 median
Equity Reality Check
The guaranteed money (base + bonus) against the equity upside. Startup equity is illiquid — the equity figure is annualized paper value at vest, not cash in hand.
Examples of real job postings
snapshot from 2026-07-12Real postings from the research corpus behind this archetype. Click one to read the actual listing.
How to test this cheaply
Pick a public dataset and take a small model all the way to a deployed, monitored endpoint (not just a notebook with a good accuracy number) — the deployment, monitoring, and "what happens when the input distribution shifts" parts are the actual job, and a notebook-only project won't tell you anything about fit.
Given how dominant LLM and agentic work has become in current postings, also try wiring a small RAG pipeline or a single-purpose agent end-to-end — that's now a majority-pattern requirement, not an optional extra. Also read Chip Huyen's ML Interviews Book chapter comparing MLE to SWE and to data scientist — it's a fast, direct way to see whether the actual center of gravity (engineering discipline over algorithmic novelty) matches what drew you to ML in the first place.
Do this role, or hire for it? Rate how much each trait actually matters. Role-holder and hiring-manager ratings are kept separate, and no single rating changes the model; ratings are aggregated with anti-gaming thresholds before they factor in.