Every week a founder sends me a job description titled "AI Engineer" that actually describes three different jobs: a researcher who trains models, an engineer who ships LLM features, and someone who keeps GPUs busy. That's not a hiring problem. It's a vocabulary problem.
AI and machine learning job titles have multiplied faster than anyone agreed on what they mean. The same title means different things at Meta, at a frontier lab and at a Series A startup. If you are building an AI team, or trying to land a role on one, you need a map first.
Here is mine, from years of sourcing ML and AI talent in Big Tech and startups.
The short version: four families of AI and ML roles
Almost every AI or ML title falls into one of four families:
- Research — invents new methods and pushes model quality forward.
- Modeling and applied ML — trains, fine-tunes and evaluates models for a real product.
- AI product engineering — builds products on top of models, often LLMs, without training them from scratch.
- ML infrastructure — makes training and inference fast, cheap and reliable at scale.
When a title is confusing, ask one question: which of these four does the person spend most of their week on? That answer matters more than the title.
Research roles
Research Scientist
Research scientists explore new model architectures, training methods and algorithms. They read and write papers, run experiments, and decide which ideas are worth scaling. Most have a PhD in machine learning, computer science, statistics or physics, and a publication record at NeurIPS, ICML, ICLR, ACL or CVPR.
You need one if your product depends on a capability that doesn't exist yet. Most startups don't.
Research Engineer
Research engineers turn research ideas into working code at scale. They write training pipelines, run large experiments, debug distributed training and make results reproducible. At frontier labs this is one of the most important and hardest-to-hire roles: strong software engineering plus a deep understanding of deep learning.
AI Safety, Alignment and Interpretability Researcher
A specialised research track focused on making models safe, controllable and understandable: alignment techniques, interpretability, robustness and red teaming. Mostly at frontier labs and safety-focused organisations, increasingly at large AI product companies.
Modeling and applied ML roles
Machine Learning Engineer (ML Engineer)
The broadest and most common title. A machine learning engineer builds, trains, deploys and maintains ML models in production. Depending on the company, an ML engineer might work on recommendation systems, ranking, fraud detection, forecasting, computer vision or LLM fine-tuning.
The key distinction: an ML engineer owns models in production, not just in a notebook. Look for experience with PyTorch, feature pipelines, model serving, monitoring and A/B testing.
Applied Scientist
A title used heavily at Amazon, Microsoft and a growing number of startups. Applied scientists sit between research and engineering: they adapt state-of-the-art methods to a specific business problem and measure the impact. Expect strong ML fundamentals, experimentation and statistics, often a PhD or a research-heavy Master's.
Applied scientist vs research scientist: the applied scientist is measured on product impact; the research scientist on advancing the state of the art.
Data Scientist
Data scientists answer questions with data: analysis, experimentation, statistical modeling and sometimes predictive models. In 2026 the title covers everything from product analytics to building ML models, so read the job description carefully. If the role requires shipping models to production, it is closer to an ML engineer.
Domain specialists: Computer Vision, NLP, Speech, Recommender Systems
Many ML engineers specialise by domain:
- Computer Vision Engineer — image and video models: detection, segmentation, generation, multimodal.
- NLP Engineer — language models, text classification, search, information extraction. Much of this work now overlaps with LLM engineering.
- Speech / Audio ML Engineer — speech recognition, text-to-speech, voice agents.
- Recommender Systems / Ranking Engineer — feeds, search ranking, ads, personalisation. Still one of the largest ML job markets.
- Robotics and Embodied AI Engineer — perception, control and learning for physical systems.
Evaluation (Evals) Engineer
A role that barely existed a few years ago. Evals engineers design benchmarks, test sets and automated evaluation pipelines to measure whether a model or an AI feature actually works — accuracy, safety, regressions, hallucinations. If you ship LLM products, someone has to own evals. Increasingly that is a dedicated hire.
AI product engineering roles
AI Engineer
The title of the moment, and the most overloaded. In most startups an AI engineer is a software engineer who builds products on top of foundation models: retrieval-augmented generation (RAG), agents, tool use, prompt and context design, evals and integrations. They usually don't train models from scratch.
AI engineer vs ML engineer: the ML engineer trains and owns models; the AI engineer composes existing models into a product. Both are valuable. Hiring one when you need the other is the most common mistake I see.
LLM Engineer
Often used interchangeably with AI engineer, but usually with deeper model work: fine-tuning, RLHF and preference optimisation, distillation, quantisation and building LLM-specific evaluation. A good LLM engineer understands both the product layer and what happens inside the model.
Applied AI Engineer / Forward Deployed Engineer
Engineers who work directly with customers to deploy AI systems in their environment: integrating models with customer data, building custom workflows and agents, and feeding product learnings back to the core team. Common at AI platform companies and enterprise AI startups.
AI Product Engineer and Founding AI Engineer
At early-stage startups, titles collapse. A founding AI engineer or AI product engineer owns everything from the user interface to prompts, evals and sometimes fine-tuning. You are hiring range and judgment, not a narrow specialist.
AI Product Manager
Product managers for AI products need to understand model capabilities and limits, evaluation, data and cost trade-offs. Not an engineering role, but increasingly one of the hardest AI hires.
ML infrastructure roles
MLOps Engineer and ML Platform Engineer
MLOps and ML platform engineers build the systems that let ML teams train, deploy and monitor models reliably: feature stores, experiment tracking, model registries, CI/CD for models, orchestration and observability. At scale, this is what turns a research demo into a product.
ML Infrastructure / Training Infrastructure Engineer
Engineers who build and run large-scale distributed training: GPU clusters, schedulers, data loaders, checkpointing and fault tolerance. Deep experience with distributed systems, Kubernetes or Slurm, and frameworks like PyTorch distributed, DeepSpeed or JAX.
Inference Engineer / ML Systems Engineer
Inference engineers make models fast and cheap to serve: batching, caching, KV-cache optimisation, quantisation, speculative decoding and serving frameworks such as vLLM, TensorRT-LLM or SGLang. For any company paying a serious GPU bill, this role pays for itself.
GPU Kernel / ML Compiler Engineer
The deepest layer: engineers who write CUDA or Triton kernels, work on ML compilers and optimise models for specific hardware. Extremely scarce talent, concentrated in frontier labs, chip companies and inference startups.
Data Engineer for ML
Data engineers who specialise in ML build the pipelines behind training and evaluation data: collection, cleaning, deduplication, labeling workflows and data quality. As models commoditise, data becomes the differentiator — and so does this role.
How to navigate AI and ML titles
Titles mean different things at different companies. A "Research Engineer" at a frontier lab writes training code for the largest models; at a startup it may mean a prototyping ML engineer. A "Data Scientist" at one company ships models, at another builds dashboards. Always read the responsibilities, not the title.
Big Tech, frontier labs and startups use different ladders. Big Tech has well-defined levels and specialised roles. Frontier labs blur research and engineering. Startups hire generalists and give them broad titles. When comparing candidates or offers, map the actual scope.
Seniority shows up as scope, not tooling. A senior ML engineer chooses the problem, defines the metric and owns the outcome. A junior one implements a given model well. The technologies on a resume tell you much less than the decisions a person made.
New titles appear faster than the market agrees on them. Prompt engineer peaked and faded; evals engineer and inference engineer are rising. If a title is new, define it in your job description in one sentence.
How to choose the right AI role to hire
Before you open a role, answer three questions:
- Do you need to train or change models, or build on top of existing ones? Training points to ML engineers, applied scientists or LLM engineers. Building on top points to AI engineers.
- Is your bottleneck quality, product or cost? Quality needs modeling and evals. Product needs AI engineering. Cost and latency need inference and infrastructure.
- What will this person own in their first 90 days? If you can't describe it concretely, you are not ready to choose a title.
Then write the job description around outcomes, not buzzwords. "Reduce inference cost per request while keeping quality" will attract the right inference engineer. "Passionate about AI" will attract everyone.
FAQ: AI and ML roles
What is the difference between an AI engineer and an ML engineer?
An ML engineer builds, trains and deploys machine learning models. An AI engineer builds applications on top of existing foundation models using RAG, agents, prompting and evals. The ML engineer works inside the model; the AI engineer works around it.
What is the difference between a research scientist and an applied scientist?
A research scientist advances the state of the art and is measured on new methods and publications. An applied scientist adapts existing methods to a business problem and is measured on product impact.
What does an MLOps engineer do?
An MLOps engineer builds the platform for training, deploying, monitoring and versioning ML models, so ML teams can ship reliably and repeatedly.
What does an inference engineer do?
An inference engineer optimises how models are served in production — latency, throughput and cost — using techniques like batching, quantisation and KV-cache optimisation.
Which AI roles are hardest to hire in 2026?
In my experience: research engineers, inference and GPU kernel engineers, LLM engineers with real fine-tuning experience, and evals engineers. They are scarce, rarely looking, and often invisible to keyword search.
The takeaway
AI and ML titles are a moving target. Use the four families — research, applied ML, AI product engineering and ML infrastructure — to translate any title into what the work really is. Then hire for the outcome you need, not the title that's trending.