If your product calls a foundation model through an API, hire an LLM engineer first. If it needs a model trained on your own data, like fraud scoring, demand forecasting or image detection, hire an ML-focused AI engineer first. Most startups in 2026 are in the first group and hire as if they were in the second, then wonder why their expensive model specialist spends the week writing glue code.
One honest caveat before the rule: there is no official definition of either title. The US Bureau of Labor Statistics has no occupation code for an LLM engineer or an AI engineer, and no standards body defines them. What follows is our working definition, built from what we see when we match these roles.
Key Takeaways
- In our working definition, an LLM engineer builds products on top of foundation models: prompts, retrieval, evaluation, cost and latency.
- An AI engineer, as the title is used, often means someone who trains and ships models, though many job posts use it to mean an LLM engineer.
- Decision rule: API-based product, hire the LLM engineer. Custom model on proprietary data, hire the ML-focused AI engineer.
- Screen on evaluation. The best signal for either role is how they measure whether the system got worse.
LLM Engineer vs AI Engineer: Our Working Definition
| LLM engineer | ML-focused AI engineer | |
|---|---|---|
| Core job | Ships features on model APIs | Trains, tunes and serves models |
| Daily work | Prompts, retrieval, test sets, guardrails, latency and token cost | Data prep, training runs, model metrics, serving |
| Background | Strong backend engineer who moved into AI | Data science or ML, often heavier on math |
| First 90 days look like | A feature live with users and an eval suite | A baseline model and a pipeline to improve it |
| Hire first when | Your product calls a model API | Your edge is a model nobody can buy |
The overlap is real. A good LLM engineer understands embeddings and model behavior. A good ML engineer can wire an API. But the instincts differ: one optimizes a system around a model they do not control, the other optimizes the model itself.
Interview Questions That Separate Them
Ask both candidates the same three questions and listen for where they go.
1. "Our AI answers got worse after a prompt change. How do you find out, and how fast?" A strong LLM engineer describes a fixed test set, scoring and a release gate. A pure researcher talks about retraining.
2. "Our model bill doubled. What do you check?" Look for token counts per request, caching, routing simple requests to a cheaper model.
3. "When would you train our own model instead of calling an API?" Both should have an answer. The LLM engineer should still say "not yet" for most products.
For a deeper interview kit on retrieval-heavy work, see how to hire RAG developers, and for agent-style products, how to hire an AI agent developer.
Red Flags When You Hire LLM Engineers
- Demos without numbers. "It works great" with no test set is a demo, not a feature.
- No opinion on cost. At current list prices a cheap model is a fraction of a big one per token, and a senior LLM engineer should know when each is enough.
- Fine-tuning as the first answer to every problem.
- A resume full of model training and no shipped product, for a role that is mostly product work.
A Concrete Version
A 15-person legal-tech startup wants to answer customer questions from its own document library. The founder's first instinct is an ML engineer to "build our model."
Instead it hires one senior LLM engineer. Weeks 1 to 4: retrieval over the library, a test set of 200 real customer questions with approved answers. Weeks 5 to 8: answers scored on every release, cost per answer tracked.
The cost math stays small. At Anthropic's published prices as of October 2026, Sonnet 5.5 costs $2 per million input tokens and $10 per million output. At 10,000 questions a month, each with 4,000 input tokens of retrieved context and 400 output tokens, that is 40 million input tokens ($80) plus 4 million output tokens ($40): about $120 a month.
Six months in, the team has a measurable ceiling: 8% of questions need a clause-classification step the API handles poorly. Now there is evidence for an ML-focused hire, scoped to one problem.
The Honest Counterpoint
Some products really do need the model specialist first. If your core value is a prediction nobody can buy, like risk scoring on your own transaction history or vision on your own factory images, an LLM engineer will hack around the gap with prompts that never get good enough. Hire the ML engineer.
Also, titles in the market are a mess. Plenty of excellent engineers called "AI engineer" on their resume are exactly the LLM engineer described here. Read the work, not the title. See how to hire an AI engineer for that profile.
Frequently Asked Questions
What does an LLM engineer do?
In our working definition, they build product features on foundation models: prompts, retrieval, evaluation, guardrails, latency and cost. There is no official job definition.
Should a startup hire an LLM engineer or an ML engineer first?
If the product calls a model API, the LLM engineer. If the product's edge is a custom model on proprietary data, the ML engineer.
Where can I hire LLM engineers quickly?
Senior engineers who have shipped on model APIs are on nearshore benches in Latin America, working US hours. Through staff augmentation they often cost 40 to 60% below a fully loaded US hire.
The Bottom Line
Hire for the product you have, not the one in the pitch deck. To hire LLM engineers this month, see AI engineers available in LATAM or describe the role to Sol for a vetted shortlist within 72 hours. Here is how Sol works.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
