Start typing to search this publication.
techcirkle
Open menu

Subscribe to techcirkle

Get new posts delivered straight to your inbox.

The Four AI Developer Archetypes (And Which One Your Team Needs)

One job title hides four different engineers. Pick the wrong archetype and your AI hire fails for reasons that have nothing to do with their ability.

techcirkle
5 min read
Cover image for The Four AI Developer Archetypes (And Which One Your Team Needs)
Engineer working on AI systems across multiple screens
Engineer working on AI systems across multiple screens

Open any job board and you will find a hundred adverts for an AI developer, and no two of them describe the same job. This is not sloppiness so much as a genuine taxonomy problem: the field moved fast enough that the vocabulary never settled.

The practical consequence is that companies hire mismatched profiles. The person is capable, the requisition was funded, and six months later the engagement is quietly not working — because the archetype was wrong for the stage the company was at. Here are the four that actually exist.

1. The AI product engineer

A strong application developer who happens to have shipped AI features end to end.

Their day is spent on streaming responses, retry semantics, caching layers, structured output handling, latency budgets and graceful degradation when the model call fails. The model is one component among many, and mostly they are thinking about everything around it.

When you need them: almost always first. If your roadmap is a list of product features that happen to use models, this is the hire.

Why they are hard to find: the profile demands genuine full-stack ability plus model fluency, and most candidates have one or the other. People with a research background often find the work insufficiently interesting; strong application engineers sometimes lack the instinct for how probabilistic components misbehave.

2. The applied ML engineer

Trains, tunes and serves models — frequently classical ones rather than deep learning.

Their problems are genuinely predictive: churn scoring, demand forecasting, fraud detection, recommendation ranking. The work involves feature engineering, training pipelines, drift monitoring and serving infrastructure.

When you need them: when the problem is a prediction over your own data, not a language task. A gradient-boosted model on well-prepared features still beats an LLM on most tabular prediction problems, at a fraction of the inference cost.

Common mis-hire: bringing one in to build a document assistant. They can do it, they will be bored, and you have paid a premium for skills the job does not use.

3. The AI platform engineer

Owns the shared layer beneath the features: model gateways, evaluation pipelines, vector stores, prompt versioning, cost attribution, GPU scheduling if you have your own.

When you need them: not for your first feature. Badly by your fifth, when four teams are each paying for their own inference through their own integration and nobody can explain the bill or upgrade a model without breaking someone.

The signal you are late: two teams have independently built retrieval pipelines, and neither can tell you whether theirs retrieves the right documents.

Interview panel evaluating a candidate
Interview panel evaluating a candidate

4. The agentic systems engineer

Designs multi-step workflows where models call tools, take actions and have to recover when reality disagrees with them.

This is the newest archetype and the worst served by conventional interviews, because the core skill is failure design rather than construction. The interesting question is never how to make step one work. It is what happens on step seven when an API returns something the model did not anticipate, whether the retry is idempotent, what compensates for a side effect that already fired, and where a human gets pulled in.

When you need them: when the system takes actions rather than producing text. The moment a model can write to a database, send a message or spend money, you are in distributed systems territory.

Where to find them: often not from AI backgrounds at all. Engineers with real distributed systems or payments experience adapt to this quickly, because the failure vocabulary is already theirs.

Matching archetype to stage

A rough progression that holds for most product companies:

  • First AI feature: one AI product engineer.

  • Genuinely predictive problem: add an applied ML engineer, or hire one instead if that is the whole problem.

  • Third or fourth feature: add an AI platform engineer before the duplication becomes structural.

  • System takes actions: agentic systems experience becomes non-negotiable.

The failure mode to avoid is writing a requisition that lists all four. It reads as thorough and it attracts generalists who are shallow in every direction. If you genuinely need coverage across several archetypes, that is an argument for a small team or an external partner, not for finding one person who can do everything.

The screening question that cuts across all four

Whichever archetype you are hiring, ask this: what did your last AI feature cost to run per thousand requests, and what did you do to bring it down?

In conventional software the marginal cost of a request rounds to zero, so engineers are trained not to think about it. In AI systems it is real money, set largely by decisions made in the first fortnight, and it scales with success. Routing everything to a frontier model is the default that quietly destroys margins; classify cheaply, cache aggressively, handle the easy majority with something small, and escalate only where it matters, and the same output costs a fraction as much.

Engineers who have operated a system answer with numbers and a list of levers. Everyone else answers with a model recommendation.

The full guide covers rate benchmarks by market, a take-home format that predicts performance, and the in-house versus partner decision: Hire AI Developer in 2026. For teams whose next step is the fourth archetype, our agentic workflow development page covers how we approach that work.

Frequently Asked Questions

Which archetype should a startup hire first?

An AI product engineer, in almost every case. They cover the widest range of early needs, and until you have several features in production the platform and agentic specialisms have nothing to work on.

Can one person cover two archetypes?

Product engineer plus agentic systems overlap substantially and are often the same person. Platform and applied ML are genuinely different disciplines, and hiring someone who claims all four usually means finding someone shallow in each.

Do I need an ML engineer if I am only using foundation models?

Generally no. Foundation model application work is systems engineering with a statistics component. Bring in an ML engineer when you have a predictive problem over your own data that no off-the-shelf model addresses.

How do I know when to hire a platform engineer?

When two or more teams have independently solved the same infrastructure problem, when nobody can attribute inference spend to a feature, or when upgrading a model requires coordinating with several teams. Any one of those is the signal.

Is agentic systems experience worth paying a premium for?

Only if your system takes actions. If it produces text for a human to act on, ordinary product engineering plus good failure handling is sufficient, and you will overpay for a specialism you do not use.

What if a strong candidate does not fit any archetype cleanly?

That is common and often fine. The archetypes are a tool for writing a focused requisition, not a taxonomy to enforce during interviews. Judge the candidate against the actual work; use the archetypes to make sure you defined that work narrowly enough.

Subscribe to techcirkle