GlossaryRoles
Roles
Who actually does this work?
Titles keep multiplying, many are less than two years old, and one advert can call three different things by the same name. Each role is placed here by the object it answers for and by how close it stands to the machine, along with what the work actually fills a week with.
No page carries a salary or a range: those figures change from one country and one city to the next, and a number published here would be wrong for most of the people reading it.
15 roles published
The map
Each column is a floor of the site, each row a depth.| Depth | 1The Model | 2The Harness | 3The Agent | 4The Organisation |
|---|---|---|---|---|
| P4Writes the model | Research scientistResearch engineerML engineer | |||
| P3Writes the system | LLMOps engineer | |||
| P2Writes the application | AI engineerContext engineerEvaluation engineer | Forward deployed engineerAgent engineer | Model validator | |
| P1Writes the brief, reads the code | Human data specialist | Prompt engineer | AI red teamer | AI compliance lead |
| P0Does not touch code | AI governance lead |
Depth is a distance, not a rank: a human data specialist at P1 helps build the model, and a governance lead at P0 decides everything.
Floor 1 · The Model
4 roles published- Research scientistThe one chasing an answer nobody has yet
A result others can reproduce, and the question that made it possible. If they leave, a line of research stops, not a service.
- Research
- Build
- Operate
- Verify
- Lead
- Research engineerThe one who makes the experiment hold at scale
An experiment that runs, replays, and gives the same number twice. If they leave, research slows by a factor measured in months.
- Research
- Build
- Operate
- Verify
- Lead
- ML engineerThe one who trains a model and keeps it in service
A model trained on real data, served to users, with its drift watched. If they leave, a service degrades without anything stopping.
- Research
- Build
- Operate
- Verify
- Lead
- Human data specialistThe expert whose judgement becomes the model’s material
A body of preferences and reference answers, in one precise field. If they leave, the model lacks expert judgement, and nobody sees it before it is in use.
- Research
- Build
- Operate
- Verify
- Lead
Floor 2 · The Harness
5 roles published- LLMOps engineerThe one who keeps in production what never answers the same way twice
A platform where other people deploy without breaking things: instruction versions, quotas, traces, attributed cost. If they leave, nobody knows what is running, or what it costs.
- Research
- Build
- Operate
- Verify
- Lead
- AI engineerThe one who puts a model they did not train to work
A feature that answers real users, with its cost per call and its quality measurement. If they leave, the product stops moving.
- Research
- Build
- Operate
- Verify
- Lead
- Context engineerThe one who decides everything that enters the window
The chain that assembles a window: what gets retrieved, what gets cut, what carries over from one conversation to the next. If they leave, answers degrade without anything breaking.
- Research
- Build
- Operate
- Verify
- Lead
- Evaluation engineerThe one who writes the mark scheme before reading the paper
A set of tests and their mark scheme, plus the number that does or does not clear a release. If they leave, nobody can say whether the last change improved anything at all.
- Research
- Build
- Operate
- Verify
- Lead
- Prompt engineerThe title that went out fabricating three others
None, and that is the cause of its disappearance. The role produced wordings nobody received as a deliverable: not a system, not a verdict, not a decision. A trade with no deliverable gets shared out.
- Research
- Build
- Operate
- Verify
- Lead
Floor 3 · The Agent
3 roles published- Forward deployed engineerThe one who goes and writes the code at the customer’s
A system running at a named customer, and the pieces of that system that make their way back into the standard product. If they leave, the integration stops.
- Research
- Build
- Operate
- Verify
- Lead
- Agent engineerThe one who makes a self-deciding loop hold together
A loop that carries a task through unsupervised, with its trace, its stopping points and what it is allowed to do. If they leave, nobody can say why it stops.
- Research
- Build
- Operate
- Verify
- Lead
- AI red teamerThe one who derails the system before anyone else does
A catalogue of reproducible failures, with what triggers each one. If they leave, those failures are only discovered in public.
- Research
- Build
- Operate
- Verify
- Lead
Floor 4 · The Organisation
3 roles published- Model validatorThe one entitled to say no before it goes live
A reasoned opinion that holds, clearing or refusing a release. If they leave, the decision chain loses its stopping point.
- Research
- Build
- Operate
- Verify
- Lead
- AI compliance leadThe one who proves to an outsider that the rule is being kept
A file that holds in front of a third party: classification of uses, evidence, a log of decisions. If they leave, the organisation can no longer demonstrate what it claims.
- Research
- Build
- Operate
- Verify
- Lead
- AI governance leadThe one who decides who may deploy what
A written framework, known and applied: what is permitted, by whom, on which data. If they leave, everyone decides alone, and nobody knows any more who decided what.
- Research
- Build
- Operate
- Verify
- Lead