AI engineering level is not the number of frameworks someone can name. Assess the scope, autonomy, complexity, and impact of work they can repeatedly deliver. Junior engineers implement bounded components with guidance; mid-level engineers own a production feature; senior engineers design and operate systems across teams; staff engineers set reusable technical direction and improve organizational capability.
Use this matrix for growth and evidence—not as a universal job-title standard.
Level definitions
| Level | Scope | Autonomy | Typical evidence |
|---|---|---|---|
| Junior | Component or test fixture | Works from clear design with review | Correct implementation, tests, documentation |
| Mid-level | Production feature/workflow | Owns delivery within team boundaries | Deployed feature with evals and operations |
| Senior | Multi-component system | Resolves ambiguity and leads design | Reliable architecture and cross-team outcomes |
| Staff | Multi-team technical domain | Sets direction and raises leverage | Standards, platforms, risk decisions, capability growth |
SFIA 9 similarly separates professional skill from increasing responsibility through autonomy, influence, complexity, knowledge, and business behavior. Its published ML profile progresses from assisted tasks through independent production work, end-to-end architecture, standards, and organizational strategy. This JoinAI matrix adapts that concept to foundation-model applications.
Competency matrix
| Domain | Junior | Mid-level | Senior | Staff |
|---|---|---|---|---|
| Software engineering | Implements typed APIs and tests | Owns service, CI, errors, and deployment | Designs boundaries, performance, and migration | Establishes patterns across teams |
| Data and retrieval | Builds ingestion/chunking fixtures | Owns provenance, retrieval, and access tests | Designs lifecycle, ranking, and multi-tenant controls | Sets data/retrieval platform direction |
| Models and prompting | Uses APIs with versioned settings | Selects routes from measured quality/cost | Designs routing, fallback, and model-change policy | Defines model portfolio and adoption gates |
| Agent/tool systems | Implements one strict tool | Owns bounded agent and side-effect contracts | Designs state, recovery, approval, and orchestration | Sets autonomy/agent architecture standards |
| Evaluation | Adds deterministic cases | Owns dataset, scorers, and release gate | Calibrates slices, judges, and online loop | Defines evaluation governance and investment |
| Security/privacy | Follows handling and secret rules | Threat-models feature boundaries | Designs identity, policy, isolation, incident response | Sets risk architecture with security/legal leaders |
| Operations | Adds trace fields and alerts | Owns objectives, dashboards, fallback, runbook | Designs reliability/capacity and leads incidents | Improves operational system across teams |
| Product judgment | Understands task acceptance | Connects technical metrics to user outcome | Shapes scope and rejects weak AI use cases | Directs portfolio tradeoffs and platform leverage |
| Communication/leadership | Documents work and asks early | Coordinates feature stakeholders | Leads reviews and mentors engineers | Influences organization and grows senior leaders |
Not every role needs equal depth in every domain. A retrieval specialist and an agent-platform engineer can both be senior with different shapes. The level comes from sustained responsibility and impact, not filling every cell.
Evidence rubric
Score each claim from 0 to 3:
0 — no evidence yet1 — completed once with close guidance2 — completed independently more than once3 — improved the practice for other engineers or teams
For every score, link an artifact:
- design record with alternatives and constraints;
- pull request or code sample with tests;
- evaluation dataset and release result;
- production trace/dashboard with privacy controls;
- incident analysis and permanent regression;
- migration, reliability, cost, or latency outcome;
- mentoring plan, reusable standard, or adoption evidence.
Certificates and course completion can support knowledge claims. They do not independently prove production responsibility.
Promotion is a change in scope
A promotion case should answer:
- What larger or more ambiguous scope does the engineer now own?
- Which decisions do they make independently and correctly?
- What production outcomes have they sustained?
- How do they handle failure, risk, and tradeoffs?
- Whose effectiveness improves because of their work?
Avoid promoting someone for doing more tickets at the same responsibility level. Conversely, do not require staff candidates to be the best individual coder in every domain; staff impact commonly arrives through architecture, standards, alignment, and leverage.
Self-assessment worksheet
| Domain | Current evidence | Score | Next-level gap | 30-day practice | Reviewer |
|---|---|---|---|---|---|
| Software | |||||
| Retrieval/data | |||||
| Models/agents | |||||
| Evaluation | |||||
| Security | |||||
| Operations | |||||
| Product/leadership |
Choose one technical gap and one scope/communication gap per cycle. “Learn Kubernetes” is not a useful next step; “deploy the evaluation service with rollback, capacity objectives, and an on-call handoff” is observable.
Example growth projects
Junior → Mid: own a small RAG or tool-calling feature from task contract through deployment, with 30 reviewed cases, strict tool/data boundaries, traces, fallback, and a support note.
Mid → Senior: redesign a failing workflow across retrieval, agent state, evaluation, and operations; lead the migration; demonstrate improved quality and cost per successful task; encode incidents as regressions.
Senior → Staff: establish shared evaluation, tool, or observability contracts across several teams; drive adoption; define migration and exception policy; measure whether delivery and incident outcomes improve.
Google Cloud’s current Professional ML Engineer scope includes building, deploying, operationalizing, automating, scaling, and monitoring traditional and generative AI solutions. SFIA’s AI/ML role guidance also combines ML, data engineering, software development, testing, integration, deployment, ethics, and operations. Both reinforce that the job is a system discipline, not prompting alone.
Start with what an AI engineer actually does, then use the AI engineer roadmap to turn the two highest gaps into a learning plan. Build evidence with the portfolio project rubric instead of listing unverified skills.
If your current role is data science, use the data scientist to AI engineer transition roadmap to turn the two lowest required domains into one end-to-end bridge project.
