24. Career Decision Matrix
Choose the ownership boundary first, then compare titles. In a small AI-augmented team, a credible profile combines one deep specialty with adjacent product, verification, and agent-supervision skills.
| Current strength | Closest role family | Evidence to build |
|---|---|---|
| Product or backend engineering | AI Engineer or Applied AI Engineer | A shipped workflow with evals, observability, failure analysis, and cost data |
| Model training and experimentation | ML Engineer or LLM Engineer | Reproducible experiments, model comparisons, and deployment constraints |
| Cloud, SRE, or internal platforms | AI Platform Engineer or MLOps Engineer | A reusable deployment path with identity, tracing, rollback, and cost controls |
| Application or cloud security | AI Security Engineer | An AI threat model, adversarial tests, mitigations, and retest evidence |
| GRC, data governance, or control automation | AI Governance Engineer | A working inventory, approval gate, lineage path, or policy-as-code control |
| Statistics, QA, or experiment design | AI Evaluation Engineer | A versioned dataset, validated graders, regression report, and release gate |
| Customer-facing technical delivery | Forward-Deployed Engineer | Production integration, adoption result, handoff, and reusable field feedback |
| Product strategy and user research | AI Product Manager | Acceptance criteria for variable behavior and explicit quality, cost, latency, and fallback decisions |
| IAM or platform security | AI Security or AI Platform Engineer with agent-identity depth | Delegated authorization, session attribution, least privilege, and revocation design |
| Technical writing or requirements analysis | Product, QA, or Applied AI role with specification depth | A versioned specification linked to tests and implementation evidence |
Portfolio proof
- Build a real workflow against a model API or open model.
- Define the task, failure modes, and acceptance criteria before polishing the demo.
- Add a versioned evaluation set and report false positives, false negatives, and uncovered cases.
- Instrument latency, cost, errors, tool actions, and fallback behavior.
- Document one incident or failed approach and the evidence that justified the correction.
These artifacts demonstrate production judgment. They do not substitute for domain, research, security, or regulatory depth where the role requires it.