AI Platform Engineering as a Discipline

By 2026, AI platform engineering has emerged as a distinct discipline. It sits at the intersection of ML engineering, DevOps / SRE, and developer platform engineering. Different from pure ML (focuses on infrastructure not models) and different from pure DevOps (specific AI primitives matter).

Scope

• Serving infrastructure: vLLM, GPU servers, observability stack
• Model lifecycle: deployment, rollout, deprecation, fine-tuning ops
• Eval infrastructure: harnesses, automation, drift detection
• Cost engineering: caching, right-sizing, monitoring
• Developer experience: OpenAI-compatible APIs, prompt management, feature flags
• Multi-tenant operations: per-tenant routing, billing attribution, isolation
• Compliance: audit logging, data residency, regulatory scope

Skills

The composite skill set:

• Linux + Docker + Kubernetes: standard infrastructure
• NVIDIA GPU ops: drivers, CUDA, DCGM, troubleshooting
• Python production: FastAPI, async, packaging
• Observability stack: Prometheus, Grafana, Loki, OpenTelemetry
• vLLM / TGI / TensorRT-LLM: tuning, deployment, troubleshooting
• Vector stores: Qdrant / Weaviate / pgvector operations
• HuggingFace ecosystem: Hub, transformers, datasets, TRL
• Standard SRE: on-call, incident response, capacity planning

vs DevOps / ML

• vs ML engineer: less focus on model architecture / training research; more on infrastructure
• vs DevOps / SRE: same skills + LLM-specific tooling + GPU operations + ML lifecycle
• vs platform engineer: same scope + AI-specific extensions
• vs MLOps engineer: more focus on serving / inference; less on training pipelines

Verdict

AI platform engineering is a real discipline emerging in 2026. Hire / develop people with the composite skill set; don't expect any single existing role (ML engineer, DevOps engineer, backend engineer) to fully cover it. The teams that recognise this and build for it have materially smoother AI production deployments.