Inference engineer vs ML engineer, ML platform and kernel engineer

A machine learning engineer builds, trains, evaluates and ships models for a product problem: data, features, objectives, evaluation and iteration. Interviews test modelling judgement, ML fundamentals and system design for recommendation, ranking or retrieval. An inference engineer starts where the model ends and is interviewed on memory, schedulers, kernels and cost per token. The overlap is real at small companies, where one person does both, and shrinks fast at scale.

ML platform work runs the infrastructure models are built and deployed on: training pipelines, clusters and job scheduling, feature and data systems, model registries, deployment and monitoring. Interviews look like infrastructure and distributed systems interviews with an ML flavour. Inference engineering is the performance-critical slice of serving inside that platform, and it goes deeper into the GPU and the engine than a platform role usually does.

A research engineer turns research ideas into working, scalable code, mostly for training and experiments. Interviews emphasize implementing papers, debugging training and ML depth. Inference research engineers exist, working on speculative decoding methods, quantization algorithms or new attention variants, and their interviews blend both roles.

Kernel engineers write and tune the CUDA, Triton or CUTLASS code underneath frameworks, often for a specific hardware generation, and are interviewed on the memory hierarchy, tensor-core programming and profiler evidence. Inference-focused kernel work is one of the main entry points into inference engineering, and the boundary between the two titles is thin: the kernel engineer goes deeper on one layer, the inference engineer spans the stack.

Read the nouns. Serving engines, latency and throughput targets, KV cache, batching, quantization and kernels mean inference engineering. Features, experiments and model quality mean ML engineering. Pipelines, clusters, registries and deployment mean ML platform. When a posting lists all of them, ask the recruiter which two the team spends most of its time on, because that decides the loop.