AI Platform Engineer
Pubblicato il 10-10-2026 - Reply in Torino
Are you an AI Platform Engineer expert in the platform that turns a researcher's idea into a running job on a cluster ?
Join Reply
WHO WE ARE
Reply is a network of highly specialized companies which support leading industrial groups in defining and developing business models using new technology and communication paradigms, such as Big Data, Cloud Computing, Digital Communication, Internet of Things, Mobile and Social Networking. Reply focuses on Consultancy, System Integration and Application Management, covering three areas of competence: Processes, Applications and Technologies.
WHAT YOU WILL FIND
- Core activities . You will keep the core of the Model Factory platform standing, and you will make it grow. You will orchestrate GPU workloads and distributed training that hold up under load, build versioning and governance across data, models and evaluation runs so results are still reproducible months later, and shorten the distance between a researcher's idea and a running job on the cluster. Every model we train,
distil or optimise — for our own portfolio and for client engagements — runs on what you build.
- Tech & tools stack. You will work SkyPilot, Dagster, NVIDIA NeMo, Kubernetes, MLflow and Unity Catalog, the core of the platform. Around them: the NVIDIA stack (CUDA, NCCL), job scheduling across national (Italian) and European HPC clusters and cloud (such as Scaleway or Nebius), infrastructure as code and GitOps CI/CD, vLLM for inference serving, and Prometheus and Grafana observability that tells us what a run cost and why it failed.
- Mentoring, collaboration, and continuous growth. You will work alongside the AI research engineers who train the models, the model design engineers who specify them, the project managers who commit the dates, and cloud and security specialists — plus client platform teams when the environment lands on their side. Your first customers are internal: if the platform is slow, opaque or fragile, they feel it long before any clien