AI Research Engineer
Pubblicato il 23-09-2026 - Reply in Roma
Are you an
AI Research Engineer
expert in
training custom AI ?
Join
Reply
Se sta pensando di inviare una candidatura, si ricordi di premere il pulsante apposito dopo aver letto l'intera descrizione.
WHO WE ARE
Reply specialises in the design and implementation of solutions based on new communication channels and digital media. As a network of highly specialised companies, Reply supports major industrial groups in the telecom and media; industry and services; banking and insurance and public sectors in defining and developing business models enabled by the new paradigms of AI, cloud computing, digital media and the internet of things. Reply's services include: consulting, system integration and digital services.
WHAT WILL YOU DO?
Core activities . You will design synthetic dataset generation pipelines from task descriptions, teacher-student distillation or production traces. You will build auto-research harnesses using AI coding to navigate the design space across recipes, architectural design, dataset mixes and ablations. You will run post-training workflows,
with a focus on RL and agentic reasoning, either via RL and distillation, on custom RL environments across domains for professional use on the Pareto frontier. You will schedule jobs on a distributed cluster across local compute and cloud compute and optimize serving confiturations either for online RL, evaluation or synthetic generation. You will share the core insights of your work via blog posts, open-source repositories, and on model and dataset hubs, like Hugging Face.
Tech & Tools Stack.
You will work with cloud compute clusters (Lambda, Nebius, AWS, GCP), local compute clusters (NVIDIA DGX), AI workload managers (SkyPilot), RL environments and stack (NeMo-RL, SkyRL, TRL), synthetic dataset generation pipelines (DataFlow, DataTrove), experiment trackers (MLflow), model and dataset hubs (Hugging Face Hub), evaluation metrics and environments (LM Evaluation Harness), kernels (Triton), serving (vLLM,