Senior AI Engineer
Pubblicato il 07-09-2026 - Principled Intelligence in La Spezia
Fully remote, with an office in the center of Rome you are welcome in as often or as rarely as suits you
We are hiring a Senior AI Engineer to train the next generation of our models, including the small language models that carry most of the load in production. Anyone who has taken a general-purpose model into production has reached the point where the remaining improvements stop coming from prompts and context and start coming from training: from data curation to continual pre-training, post-training with reinforcement learning, model merging, and model evaluation. This role owns that work end to end. We firmly believe the next move in the state of the art will not necessarily come from whoever has the most compute, and we would like to be one of the places it comes from instead.
What you'll do
- Train, fine-tune and distil new model families, from dataset construction through to the evaluations that decide whether a run ships.
- Build the training and evaluation infrastructure that makes results reproducible,
including harnesses that catch regressions in behaviour a benchmark score will not surface.
- Curate and generate training data, and decide what a given capability actually requires: more data, better data, a different teacher model, or a change in objective.
- Take models the last distance into production, working with the engineers who serve them on quantisation, latency and the cost per answer at volume.
- Choose which questions are worth pursuing, since several of the ones in front of us have no settled answer in the literature yet.
Minimum qualifications
- Five or more years of engineering experience, with at least one spent training or fine-tuning language models.
- Fluency in PyTorch and experience with distributed training across multiple GPUs or nodes.
- Practical experience with post-training methods, including supervised fine-tuning and at least one preference optimisation approach such as DPO.
- Experience designing evaluations for your own mode