Research Engineer, Mid-Training

About the role

We are an applied AI lab building end‑to‑end software agents, creator of Devin – the first AI software engineer. The Mid‑Training Research Engineer will focus on the critical stage between pre‑training and post‑training, shaping the model’s ability to reason, generalize and support downstream tasks.

Key responsibilities

• Design and iterate high‑quality data mixtures for late‑stage and annealing training runs, including sourcing, filtering and weighting strategies.
• Drive capability injection in coding, mathematics and long‑horizon reasoning through curated data and training interventions.
• Develop and evaluate synthetic data pipelines that generate scalable training signals, while understanding their limits.
• Research and optimise multi‑stage learning‑rate schedules, warm‑up strategies and compute allocation across training phases.
• Extend effective context length without harming short‑context performance via positional‑encoding and data‑construction techniques.
• Build robust evaluation suites to distinguish genuine capability gains from benchmark overfitting and close the feedback loop.
• Measure how mid‑training interventions scale with compute and data, proposing new methodologies for large‑scale deployment.

Required profile

• Strong research mindset combined with solid engineering execution on large‑scale model training.
• Experience designing data pipelines and training schedules for deep‑learning models.
• Ability to translate research insights into measurable performance improvements.
• Analytical skills to evaluate synthetic data approaches and scaling behaviours.
• Collaborative attitude to work across pre‑training and post‑training teams.

Required skills