Staff Engineer, Merlin

• You want to build a model that understands what happens next. Not a classifier, not a chatbot — a learned model of an aircraft moving through an environment, good enough that a planner can imagine several futures and pick one
• You are as interested in where your model is wrong as in where it is right, because in this application the second question is the one that matters
• You'll own the learned predictive models of the aircraft, its environment and other actors in it — and the rollout machinery that turns those models into evaluated candidate futures
• Degree in Computer Science, Artificial Intelligence, Data Science, Computer Engineering, Applied Math, or a related subject
• 10+ years building and training deep learning models, with meaningful work on sequence, dynamics, video or trajectory prediction
• Strong PyTorch and a solid grasp of the training stack — distributed training, mixed precision, experiment tracking, debugging a run that has silently gone wrong
• Working knowledge of dynamical systems, state estimation or control; you can read a flight dynamics model and know what your network is and is not replacing
• Demonstrated rigor in evaluation and uncertainty quantification
• Comfort with messy real-world sensor and telemetry data

Desirable

• You want to build a model that understands what happens next. Not a classifier, not a chatbot — a learned model of an aircraft moving through an environment, good enough that a planner can imagine several futures and pick one
• You are as interested in where your model is wrong as in where it is right, because in this application the second question is the one that matters
• You'll own the learned predictive models of the aircraft, its environment and other actors in it — and the rollout machinery that turns those models into evaluated candidate futures
• Design, train and evaluate world models that predict the evolution of aircraft state, environment and other traffic under candidate action sequences
• Build the rollout and imagination machinery that lets a planner evaluate candidate plans against predicted outcomes before committing to one
• Own uncertainty: calibrated predictive uncertainty and out-of-distribution signals are required outputs of your models, consumed directly by the safety monitoring layer
• Characterize model validity boundaries rigorously — the flight regimes, weather, traffic densities and configurations where predictions can be trusted, and where they degrade
• Work the sim-to-real gap in both directions with the Data/Sim team: train in simulation, validate against flight data, and feed discrepancies back into simulator fidelity
• Benchmark learned dynamics against Merlin's existing physics-based models and flight-controls models; be honest about where the classical approach wins
• Contribute to the shared eval harness and hold the line on reproducibility: every result traceable to a data version, a config and a seed