Senior Manager, AI Foundation Model

Merlin Labs•Boston, MA

18h•Onsite

About The Position

Merlin is seeking a Senior Manager, AI Foundation Model to own the company's foundation and world-model work. This role involves technical strategy, team leadership, system interface design, evaluation definition, safety-relevant output establishment, benchmarking, certification partnership, and research judgment. The ideal candidate has built learned decision-making systems that have transitioned from the lab to real hardware, possesses fluency in modern model architectures and post-training techniques, and understands the critical difference between a model that performs well and one whose behavior can be characterized for safety. This position is crucial for advancing Merlin's autonomy platform and ensuring the safety and defensibility of AI components in regulated aviation environments.

Requirements

Degree in Computer Science, Artificial Intelligence, Data Science, Computer Engineering, Applied Math, or a related subject.
8+ years building AI systems, with 3+ years leading technical teams or owning a major model program.
Proven team management experience shipping high-tech, AI-powered models into production — hiring and developing AI engineers, setting technical direction and priorities, and owning delivery from research through deployment.
Demonstrated ownership of a learned system that shipped into a physical, real-time product — robotics, autonomous vehicles, aerospace, or industrial autonomy.
Depth in at least two of: world models and learned dynamics; sequence models applied to planning or control; post-training (SFT, preference optimization, RL fine-tuning); structured or constrained generation.
Rigorous evaluation practice: you have built eval harnesses that caught regressions before customers did, and you can explain why a model's aggregate metric improved while a specific behavior got worse.
Strong PyTorch; comfortable reading and reasoning about the C++ real-time systems your models feed.
You write clearly. Architecture decisions here get read by systems engineers, safety engineers, and regulators — not only by other AI engineers.
Applicants must be authorized to work in the United States without the need for visa sponsorship now or in the future.

Nice To Haves

Experience with learned components in a certified or regulated product (DO-178C, ISO 26262, IEC 62304).
Background in classical planning, behavior trees, MCTS, or hierarchical task networks — you'll be replacing and interoperating with exactly these.
Familiarity with aviation domain structure: flight phases, ARINC 424 procedures, ATC phraseology.
Publications or open-source contributions in embodied AI, world models, or robot learning.

Responsibilities

Technical strategy: own Merlin's foundation and world-model work — architecture selection, build-vs-adapt decisions, post-training approach, and the capability roadmap that supports it.
Team leadership: lead and mentor a small team of world-model and post-training engineers; set the technical bar and the review culture for model work across AI Core.
System interface: design the model interface to the rest of the autonomy stack — structured, schema-constrained plan outputs that a deterministic verifier can accept or reject, never free-form actuator authority.
Evaluation: define what “good” means before training begins — build the evaluation harness, capability taxonomy, and regression suite that gate every model release, in partnership with the Data/Sim/Release pillar.
Safety-relevant outputs: establish uncertainty quantification and out-of-distribution detection as first-class model outputs, not afterthoughts — downstream safety monitoring depends on them.
Benchmarking: deliver an honest, reproducible comparison between learned planning and Merlin's current rule-based behavior planning across representative mission profiles, including the cases where the learned approach loses.
Certification partnership: work with Systems Engineering, Certification, and the Chief Architect to keep model design inside what is defensible to a regulator, and to shape what “defensible” will mean for learned components.
Research judgment: track the external research frontier and make disciplined calls about what Merlin adopts, builds, or ignores.