Anduril Industries is a defense technology company with a mission to transform U.S. and allied military capabilities with advanced technology. By bringing the expertise, technology, and business model of the 21st century’s most innovative companies to the defense industry, Anduril is changing how military systems are designed, built and sold. Anduril’s family of systems is powered by Lattice OS, an AI-powered operating system that turns thousands of data streams into a realtime, 3D command and control center. As the world enters an era of strategic competition, Anduril is committed to bringing cutting-edge autonomy, AI, computer vision, sensor fusion, and networking technology to the military in months, not years. ABOUT THE TEAM Maritime Digital Production (MDP) is the software and digital systems function within Anduril's Heavy Metal division. We build and deploy the full technology stack that powers Anduril's shipbuilding factories: the data infrastructure that makes every machine and sensor visible in real time, the manufacturing execution system (ArsenalOS) that workers and planners use every shift, the scheduling engine that replans production in minutes instead of days, and the AI systems that eliminate manual toil from both the shop floor and business operations. MDP operates at the boundary between Operational Technology and Information Technology. Our systems live where factory-floor machines, edge compute, and OT networks meet enterprise platforms and cloud infrastructure. We incubate solutions close to the production line, validate them with real operators building real hardware, harden them for reliability and security, and then scale them across multiple sites. The environment is fast, physical, and consequential. When our systems go down, production stops. The output of our work is not a dashboard: it is a ship. This is not a support function. It is a strategic investment by Anduril in the premise that digitizing the manufacturing lifecycle end-to-end, from engineering definition through scheduling through execution through field feedback, is how Heavy Metal will out-build, out-adapt, and out-scale the traditional defense industrial base. MDP is scaling from a founding team to 70+ engineers across multiple U.S. sites. You will be joining early, working on hard problems with real operational stakes, and shaping how manufacturing software is built at Anduril from the ground up. ABOUT THE JOB As a Senior Numerical Optimization Engineer on this unique initiative, you will own the scheduling optimization core that decides when and how ships get built. This project represents a shipyard's work as a graph of 100,000+ manufacturing operations carrying precedence, spatial, resource, and qualification constraints, then continuously replans as the factory changes. You will formulate that problem, build and benchmark the solvers that attack it, and keep them fast and trustworthy enough that a new plan can reach the floor within minutes of a disruption. This is a research-grade problem with production consequences. Resource-constrained project scheduling at this scale is NP-hard, there is no ground truth to measure against, and a schedule has to be stable enough that the shop floor is not thrashed every time it changes. You will have the latitude to develop novel approaches and the obligation to ship them. WHAT YOU'LL DO Formulate shipyard production scheduling as a mathematical optimization problem covering precedence with lag, disjunctive spatial exclusion, multi-resource capacity with qualification matching, calendars and shifts, and material availability. Own the solver-agnostic scheduling interface, integrate commercial and open-source solvers behind it, and benchmark them against each other on makespan, stability, resource utilization, and solve time. Build the two-tier replan path: fast local repair that returns a feasible schedule in seconds for a bounded affected subgraph, and background global re-solve that runs for minutes and swaps in when it beats the incumbent plan. Develop the stability objective that keeps a replan from needlessly moving work the factory has already staged, and the policy that decides when a better plan is worth the churn. Establish what "good enough" means when there is no ground truth: LP relaxation lower bounds, best-of-N ensemble upper bounds, quality ratios, and solution-quality regression tests that run on every change. Research and evaluate approaches beyond the V1 solver, including decomposition, metaheuristics, rolling-horizon methods, and non-traditional hardware (wafer-scale compute, quantum annealing) where they earn their place. Turn disruption events into replans: determine blast radius across the operation graph, scope the re-solve, merge the result into the live schedule, and resolve boundary conflicts. Build adaptive planning templates that generate recovery work (E.g. a damaged part that must be removed, replaced, and reinspected) by composing geometric qu