Data Scientist, Hadrian

• Forecasting and prediction on real, messy manufacturing data, with honest uncertainty
• Representation learning and embeddings; similarity and retrieval; transfer/few-shot for sparse data
• Deep learning that ships (PyTorch), and the judgment to know when not to use it
• Strong classical ML and statistics (GBMs, Bayesian/hierarchical, survival, causal)
• Validation done right: backtesting, leakage control (time and part-family), calibration
• Python; turns a messy process into features and a model into a decision an operator or a downstream system can consume
• Deploys and monitors models; thinks about pipelines and drift from the start, not after
• Works with limited, high-value data and knows how to borrow strength

Desirable

• This is the modeling half of manufacturing data science at Hadrian
• The factory turns geometry into parts: a CAD model, a material, a set of tolerances, a route through stations
• This role predicts what that process will do before it runs, and gets better at it with every part that goes through
• Our factories generate rich process data on high-mix, low-volume aerospace parts, but most parts are near-unique, so the classic "lots of history per SKU" playbook doesn't apply
• The leverage is representation: embed a part by its geometry, material, tolerances, and route, then predict cycle time, cost, tool wear, quality, and triage risk from the parts like it, before the first chip is cut
• The work spans forecasting and prediction (cycle time, tool life, quality and yield, demand, queue and lead time, always with calibrated uncertainty), representation learning (part and operation embeddings so a part with no history inherits the behavior of its neighbors), and geometric modeling (features and models straight off CAD, mesh, and point cloud)
• Deep models where they earn their keep, classical where it wins
• Those predictions feed quoting, scheduling, capacity, and DFM, and you'll own the pipelines that serve them, partnering with ML Platform to deploy and Data Engineering on features
• Build and ship production models for cycle time, tool life, quality, and demand, using calibrated uncertainty (quantile, conformal, or Bayesian) rather than point estimates alone
• Model directly off geometry by engineering features and building geometric/graph models that predict cycle time, cost, DFM and tolerance risk, and triage probability
• Build a part and operation embedding layer that represents a part by geometry, material, tolerances, and route, retrieves similar parts, and transfers their behavior to cold-start new ones
• Validate honestly through backtesting that respects time ordering and part-family leakage, and make a defensible case for deep versus classical methods on each problem
• Own models end to end on the platform, including reproducible training, serving, monitoring, and retraining, in partnership with ML Platform and Data Engineering
• Close the loop in production by detecting drift and quality anomalies so predictions improve as new data lands
• Turn predictions into decisions for quoting, scheduling, capacity, and DFM; design experiments and A/B tests to measure real impact, then document and hand off to operations