The Commodities tribe at Kpler runs production ML models that predict what cargo a vessel is carrying (Product Estimation) and where in-transit vessels are headed (Destination Forecast), and where they are expected to arrive (ETA)— across LNG, DRY, LPG, and LIQUIDS. These predictions feed directly into Kpler's cargo intelligence platform, consumed by market analysts, trading desks, and external customers worldwide.
You will own the science behind these models: designing and evaluating features from maritime AIS data, H3 geospatial routing distributions, transit statistics, and commodity-specific signals; running structured experiments on ML Flow based platform; and pushing the accuracy, coverage, and reliability of predictions forward.
You are not handed a Jupyter notebook and a dataset. You work in a production system with real-time inference running every 1–3 hours across 4 commodity types, and your model changes need to be validated against a running parallel baseline before they go live. The new platform is being built specifically to make the experiment loop fast enough that this level of rigour does not slow you down.
Key Responsibilities
- Own the feature engineering roadmap for ETA & Destination Forecast across all 4 commodity types — propose and implement new features as dbt models using Airflow to orchestrate the data pipelines, and validate their impact through structured experiments.
- Design and run experiments using kpler-ml framework, logging all runs from train to evaluation to MLflow and producing structured comparison reports against the production baseline before any promotion.
- Work directly with Commodities Market Analysts and product stakeholders to understand where prediction quality matters most commercially — and use that to prioritise the experiment backlog.
- Contribute to the drift monitoring setup — validate PSI/KS thresholds using MLFlow against historical inference batches; define what constitutes a meaningful drift signal for PE and DF specifically.
- Document experiment decisions in MLflow and Confluence documents — the experiment history is a first-class artifact, not an afterthought.
Experience & Background
2+ years applying ML to real-world production problems — not research or hackathon work, but models running in production with real consequences for errors
Experience with geospatial or sequential data — vessel trajectories, routing patterns, H3/S2 grid systems, or equivalent spatial representations
Python proficiency at a level sufficient to implement new features, write dbt models, and script experiments — not just use notebooks
Familiarity with MLflow or equivalent experiment tracking (Weights & Biases, Neptune, etc.)
Desirable:
Domain knowledge of maritime shipping, commodity trading, or cargo intelligence — understanding what a port call sequence or a vessel's draught profile means physically, not just statistically
Familiarity with Redshift or columnar warehouses for large-scale feature queries and dbt (authoring or reading SQL models)
