Commodity Sourcing Data Scientist

Atmospheric AI; Greater London; Hybrid; GBP 75,000 - 110,000; Full time

Job summary

Atmospheric AI Limited is building quantitative models to infer supplier networks, origins, and flows from incomplete evidence in a 16 trillion-dollar materials economy. You’ll design probabilistic, network-based approaches that respect physical constraints and deliver traceable results that scale across suppliers and facilities.

You’ll own the sourcing model, product-origin inference, and the evidence layer, choose appropriate validation strategies, and ship production-grade Python modules with

Qualifications

• Strong background in probabilistic modelling and network analysis.
• Experience with validation under limited ground truth.
• Ability to reason about mass balance, capacity and transport.
• Proven production-grade Python development with tests and documentation.

Responsibilities

• Design and implement sourcing models to infer suppliers and flows.
• Create product-origin inferences from finished products to facilities and converters.
• Ingest and reconcile production, capacity, trade and shipment data for the evidence layer.
• Choose validation strategies and ensure credible tests for mostly unobserved targets.
• Deliver reproducible Python modules maintained by the broader team.

Skills

Network modelling

Probabilistic modelling

Validation mindset

Physical-system reasoning

Production-quality Python

Education

Master's degree in a quantitative field

PhD in a quantitative field

Tools

Python

PostgreSQL

Dagster

AWS

Job description

Everything physical is made from raw materials: polymers, chemicals, metals and minerals, a $16T economy that software has barely modelled. Atmospheric AI computes what those materials should cost, what carbon they carry and where their supply is exposed, from the physics of how they're made, so manufacturers can decide and negotiate from ground truth.

The problem

A manufacturer knows its direct suppliers. It rarely knows who supplies them, which plants a grade came from, or how exposed it is two tiers up. Most of those flows are never observed directly. They have to be reconstructed from incomplete and contradictory production, trade and shipment evidence. You'll build the models that do it, and say plainly where the evidence runs out.

What you'll own

• The sourcing model. Identify feasible suppliers, infer facility-to-facility relationships and allocate volumes, within the physical and commercial constraints that decide who can supply whom.
• Product-to-origin inference. Work backwards from a finished product or grade to its plausible converters, producers and facilities.
• The evidence layer. Ingest and reconcile production, capacity, trade, customs and shipment data, and stay accountable for whether the evidence supports the flows.
• Method choice. Network flow, optimisation, Bayesian inference, discrete choice, simulation. You pick on validation performance and add complexity only when it pays.
• Validation. Credible tests for targets that are mostly unobserved, using known relationships, held-out shipments and aggregate constraints.
• Useful uncertainty, in production. Best estimates with relationship probabilities and confidence, shipped as tested Python modules the wider team can maintain.

What we're looking for

• You've recovered hidden relationships or flows from incomplete, contradictory evidence, at scale. Walk us through a model you built and how you knew it was right. That's where every conversation with us starts.
• Network and allocation modelling. Probabilistic modelling, network science, operations research, optimisation or discrete choice.
• A validation mindset. You know how to evaluate a model when ground truth is scarce, and you recognise when a problem is underdetermined.
• Physical-system reasoning. You use capacity, compatibility, mass balance and transport to rule out results that are statistically convenient and physically impossible.
• Production-quality Python. Interfaces, tests, review and reproducibility, and comfort with large, messy data and inconsistent identifiers.

Nice signals: physical commodity flows or global trade; customs records, bills of lading or vessel tracking; agent-based modelling; Databricks and MLflow or equivalent; a Master's or PhD in a quantitative field.

Stack: Python, Postgres, Dagster, AWS. A signal, never a requirement.