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
Before a polymer or a metal reaches a converter, it has been extracted, refined and transformed somewhere specific, with technology that sets both its cost and its carbon. A buyer asking why two suppliers' footprints differ needs that upstream picture by location and technology, with the uncertainty stated. You'll build it as validated data and Python models that run inside the product.
What you'll own
- New feedstock categories. Map pathways from extraction through refining into industrial transformation, with the assets, technologies and system boundaries involved, to a design ready to build.
- The upstream model layer. Datasets and models for where materials are produced, how production varies by location and technology, and the carbon intensity that goes with it.
- License, adapt or build. Judge materiality, coverage and resolution before choosing, and build new methods selectively where existing sources fall short.
- Upstream data products. Evaluate, ingest and reconcile open and commercial datasets, resolve assets across conflicting sources, and ship documented, versioned outputs.
- Validation and provenance. External benchmarks, quantified uncertainty, and each important output carrying its source, vintage, allocation choices and known limits.
- Upstream authority. The person the team asks whether a pathway, production estimate or carbon result is physically and commercially plausible.
What we're looking for
- Deep upstream expertise in at least one raw-material system: oil and gas, mining and metals, forestry, agriculture or land use. Deep enough to challenge an answer that is statistically plausible and physically wrong. We can teach our stack; we can't teach years inside these systems.
- The ability to transfer it. You learn unfamiliar commodity systems fast and build a defensible model of them.
- Strong GHG and LCA knowledge. System boundaries, allocation, land-use change, methane and other non-CO2 emissions, and where common datasets fall short.
- Applied data science in Python, documented well enough to be reviewed and productionised.
- You've owned an important dataset or model that others relied on: its methodology, quality, evaluation and maintenance. Walk us through it. That's where every conversation with us starts.
Nice signals: geospatial, raster or satellite-derived modelling; familiarity with OPGEE, PRELIM, OCI+, ClimateTRACE, GEM or Trase; more than one commodity system; production software experience; a Master's or PhD in a relevant field.
Stack: Python, FastAPI, Postgres, Dagster, AWS. A signal, never a requirement.
