A carbon model for campus AI use, built to be argued with

The model runs from people to prompts to energy to carbon: population, queries per person per day and adoption per tool give a daily query volume; energy per query turns that into kilowatt hours, and the grid’s carbon intensity turns those into kilograms of CO₂. The grid figure is New England’s from the EPA’s eGRID; energy per query sits in the published range from Samsi et al. (2023), cited on the page where it is used.

Because every number flows from the same small set of inputs, the tool is a way of making a set of assumptions legible and adjustable, and it reports its own uncertainty, plus or minus 40 per cent, rather than presenting estimates as measurements.

The state-level choropleth is drawn without a mapping library: TopoJSON features are projected to path strings and rendered as raw SVG, with the topology client imported dynamically in parallel with the data fetch. It carries a cancellation guard, and if the map request fails it still renders the data rather than an empty box.

The three model-backed routes check for an API key before constructing a client, so every page renders fully without one — a prototype anyone can run on handover.