Traditional actuarial risk leans on unimodal data and linear risk pools. I modelled it instead as the interaction of two vectors: a policyholder's traits and an insurance plan's characteristics that are fed into a two-tower neural network that predicts expected coverage at the micro-level across 407 benefit types. It's a research project, so there's no product link — instead, the model I trained is running live, right here in your browser.

A random policyholder is drawn and scored against a sample of real ACA marketplace plans. Each plan's coverage-match score is the model's summed softmax probability over the benefits that plan actually covers — best vs. worst shows how much coverage the network recovers. Inference runs entirely client-side from an int8-quantized bundle, within 0.0016 of the original model.