I write at Donuts with Chopsticks on Substack.
Five things true at once: the record, the ceiling, the wall, and the instrument — the full landscape of healthcare AI evaluation in four parts.
A scroll-driven visual essay: healthcare AI benchmarks are saturated, and the evaluation that answers “does this help?” looks like product analytics — in private.
The precharting funnel rendered as live particles — swarm → the doctor → extraction and reasoning lanes → the usage log.
Capability benchmarks are saturated; the evaluation that answers “does this help?” looks like product analytics, and it happens in private.
Ontologies as data poetry — reasoning over biomedical knowledge graphs in Python, landing on a rare-disease example.
A playful Altair analysis built from MTA data.