A behind-the-scenes look at the intake model we built for NimbusHealth's telehealth platform.
Telehealth intake is deceptively hard to automate well. Get it wrong and you either overwhelm clinicians with false urgency flags or miss genuinely urgent cases.
For NimbusHealth, we trained a triage model on de-identified historical intake data, validated continuously against clinician overrides rather than a single offline accuracy number.
The model doesn't make final decisions — it routes patients to the right queue and surfaces relevant history to the provider before the call starts, cutting the time clinicians spend reading charts mid-consult.
Clinician trust mattered more than raw model accuracy. We shipped an early version with conservative thresholds and tightened them only after providers had months of evidence the system wasn't missing urgent cases.
The result was a 58% reduction in average consult wait time, driven less by the model itself and more by how much manual triage work it removed from the front of the queue.
Divya leads applied AI initiatives, helping clients turn data into production-grade machine learning systems.