Selected work
UCSF Health × Stanford MedicineAI Software Engineer – Healthcare SystemsJul – Nov 2025

Caduceus: Clinical AI Platform

Caduceus is the Go backend for a stroke-rehab telehealth platform. Patients train in a mixed-reality headset, a multimodal model scores each rep from the video, and the result lands in their clinical record while the session is still running. Behind it, a Kafka, Redis, and Cassandra pipeline with idempotent writes, because a duplicate rep is a corrupted chart, not a minor bug.

Architecture
01Headset frames02ffmpeg + GCS03Multimodal rep eval04Postgres + event pipeline

Decisions

Exactly-once by composite key

Results were keyed on exercise and rep number, so a retry from a flaky headset could overwrite a different patient's row. Widening the conflict key to exercise, patient, rep, and step makes a retry idempotent instead of destructive.

Let the compiler judge the model

New exercises are generated as Scenic programs that must compile against a fixed library of allowed objects and actions. Prose can lie fluently. A program either compiles or it does not.

A boolean you can chart, reasoning a clinician can read

Each rep evaluation returns a bool for the dashboard and a paragraph for the edge case review. One is measurable across model versions, the other is accountable to a human.

Boring beats clever under load

Circuit breakers and retries keep one slow dependency from cascading, and the target is p99, not the mean. Trust lives in the slowest one percent.

— Amisha

Go · Postgres · Google Cloud Storage · Gemini · Kafka · Cassandra