How RadAI works

The model

CheXagent-2 (3B) is a multimodal vision-language model from Stanford AIMI, trained to read chest X-rays and describe findings in natural clinical language. RadAI prompts it once per anatomical region, then asks it to condense those findings into a single impression — mirroring how a radiologist structures a read.

The request pipeline

1. 1

   Upload, in the browser

   Your X-ray is downscaled client-side and sent to a Next.js API route. AWS credentials live only on the server — they never reach the browser.

2. 2

   Asynchronous invocation

   The server stores the image in S3 and calls SageMaker Asynchronous Inference. Async (rather than real-time) suits a model whose full report takes a minute or more, and it's what makes scale-to-zero possible.

3. 3

   Section-by-section inference

   On the GPU, CheXagent standardizes the image and generates a finding for each of the five regions, then summarizes them into an impression.

4. 4

   Poll & render

   The browser polls for the result in S3 and renders the structured report. The bundled sample skips all of this with a precomputed result, so the demo is instant.

Serving that scales to zero

A 3-billion-parameter model needs a GPU, and GPUs are expensive to leave running. So the endpoint is configured to scale to zero instances when idle: there is no cost when no one is using it. When a request arrives, autoscaling spins up a GPU, runs the analysis, and scales back down afterward.

The trade-off is the cold start: the first scan after a quiet period waits while a GPU boots and the model loads. For a portfolio-scale tool, that's a worthwhile price for near-zero idle cost.

Stack