Work 50,000 scans, and a pipeline that had to trust its own labels. Wake Forest CAIR, May to August 2025.

A summer at the Center for Artificial Intelligence Research, serving PyTorch models on GCP against a real hospital-scale imaging dataset, where the hard part wasn't the model, it was making 50,000 scans and 10 TB of data trustworthy enough to train on.

Where: Wake Forest CAIR · Winston-Salem, NC
Dates: May 2025 – August 2025

50,000+imaging scans processed
10 TBof data
95%radiologist concordance

Wait Chapel and the quad on Wake Forest's campus

01 · Model serving at hospital scale

Cloud-native AI services served PyTorch models on GCP with Terraform, processing more than 50,000 medical-imaging scans and 10 TB of data, not as a one-off batch job but as infrastructure meant to keep running.

02 · The real work was the labels

Automated data curation with MNI152 standardization, using Python, NiBabel and SimpleITK, reduced preprocessing from 4 hours to 18 minutes per case and reached 95% radiologist concordance, the number that actually mattered.

03 · Monitoring that earns trust

Delivered monitoring tools alongside the pipeline, holding P99 latency at 1.6 seconds, so the team could see the system behave in production, not just read a paper metric.