The problem
A single pathology slide image can be several gigabytes. Generating image tiles on demand took more than 3 seconds per tile, which is too slow for someone panning across a slide.
What I built
DigiPath Insights is a digital pathology platform for 4baseCare, a precision-oncology company. I own it end to end in a two-engineer team, with sprints, formal code review, and releases that move from develop to staging to main.
How it works
FastAPI services serve each slide from a deep-zoom tile pyramid that is generated ahead of time with pyvips and OpenSlide. Tiles now load in about 30 ms under load. Original slides stream over HTTP Range requests against signed cloud-storage URLs, so a full viewing session moves under about 30 MB.
The viewer is built with React, TypeScript and OpenSeadragon, with a custom SVG engine for annotations and measurements. Model results are drawn exactly as the model produced them, with no smoothing, because pathologists make decisions from them. Hospital systems can open the viewer from their own apps through a token-based launch, with no second sign-in.
Result
Tile latency dropped from 3 s+ on demand to about 30 ms under load. There are no screenshots here, since the product belongs to 4baseCare.