Three AI workloads laid out as Modal apps, one per load shape. For each: how the work splits into functions, classes and storage, where one request spends its time, and how cost and latency move with load and cold starts. Each app has a naive and a tuned build; switch between them with the toggle at the top right of any app page.
| App | Shape | Governing number | What sets it |
|---|---|---|---|
| ocr-pipeline | Throughput-bound | pages/s per GPU, $ per 1k pages | Fan-out across CPU and GPU stages, in-flight requests matched to the engine batch |
| voice-edge | Latency-bound | end of speech to first audio | Streaming STT and TTS on one socket, one container slot per live call, cold start as the worst case |
| text-serve | Concurrency-bound | time to first token, $ per 1M tokens | Continuous batching on a shared engine, KV cache against context length |
Prices are Modal list prices per hour. Code on the Stack tab is a sketch of the configuration, not a deployed app.