GPUs on demand
Modal
Every model that needs a GPU runs on Modal. Each app scales to zero, so it costs nothing while no one uses it. Terraform deploys the walking tour apps with the rest of the backend.
The narrator’s voice
Kokoro text-to-speech on T4 GPUs turns each guide into MP3s, with up to three containers at once. Modal checks a proxy token before any GPU starts.
Walking ToursA guide you can talk to
A live speech pipeline on one T4 in the EU. Silero, Smart Turn and Parakeet listen, Claude Haiku answers and Kokoro speaks, over WebRTC. Six walkers share one container.
Walking ToursTraining a sign language model
Staged training of a pose encoder and a LoRA adapter on Llama 3.2 1B, on L4 and A10G GPUs, with data and checkpoints on a Modal Volume. A full run costs $4–10 of a $100 budget.
BSLScribe
A lesson learned: cuDNN 9.5 keeps Kokoro on the GPU, at 0.2 seconds a sentence. Newer versions quietly fall back to the CPU, at 1.5 seconds.