Surfaces

Create a course from any text sequence. Define readings in a target language (Spanish, Icelandic, Latin), extract vocabulary, generate dialogs, and run thematic RP conversations.

Errors feed a morphosyntactic taxonomy wired to Thompson Sampling: which correction strategies produce retention, which produce frustration.

This feedback will eventually power an adaptive learning layer that understands a user's specific points of confusion.

Mobile and web frontend via Expo. Federated GraphQL via Hive Gateway. Tofu-managed multi-service mesh on Cloud Run, each with isolated schemas in Supabase.

Hub-and-spoke reference architecture for multi-domain health applications. New data sources plug in without touching existing services.

Equipping agents with fully-featured access to external surfaces where the stakes are real.

The domain engineering is deliberately heavy: hallucinations posted publicly are not recoverable, so output space has to be bounded before anything goes live.

The longer-term question is how to relax these constraints as the learning layer matures.

35+ MCP tools over a general interface to any model with ComfyUI nodes. Distributed GPU across Fly.io and RunPod. 467 tests.