JuanMonsalve

A full turn, from microphone to spoken answer, takes 135 milliseconds measured on the target machine. All of it happens there: offline, with no third-party services, and without a single byte of audio leaving the device.

That number is the result of an architectural decision: the language model is not on the critical path. Commands resolve through four stages of increasing cost, and only the ambiguous ones reach the model, so what governs latency and spend is not the model's accuracy but the share of commands that need it at all. That share is measured on every release.

Execution is allowlisted: the model picks an intent and its arguments inside a validated schema, and never generates commands. The project carries 307 automated tests, and the full intent benchmark runs before any threshold is touched.

An amateur football league runs its tournaments, its collections, and its payroll on this, every day and in production. Twenty-two domain models split across an API and a desktop client, both strictly typed. It began as a monolith refactor; today it is a product that installs, updates, and maintains itself.

The decision that changed the product most was about the domain, not the code. Every football schema models a match as home versus away, and that is how this was built; but amateur leagues rent neutral fields and both teams travel to the same place. The distinction meant nothing and forced the administrator into an arbitrary choice on every match, so it was removed, data migration included.

Deployment and operations are mine as well: containers on a self-hosted server, a certificate that renews itself, a database with no exposed port, daily backups, and a script that restores the latest one into a separate database to prove the backups work. The client updates from that same server with no configuration from the user, and permissions follow the job: the administrator handles accounts and payroll, the scorekeeper only sees the day's cash box.

A service that receives contact forms, drops whatever does not need a model using cheap rules, and classifies the rest through the Anthropic API, validating the response against a schema before storing it.

The goal is cost control: every call is recorded with its token count and spend, and an endpoint reports what share of submissions needed the model at all. It is the same judgment as Apollo, applied to a service that bills per use.