Life Coach — privacy-first personal AI
A coaching agent with months-long memory — temporal knowledge graph, contradiction resolution, and a three-lane privacy architecture that keeps the private parts local.
A personal AI coach that remembers across months instead of messages: conversations, energy and mood signals, goals, and interventions live in a temporal knowledge graph with validity windows and automatic contradiction handling. Provenance is explicit — every entity records whether the user or the coach said it, which killed a real echo-chamber failure mode where the coach was citing itself as evidence.
The architecture decision I'd defend in any design review: privacy lanes. Cloud-lane conversation runs through aliasing (names and employers never leave home in the clear); a local-GPU lane handles sensitive reasoning on my own hardware; a private lane prefix guarantees a turn is never sent anywhere. Privacy by architecture, not by policy.
It's also the most operationally hardened service on the platform — a 13-finding audit (dead salience scorer, untested formatter that silently killed context-building for 36 hours, coach out-talking user 25:15) had 11 findings closed within 48 hours, verified by telemetry.
Status & limits — shown here as architecture only; the data is my life and stays private. A reliability sprint on one degraded dependency is in progress.