Milo — Screenshot 1 of 5
Screenshot 1 of 5
AI Agent

Milo

Staying on track without having to think about it.

Overview

Four apps record my day: Hevy my training, MyFitnessPal my food, the scale my weight, the watch my sleep. Milo reads them all and tells me what they mean together. It started as a Telegram bot; today it is a native app for iPhone and iPad. It wakes me until my morning routine is done, follows up after training, checks in every evening and can be talked to live. Behind it sits a modular system that grew piece by piece: Postgres as memory, n8n for the workflows, GPT-6 Luna for the words. Built for myself, not as a product.

Highlights
  • Native SwiftUI app for iPhone and iPad: Today with weekly goals and streaks, chat, widgets and an AlarmKit alarm that rings again until the routine is done
  • Milo reaches out on occasions: a greeting after the alarm, a follow-up right after every workout from Hevy, an evening check-in, pushed to every device
  • Real-time voice conversation over WebRTC with GPT-Live, using the same prompt as the chat, built in SQL
  • SQL does the maths, never the model: every reply starts from a briefing of eight blocks, and domain knowledge comes only from 49 evidence cards with verified DOIs
  • Memory in pgvector, condensed every night; errors reported by email and message, encrypted backups every day
Tech stack
  • SwiftUI
  • n8n
  • PostgreSQL
  • pgvector
  • OpenAI API
  • WebRTC
  • Hono
  • React
  • TypeScript