Milo

I quit nutrition tracking twice and my training log once — not because it was too much work, but because I simply stopped thinking about it. Milo starts right there: every evening it checks in on Telegram, it remembers, and it points out connections no single app can see. n8n wires the pieces together, everything Milo knows lives in Postgres 16 with pgvector, and Claude puts it into words. Besides Telegram, Milo runs in a private web app with passkey login, and Whisper transcribes voice messages. Built for myself, not as a product.

• The prompt is rebuilt from eight SQL blocks on every message — SQL does the maths, such as weekly progress and gaps between sessions, never the model
• Training from Hevy via webhook; weight, food, sleep and steps from Apple Health via a Shortcut; habits ticked off with a button
• A nightly job spots comebacks, breaks, new top weights and habit streaks and stores them as memories, followed by a weekly review every Monday
• Domain knowledge only from 49 evidence cards, each with a DOI verified via Crossref
• An error workflow reports failures on Telegram; encrypted backups run daily

• n8n
• PostgreSQL
• pgvector
• Claude API
• Telegram Bot API
• Hono
• React
• TypeScript