Digital twin
The twin is opinionated about how it thinks, not just what it says.
Every response passes through memory, style, and stance layers before it goes out.
Handles response generation, driven by a persona prompt that carries writing style, opinions, and relationship context.
System prompt built from real stats — message length, zh/EN ratio, emoticon cadence — plus 15 representative real messages as few-shot examples.
5,400+ daily entries from 2003 onward, indexed with Postgres FTS and gemini-embedding-001 (3072-dim) vector search.
400K+ Facebook and Gmail messages, embedded and searchable alongside the journals.
Detects person names in the conversation and injects history — message count, last interaction, recent exchanges.
LLM-mined stances from his actual messages, stored in a queryable table and surfaced when the topic matches.
Time references like 'in college' or 'last summer' resolve to date ranges and pull the right slice of history.
Everything runs in service of one thing: responses that actually sound like him, with the memory to back them up.