After two decades of living online, the author mined his own chat logs to chart personal growth. Starting with a 2014 “Your Life in Weeks” grid that felt hollow, he moved from paper journals to Obsidian notes, then to a digital CRM built from archived ICQ, VK, Instagram, Telegram, and Twitter data. The goal: convert messages into a readable life map.

Parsing 1.2 million messages proved noisy: 41% comprised links, media, or fillers. The author filtered short tokens by sampling offsets, then manually curated a denylist of 80 common fillers across languages. After cleaning, 52,000 unique lemmas remained, with a novelty rate plateauing at 6% six years ago. Mapping nicknames like Sasha across platforms required a BERT classifier trained on his chats.

To resolve ambiguous mentions and event types, the author ran 200+ LLM sessions, totaling roughly 15–20 billion tokens and costing about $15 k on Opus. The model produced JSON manifests of bullets, sentiment tags, entity facts, and unresolved ambiguities. A deterministic script then injected these bullets into the vault, yielding a searchable, quantified life narrative that no longer relies on memory alone.