Guest named Sar in the transcript (software engineering group lead at monday.com); surname unclear. Host: Roni.
The team building Harmony, a new go-to-market product at monday.com (the same team that built monday’s AI note-taker), wanted to capture early “shadow data”: meetings, Slack, WhatsApp, late-night architecture debates, and hallway talks. v1 was one person’s n8n flow updating three files (decisions, action items, open questions). The files bloated and hallucinated, and per-person workflows wouldn’t scale to 20 people. After Karpathy’s LLM Wiki post, they built a persistent, interlinked wiki in a git repo: concepts, domains, entities (companies and roles), decisions, action items. A Claude skill on each person’s machine runs every ~30 minutes, pulls new data through their personal MCPs (note-taker, Slack, WhatsApp, docs), skips already-ingested raw items, updates many pages per meeting, and pushes. Answers worth keeping get saved back. Access: a Wikipedia-style site, and a WhatsApp agent (“Siena”, on OpenClaw) that answers anyone 24/7. Impact: an executive vision deck in hours instead of weeks, Google Ads messaging written from the brain with good leads, and cross-team alignment without syncs. Next: autonomous discipline agents built on the brain. Costs: private conversations leaked into git history (repo wiped more than once), and nothing is private anymore. Advice: start simple and local before automating at scale.
Show: startup-for-startup
Concepts: second-brain · llm-wiki · ai-guardrails