I built my own analytics to see what visitors ask my chat

I wanted to know what visitors actually ask my portfolio chat — without sending their messages to Google or Plausible.

So I built my own dashboard.

/admin on furevikstrand.cloud is a server-rendered Next.js page, basic-auth gated, that returns 404 when the env vars aren't set so it isn't discoverable. It reads from a MySQL table I log every chat event and page visit to (best-effort — if the DB is down, the request still succeeds).

13 panels, two sections.

Traffic: conversion (visits → chat sessions), top referrers, countries (resolved locally via geoip-lite, no external API), pages, device split, UTM sources.

Chat: fallback rate over 30 days, reply-source split (Claude / Ollama / fallback), locale split, top topics, session-depth distribution, sessions per day, and the most useful one — recent fallback misses: every question my keyword matcher couldn't answer.

That misses panel is the work queue. Each row has a "copy ts" button that emits a paste-ready KnowledgeEntry stub — derived keys from the message, locale + timestamp comment, TODO shells for replies in en/no/pt. Closes the loop from "spotted a gap" to "shipped an answer" in about 30 seconds.

What's the smallest piece of your stack you've pulled in-house instead of paying a SaaS for?