Build a personal work-activity assistant that indexes what a person actually did across their work tools, then answers questions like "what did I do this week?" or "what's my case for promotion?" with grouped, fully-sourced evidence — real links and dates trustworthy enough to paste into a review doc. It must never make live API calls at question time (answer only from the index) and must be honest: if something wasn't indexed, say so rather than guess. Design it in two halves. Ingestion backfills history into a shared persistent index (a SQLite database on a named volume) and keeps it fresh. Make each source its own drop-in step so new sources are easy to add later. Cover activity sources like team chat (e.g. Slack — or your workspace's chat tool), calendar (e.g. Google Calendar — or your calendar provider), code activity (e.g. GitHub — or your git host), and CRM activity (e.g. Salesforce — or your CRM), attributing activity to each user by their identifier/email. Each source runs on a short recurring cron: the first pass for a user backfills a fixed history window (~90 days), later passes only fetch what's newer than a stored per-source cursor. Also expose a "trigger backfill" route that fans out to every source for an immediate pass, so a newly registered user gets indexed right away. Ask queries only the index. Use three route/steps: (1) an AI step that takes the user's question and infers the time range and answer type (summary / highlights / promotion); (2) a read-only retrieval step that pulls candidate evidence for that user and range from the index; (3) an AI step that keeps only meaningful activity, groups it by impact/theme, resolves duplicates to a primary source, and re-verifies in code that every cited link actually came from the retrieved candidates so evidence can't be invented. Use a capable LLM (e.g. Anthropic's Claude — or your preferred model) for the AI steps. Add user-facing pages, gated to space members with identity taken from the authenticated-email header: an Ask page with an ask box, suggested questions, staged loading, answers grouped by theme with evidence cards, and an honest coverage line; and a Setup page where a user registers, sets their work email, optionally adds source handles, kicks off the backfill, and watches counts arrive live. Add a small coverage API returning what the index knows for the caller. Make it multi-user (scope everything by authenticated email), read-only against all external services (write only to your own index, send no messages, create no external records), and honest about coverage limits per source. Tools used: Slack, Google Calendar, GitHub, Salesforce, Anthropic (Claude), SQLite