A command center for a real-estate investment CRM.

The founder of a US real-estate CRM had a real ops problem hiding behind an AI problem. Their day ran across seven surfaces and a pile of one-off AI tools that did not talk to each other. Every prototype worked in isolation. The operation around them did not exist.

The founder's day-to-day lived in Linear for engineering, Slack for the team, Gmail for external mail, Google Calendar for meetings, Google Sheets for the personal journal, a pile of scattered AI tools for email drafting and meeting summaries, and their own memory for everything else. Each tool worked. None of them talked to each other.

The consequence was invisible at the surface and expensive underneath. When the founder logged a call, the transcript went to one tool. The follow-up drafted in another. The tasks it generated ended up in Linear only if someone remembered to move them. The ideas it surfaced sat in a notes app until they were forgotten. The scattered AI features were islands, and the labor of connecting them fell back on the founder every day.

The prototypes were fine. What was missing was orchestration: one place to type, one place to look, and one system that knew which context belonged to which decision.

The founder types a natural-language request. A router decides whether the question is about sales, email, ideas, tasks, meetings, calendar, rewriting, or the business as a whole, and dispatches to the specialist agent that owns that domain. Eight specialists behind one interface. Each with its own prompt, persona, memory, and rules.

Ask "what should I work on today?" and the task specialist returns a priority-sorted list built from Linear plus the personal task queue, weighted by 90-day business plan alignment. Ask "what did I discuss with the last three prospects?" and the sales specialist synthesizes from the last three call transcripts. Ask "draft a reply to the last email from the vendor" and the email specialist reads the full thread, pulls the recipient's context, and produces a personalized draft.

The email surface auto-classifies inbound mail into three buckets: high-priority-needs-reply, high-priority-informational, and promotional. Every classification can be corrected with a thumbs-up or thumbs-down plus a text reason. Corrections feed the training loop. The ideas surface validates every new idea against the 90-day plan and auto-routes: financial angle emails the CFO, technical angle books time with the CTO, crossover mails both on one thread. The daily check-in agent asks about yesterday's work, sleep, and workout, and surfaces a pattern break when three or four consecutive days go sideways.

Every feedback interaction writes a snapshot. Thumbs-down on an email classification. A manual task reorganization with a stated reason. A rejected idea escalation. Each snapshot captures timestamp, circumstances, context, and the reason the human gave. A separate training agent reads those snapshots against every specialist's prompt files (prompts.md, personas.md, memory.md, rules.md, heartbeat.md) and proposes mutations as a diff. The human accepts or rejects. Accepted diffs update the prompt. Rejected diffs discard with no state change.

The prompt files literally change every week, and every change is traceable to a specific human correction made on a specific date for a specific reason.

The command center is live and in daily use by the founder and three team members. Morning email triage, opening check-in, recurring task reorganization, the ideas surfaced throughout the day, and the follow-ups from every sales call all run through one interface. The scattered AI tools that existed before are gone.

Under the hood, every external call to Linear, Slack, Gmail, Calendar, and Sheets is wrapped in retries, timeouts, and fallbacks. Multi-user isolation partitions state and training data cleanly: the founder's training snapshots do not leak into the CTO's agents, and the CFO's rules do not affect the sales lead's classifier. Every generation is traced. Cost per call is logged. When an invoice would have spiked, the alert fires before the founder sees the bill.

The compounding is the point. Not because a model got smarter, but because the prompts, rules, and memory files get tuned to this team's operating logic every week, on their own data.

Anas