MeetPlanner — Hybrid meeting orchestration

The Problem

MeetPlanner's ops team spent the first two hours of every day stitching together meetings across four time zones, two calendar systems, and a Slack backlog that didn't sleep. The bottleneck wasn't the calendar — it was the intent. Most meetings could be a 15-minute async checkpoint, but the request shape ("can we sync?") hid the actual need.

The System

A small agentic pipeline sits between the request channel (Slack + email) and the calendar. It classifies the intent — decision, status, unblock, pairing — and routes the request to the lightest format that resolves it. Only the residual ~30% of requests reach the calendar; the rest get async-templated replies the team can edit and send.

The decision that mattered most: the system never proposes a meeting time without first proposing not having the meeting. Removing a meeting is cheaper than scheduling one, and the model surfaces that option first.

What I built

• Intent classifier on claude-3-5-sonnet with a tight schema (4 classes, no free text).
• Slack + email connectors using thread-safe deduping (one workflow per intent).
• Postgres-backed routing table per team with team_id × intent → format rules.
• A 15-minute checkpoint template generator (markdown + Resend transactional send).
• Edge function for calendar-handoff only when the routing rule says "schedule".
• An operator console that shows what got not-scheduled and why.

Outcome

Eleven of fourteen teams adopted the system in the first month. The ops lead who used to triage at 9am now reviews three operator-console digests per week. Across the org we measured a 70% drop in scheduling rounds per resolved request and an average 12 hours per week reclaimed per team.

What I'd do differently

I overweighted classifier accuracy in week one and underweighted the operator console. Teams trust a system they can audit; they don't trust a system that's quietly right. If I started over I'd ship the console on day three with a hand-labelled corpus and let model quality catch up over the next two weeks. The system's legibility moved adoption more than its precision did.