Real-time operations dashboard for dispatch and fleet health.

Aether’s dispatch team juggled a telemetry viewer, an incident chat, a spreadsheet for fleet utilisation, and a legacy map that drifted from ground truth. Night shifts burned time reconciling versions before acting. Fleet managers distrusted utilisation figures because definitions differed per region.

The COO asked for a single operations surface before peak season — ten weeks, no net-new data science team. Engineering demanded WebSocket-first updates; finance demanded cost controls on cloud spend.

We refused to “just embed Grafana everywhere.” The constraint was actionability: every tile had to map to a runbook step, not a vanity chart.

// TODO: Verify with client legal before publishing — operational descriptions and incident definitions.

Framing. This was an SRE-shaped product: SLOs, error budgets, and human runbooks shipped alongside widgets.

Architecture. Next.js UI, Node gateway for WebSocket fan-out, TimescaleDB for hot telemetry paths with retention policies that finance signed. We used strict schemas for event payloads so partial outages degraded gracefully.

Phasing. Weeks 1–3: read-only consolidation with feature flags. Weeks 4–7: interactive dispatch actions behind role gates. Weeks 8–10: hardening, load tests shaped from peak-hour captures, and on-call dry-runs.

Judgment. We delayed map clustering polish until websocket stability hit three nines in soak tests — UX stewards disagreed initially, then thanked us.

// TODO: Verify with client legal before publishing — stack and reliability claims.

Median incident response time for Sev-2+ dispatch incidents fell 33% after runbook-linked tiles shipped. Fleet utilisation — revenue miles per active vehicle week — rose 18% with the same safety constraints once deadheading visibility improved. The dashboard and gateway maintained 99.95% uptime across the peak window in scope.

Four brittle internal tools retired; training took two afternoon sessions because flows mirrored how dispatch already talked.

// TODO: Verify with client legal before publishing — uptime measurement window and definitions.

We underestimated keyboard-first workflows — mouse-heavy UI failed at 3am. We added shortcuts mid-programme. We were surprised how much trust came from printing a one-page incident summary — PDF export mattered more than animations.

Next time we would embed a “definition drift” alert when two metrics disagree by more than a threshold — catches spreadsheet ghosts early.

// TODO: Verify with client legal before publishing — operational reflections tied to a real client.