Case study

Store-first AV fleet monitoring

One fleet view across Yealink, Logitech, Poly and Neat, built by polling vendor APIs into one store and driving alerts, reports and room pre-checks from it.

Date
Role
Designed and built
  • n8n
  • Supabase
  • PostgreSQL
  • Yealink YMCS
  • Poly Lens
  • GraphQL
  • Logitech Sync
  • mTLS
  • Neat Pulse
  • Microsoft Graph
  • Docker

Problem

The meeting-room estate was spread across several vendor clouds (YMCS, Logitech Sync, Poly Lens, Neat Pulse), each with its own portal, its own idea of “offline” and its own alerting.

Answering “is this client’s estate healthy today?” meant checking several portals, and catching a faulty room before a meeting rather than during one depended on someone happening to look.

Constraints

  • Every vendor API is different: Yealink YMCS and Neat Pulse are REST, Poly Lens is GraphQL, and Logitech Sync requires mutual TLS with a client certificate.
  • Vendor APIs are rate-limited, so polling has to be scheduled and modest.
  • Vendor and client credentials must never leave the server side or end up in the app image.
  • Actions such as reboots run on live customer hardware, so they must never interrupt a meeting in progress.

Architecture

Store-first fleet monitoringScheduled pollers read Yealink, Poly, Logitech and Neat APIs into one Supabase store. The fleet app, alerting, reporting and room pre-check read from the store. Guarded actions such as reboots and firmware pushes go through a webhook dispatcher.Vendor cloudsupsertclears cooldownreboot / firmwareYealink YMCSRESTPoly LensGraphQLLogitech SyncmTLSNeat PulseRESTScheduledpollersn8nSupabasePostgreSQLOfflinedetectionevery 2 minAlert resethourlyDeduplicatedalerts30-day clientreportsM365 calendarsGraph + testcallsDaily roompre-checkFleet appCloudflareAccessWebhookdispatcherin-call guard

The core decision was store first. Scheduled pollers write normalised device state into one database, and the fleet view, alerts, reports and room pre-checks all read from that store, never from the vendors. Only the pollers and explicit, guarded actions talk to vendor APIs.

That gives one definition of “offline”, one history for reporting, and a vendor API call budget that doesn’t grow as features are added.

What I built

The AV Device Manager. A web app with one fleet view across Yealink, Logitech, Poly and Neat: firmware status, config-drift checks against an approved baseline, warranties and contracts, room utilisation and prioritised alerts. Remote reboots run behind an in-call guard, so a device in a meeting isn’t restarted, and firmware pushes go through the same path.

An n8n backend. Self-hosted n8n runs a webhook dispatcher for every app action, plus scheduled sync, alerting, heartbeat and reporting workflows, with Supabase (PostgreSQL) as the data layer.

Offline detection with deduplication and cooldown. Two workflows split the job:

  • A detection workflow runs every 2 minutes. It finds devices that have gone offline, raises one alert per device and records that the alert has fired.
  • A reset workflow runs hourly. It clears the alert state for devices that have come back, so the next genuine outage alerts again.

Splitting detection from reset stops repeat alerts: a flapping device raises one alert, not one every two minutes.

Automated client reports. 30-day client reports are generated automatically from the stored history, so they cost no extra vendor API calls.

Room pre-check. A daily check across around 100 rooms combines Microsoft Graph calendar data and test calls with stored device health and call-quality (QoE) data, so a booked room with a problem shows up before the meeting rather than during it.

Secure packaging. The app ships with Docker Compose as an nginx container with no secrets in the image, behind Cloudflare Access service-token authentication. Access is role-based with per-user hashed keys, admin actions are enforced server-side, and every call is written to an audit log.

Results

  • One fleet view across four vendors instead of four portals.
  • Offline alerts deduplicated: one alert per outage instead of repeats.
  • 30-day client reports produced automatically from stored data.
  • The daily room pre-check covers around 100 rooms, catches faults before meetings start and saves tens of hours a week.

What I’d do next

  • Add trend-based alerts (for example, a device whose call quality is slowly degrading) alongside up/down state.
  • Give clients a read-only view of their own report data instead of sending files.
  • Extend the pre-check to more rooms and feed its results back into alert priority.