MKMoustafa Kasem

Senior Integrations Engineer · Insights & Integrations Consultancy · Geotab

From side project to how we deliver

Proof-of-concept and production deployments: from months to weeks, if not days.

It started as my own side project. Now a new customer integration adds only its own business logic to a proven architecture:

  • One specone standard way to build it
  • One pipelineone way to ship it
  • One dashboardone place to watch it

Support and solutions engineering run and build on it.

apps live in production, all reporting to one dashboard, plus 4 in test
14
minutes to onboard an engineer new to the repo, with AI skills that scaffold an app from their requirements
<15
engineers needed to re-run a customer report
0

How it grewShown at an internal demo

  1. MarOne repoIntegrations move into one repo with a shared template, CI and image build.
  2. Apr–JunDashboard and rerunsA Cloud Run dashboard with Google sign-in. Rerun requests travel through Pub/Sub. Demoed in June.
  3. Jul–SepSelf-service deploysA plan-then-apply tool deploys a container from one YAML manifest. Demoed in September.
  4. OctAI scaffolds an app from requirementsSkills interview the engineer, then generate the app, its CI job and its deploy config in minutes. Architecture demo, with AI triage, in October.

1 · Why it matters

Why it was worth building

Delivery scales

AI skills scaffold a new integration from the engineer’s requirements, and it ships through one pipeline, so delivery no longer waits on one engineer.

Problems surface early

Every app reports its health to one place, and an AI triage proof of concept adds the likely cause and next steps. Fixes land once in shared clients.

Ownership spreads

Support re-runs customer reports without an engineer, and solutions engineers build on the platform.

2 · How it’s built

Built once, reused by every app

Standard parts only: Docker VMs, Vault, BigQuery, Pub/Sub and Cloud Run. A new integration adds just its business logic.

How a new integration ships

The same steps for every integration, from idea to production.

generates every app imports merge to main publishes image ships App specone YAML fileAI 8 apps · one framework Rail syncRail downtimeRail usageOdometer feedWatchdog feedStops fileTracking writerData package Shared librariesMyGeotab · GCP · Vault · AppRunner · alerting App catalogrebuilt every merge CI pipelinetests · scan · image Deploy toolplan → apply · backup Docker VMs on GCPCloud Run fns ready
Select a box to see why it matters.

What it changes

  • AI scaffolds the appA skill interviews the engineer about their requirements, writes the spec and generates the app, its CI job and its deploy config in minutes. A new engineer is onboarded in under 15 minutes.
  • Fix once, every app gets itShared MyGeotab, GCP and Vault clients. A paging bug was fixed once in the shared client, not 15 times.
  • Every merge is checkedTests and a security scan run on every MR. The app catalog is rebuilt on every merge, so it can’t drift from the code.
  • Self-service deploysA plan-then-apply deploy tool with automatic backup. I deployed the team’s first Cloud Run function through Terraform.

The operations loop

How engineers and support see and fix what’s running.

unhealthy alert diagnosis status reads sign in rerun runs it AlertsYAML per app → Chat AI triageproof of concept Support teamSolEng · engineers AppsDocker VMsCloud Run functions BigQuery contract heartbeats · every 60 srun historyrerun requests Dashboard Cloud Run Google sign-in adminrerunviewer Pub/Subrerun queue · 10 min hold proof of concept rerun loop
Select a box to see why it matters.

What it changes

  • One contract for every appEach app writes a heartbeat every 60 seconds and a start and finish row per run, so a new app appears with no custom work.
  • Role-gated accessGoogle sign-in with admin, rerun and viewer roles.
  • Reruns without an engineerA rerun waits 10 minutes so a mistake can be cancelled, then goes through Pub/Sub to the app.
  • Health is monitored, not discoveredAlerts are declared in YAML per app and posted to Chat, so integration engineers can resolve issues proactively.
  • AI triage Proof of conceptAdds the likely cause of an unhealthy app and the next steps to resolve it, so whoever picks up the alert starts with a diagnosis. Runs in a test project.

3 · Who runs it

Others run it and build on it

Run it without me

  • Customer supportRe-runs customer reports from the dashboard, with no engineer in the loop.
  • Solutions EngineeringWrote a support guide around the dashboard.

Build on it

  • A regional Solutions EngineerContributing code to a pipeline and preparing to operate it.
  • Another Solutions EngineerBuilt a bulk importer based on my import tools, for ad-hoc customer onboarding of tens of thousands of users and groups.
  • Data scientistsDeployed a backend service for a customer project with my deploy tool, to support their machine-learning analysis work.

Dashed = a manual step the platform removes

Building a new job

Before
  1. Write the job
  2. Write a scheduler (manual step)
  3. Add logging (manual step)
  4. Copy to a VM (manual step)
  5. Run it by hand (manual step)
  6. no monitoring, no alerts
Now
  1. AI skills scaffold the app
  2. Add the business logic
  3. Merge and deploy
  4. built in:
  • Scheduling
  • Monitoring
  • Chat alerts
  • Run on demand

A job goes wrong

Before
  1. Job fails quietly
  2. Customer notices
  3. Ticket (manual step)
  4. Engineer logs in to the VM (manual step)
  5. Reads the logs (manual step)
  6. Fix
  7. found late
Now
  1. Job misses a run
  2. Health turns red
  3. Alert in Chat
  4. Fix
  5. flagged in Chat

Re-running a customer’s report

Before
  1. Customer asks
  2. Ticket (manual step)
  3. Engineer with VM access (manual step)
  4. Script run by hand (manual step)
  5. needs an engineer
Now
  1. Support
  2. Dashboard
  3. Rerun
  4. self-service