DIM 9000 analytics stack from zero

The brief

DIM 9000 is a property management SaaS covering residential buildings and apartments. When I joined as an intern in June 2025, analytics meant manual SQL queries, spreadsheets, and reports built on request. No warehouse, no defined metrics, no alerting.

By October 2025 I had taken full ownership of the analytics direction as the sole analytics engineer. The job: build everything from scratch and make it reliable enough that the team could actually depend on it.

What I built

Real-time ETL pipeline

Replaced a legacy daily batch process with a Python Cloud Functions pipeline syncing production PostgreSQL to BigQuery every 15 minutes. The shift from 24-hour to 15-minute data latency unlocked operational use cases that weren’t possible before — real-time building performance monitoring, same-day financial visibility.

Star Schema data warehouse

Designed a Star Schema in BigQuery with a small set of conformed dimensions:

• dim_llc, dim_space, dim_building — the entities the business cares about
• fact_billing, fact_debt, fact_payments — the events and balances

One source of truth across departments. Every dashboard and report reads from the same definitions.

Operational and product dashboards

Shipped v1.0 dashboards for:

• Building/apartment performance — occupancy, financial KPIs, operational status across the portfolio
• CRM feature adoption — tracking which modules teams actually use and where the bottlenecks are

Fully replaced manual reporting workflows.

Automated Slack alerting

Built a custom anomaly detection pipeline: BigQuery → Cloud Scheduler → Cloud Function → Slack Webhook. Migrated the original Google Apps Script implementation to Python Cloud Functions v2 for reliability and scalability.

Designed to be quiet by default — alerts fire when something is actually wrong, not on every data point that moves.

SLA monitoring (“Sniper”) and daily digest

Two additional alerting layers on top:

• Sniper: instant Slack alert for critical support tickets stuck 60+ minutes without an assignee
• Daily heartbeat: automated management digest covering stale backlog (14+ days), overdue deadlines, and ticket volume per building complex

What I’d do differently

The first warehouse modeling pass had too many implicit assumptions baked into the SQL — things that made sense in the moment but required explanation later. The second iteration was deliberately more explicit: named columns, documented grain, no clever coalesces. It’s easier to read and easier to hand off.