From zero to live-ops analytics in 90 days
A cloud data stack for a venture-backed Roblox studio — its first analytics foundation across five branded game IPs.
| Client | A venture-backed (seed-stage) Roblox studio of ~30, building branded worlds and games for global consumer brands. Anonymized at the client's request; later acquired by a publicly traded media company. |
|---|---|
| My role | Head of Data & Analytics, reporting to the CEO and Head of Product. |
| Engagement | Full-time, ~3-month build (design → build → adoption). |
| Core stack | Google Cloud — BigQuery, Pub/Sub, Dataflow, Cloud Run / Functions, Cloud Storage, Dataform, Looker, Cloud IAM; on top of Roblox (Luau) and Google Workspace SSO. |
Context
The studio builds custom branded worlds and games on Roblox for the marketing teams of global consumer brands across fashion, beauty, music, and entertainment IP. With five new game IPs launching in a single year, its creative output had outpaced its ability to measure what players actually did inside those worlds.
The challenge
When I joined, the only window into the games was Roblox's native creator dashboard: top-line KPIs, no usable history, nothing granular enough for real product analytics. Teams couldn't follow a cohort through onboarding, see where players dropped off, or tell whether a live-ops change had moved a metric. Producers had no confident read on what was working inside a live experience. No analyst could query game data, build a chart, or investigate an issue. And there was no data-backed narrative for the brand clients and the board funding these worlds, which also left an untapped chance to package analytics as a service line.
The hard part wasn't buying a tool. There was no data platform at all and no in-house data engineering to build one. Roblox doesn't hand you clean event streams; granular telemetry has to be instrumented in-game and shipped out with schemas built to scale. And it had to work across five games launching that year, on a short runway, for a team with no SQL background.
The approach
Map the data and the questions first
I started from the games and their users rather than the tooling, working through what Roblox could capture and defining each game's questions with producers and the analyst. The key insight: most needs were common across all five games — sessions, retention, funnels, economy, engagement — with a smaller custom set on top. That split shaped the architecture.
Architect for five games, not one
After a vendor evaluation I onboarded Google Cloud and built around a shared, schema-driven pipeline: capture a common event model in real time, serve the metrics every game needs, and leave room to federate per-game metrics later. Every decision favored reuse, so each launch would plug in rather than start fresh.
Ship an end-to-end MVP
I built the full pipeline end to end for one live game first — instrumentation through to a live-ops dashboard — then templatized it across the rest.
What I built
A complete production data stack on Google Cloud, fitted onto the studio's existing Roblox and Google environment:
- Structured telemetry from the Roblox game servers (Luau), shipped to a cloud endpoint (Cloud Run / Functions) and buffered through Pub/Sub for real-time capture.
- Dataflow pipelines transform events into BigQuery, with raw events landed in Cloud Storage as a replayable source of truth.
- A schema-driven event model and SQL transformations (Dataform) producing a shared library of live-ops metrics, plus hooks for per-game custom metrics.
- Looker dashboards for a no-code producer view, and a modeled warehouse the analyst could query with plain SQL.
- Access via Cloud IAM on the existing Workspace SSO, with backups and pipeline re-runs/backfills, so a bad day never meant lost data.
Because it sat on top of what the studio already had, there was nothing to rip out and replace.
Results & impact
Inside the build window the studio went from no analytics to a live platform processing 2M+ events a day and serving 70+ live-ops metrics across five game IPs — its first live-ops analytics ever. The payoff showed up fast. When a game's leaderboard broke and players lost their rankings, the team used the data to diagnose and restore them, the exact response the platform was built for. The team also used onboarding-funnel data to find drop-off points and improve the funnel. Producers got their metrics on a dashboard, the analyst got a reliable system to query, and the Head of Product could turn it into data-backed stories for clients and the board.
I led this end to end: design, build, and adoption. The studio was later acquired by a publicly traded media company.
Why it matters
The platform was for games, but the capability is industry-agnostic: install a governed, reliable data foundation where none exists — the precondition for putting AI to work on top of a business.