What I've Been Building

LudoSense is a game analytics platform I built to capture, process, and visualise player telemetry from my thesis game, The Undoing of Ashvale. The system spans the full vertical stack: a GDScript SDK plugin that instruments the game client with event batching and retry-on-failure, a Python backend (FastAPI, Celery, Redis) that ingests and processes those events, a PostgreSQL database, and a React dashboard, all originally hosted on AWS.

The data pipeline follows a medallion architecture: raw events land in bronze tables exactly as the game client sends them, scheduled Celery tasks clean and normalise them into silver tables, and gold tables aggregate the results into hourly, daily, weekly, or monthly metrics that power the dashboard. The gold layer is fully configurable in that developers can define what their events mean (using counters, funnels, duration distributions, segmented breakdowns), and the pipeline computes the results automatically. Adding a new metric means writing a definition, not changing the ingestion path. This allows developers to build their own metrics mapped to however they've sent their raw events to the bronze tables.

The AWS infrastructure has since been taken down to avoid idle costs, but everything you see below is the real output of that pipeline, served as static JSON for the purposes of a portfolio piece. The data covers 20 synthetic players across two game versions (v0.1.0 before a fabricated performance patch and v0.2.0 after) along with 2 real thesis players, which demonstrates the full analytics loop: instrument a change, measure the result, and see if there is an improvement. The synthetic data was created based on the actual thesis play sessions and was made in a way that best showcases what LudoSense can do. As this is an ongoing project, the GitHub repository remains private, but I will be creating a specific public repo shortly for anyone interested in how the metrics are actually calculated. Keep an eye out for the updated link to the repo!

The overview below shows four key metrics at a glance, each represented as a card with a sparkline trend. Click any card to expand it and see the full chart: a line chart for counters, a stacked bar for breakdowns, a funnel for quest completion, or a duration chart for scene load times. The back button returns to the grid.

The version comparison is the centrepiece of the data story. The synthetic data contains two versions: v0.1.0 as the initial release, and v0.2.0 with a targeted performance patch. The charts below show scene load times side by side, highlighting the average, min, and max time per scene transition, so the improvement is immediately visible. Look at each scene to see how the patch affected each one.

Aggregate metrics tell you what's happening across the game, but per-player views help showcase who's doing what. The table below shows every player in the dataset with their session count, total playtime, quest progression, and retention status. Click any row to see which quests they completed and their activity timeline. Filter between thesis players (real dissertation playtesters) and synthetic players (generated data).

When aggregate metrics flag something interesting, the next step is drilling into individual sessions. The replay view lets you browse all recorded sessions, filter by version, and select one to see its complete event timeline. Every event is expandable to inspect the full payload, scene transitions, dialogue sequences, quest progress, and harvesting actions, all in chronological order.

The game client emits 59 distinct event types, from scene transitions and dialogue interactions to quest state changes and resource harvesting. The catalogue below shows every event type with its total count and an example payload. Search to find specific events or browse to see the breadth of instrumentation.