Incidents from game KPIs · Lab

A clickable prototype on mock data: a KPI dashboard flags an anomaly, four agents look for the cause, and a last step writes a root-cause report with a confidence level; Barış Ülgen built it from our conversations with game studios.

What it is

FIG. 1From an alert to a report. The four agents work in parallel; only the weighing step writes the report, or marks the alert as a false alarm.
Source: cogniscope-kpi src/data/incidents.ts and src/data/types.ts, HEAD 3cfc02d

FIG. 2The five mock incidents and how sure the report is. Made-up scenarios written into the prototype, not real incidents. The last one is the case the system should not escalate: the data was late, the players were not gone.
Source: cogniscope-kpi src/data/incidents.ts (compositeConfidence), HEAD 3cfc02d

A game's numbers move every day: daily players, retention, revenue, session length, conversion. This prototype watches them and opens an incident when one falls outside its usual range. Four agents then look for the cause at the same time: one reads community posts, one internal tools such as deploy logs, one game telemetry, and one live events and patch notes. A last step weighs what they found by how well the sources agree, how closely the timing lines up and how much signal there is. It writes a root-cause report with a confidence level, the other explanations it considered and a timeline, and you can ask follow-up questions in a chat beside it. Everything runs on mock data.

What happened

The idea came out of our conversations with game studios. Barış Ülgen built the prototype; all 35 commits are his. My part was those conversations and showing the result: I recorded a 90-second walkthrough, from a new investigation to the finished report, for the team whose problem it answered. Sent to older leads later that month, it had five viewers and reopened some old threads.

What I learned

Detection alone was the crowded part. An April 2026 review of the market found many analytics tools that already flag a KPI drop, some of them free. What made people lean in was the why after the alert, so the investigation, not the dashboard, was the part worth building.