Dungeon AnalystSabermetrics for one D&D campaign

From the table to the board

How a night of D&D on Discord turns into the numbers on this site.

The pipeline

Step 1: Record

A Discord bot records each session with one audio track per player, so every voice is already separated.

Step 2: Transcribe

Whisper turns each track into text on a laptop, not in the cloud.

Step 3: Clean and merge

Each speaker is matched to their character, and the tracks are merged into one transcript in the order things were said.

Step 4: Extract with six Claude agents

Each agent reads the transcript for one job. Every record it returns carries a confidence score and the quote it came from. Anything under 85% confidence is flagged for a person to check.

Roll ExtractorEvent ExtractorNPC TrackerSession SummarySession ReportPsychology Analyzer, never published

Step 5: Store and review

A Postgres database is the system of record. Corrections are saved as files that can be replayed, every change is logged, and each number traces back to the prompt version that produced it (assumed for sessions 9 to 11, recorded from session 12).

Step 6: Publish an anonymized release

A release strips real names and handles, then scans for them again before anything leaves the laptop. Survey answers only appear as averages of three or more people. Each release is new and never edited, so going back is one switch.

Step 7: Serve

This site runs on Cloudflare's edge and reads the release from a cloud Postgres database. The group's private site sits behind a one-time email code.

Keeping it honest

  • Every roll shows whether a person has checked it yet.
  • Rates come with a likely range, and small samples aren't ranked.
  • Since release 2, damage, initiative and other rolls, which have no pass or fail, are left out of success rates.
  • When the extraction method changes, the affected sessions are marked.
  • The database can be rebuilt from its source files and compared row by row before any change goes live.

Built with

Claude APIWhisperPythonPostgreSQLSupabase (local)NeonNext.jsCloudflare WorkersCloudflare AccessCloudflare R2 backups