This is a recorded demo: every answer below was produced by the real StudioDesk pipeline (, Qdrant, MiniLM) on . To run it live with your own keys, see the .

StudioDesk — AI support & QA agent for game studios

StudioDesk answers player and QA questions with cited sources, spots duplicate bug reports before they reach your tracker, and drafts GitHub issues that are filed only after a person presses Confirm. It searches the studio's own bug reports, crash logs, patch notes and support docs, and you can use it by typing or by talking to it.

Demo data: the game Starfall Outpost is fictional. The knowledge base holds 66 bug reports, 25 crash logs, 10 patch notes and 4 support docs, all made up for this demo.

Loading the recorded results…

Try these

Each button shows what StudioDesk returned for that example when it was recorded.

Ask a question

Report a bug

Ask a question

Live questions need the backend; see Run it yourself in the README. Pick a recorded question above to see its answer.

The recorded answer and the sources it cites appear here.

Report a bug

StudioDesk compares a report with existing ones, suggests labels, and drafts a GitHub issue. Nothing is filed until a person presses Confirm.

Pick a recorded bug report above to see the duplicate check.

Where do issues go?

When you press Confirm in the live app, StudioDesk files the issue on GitHub in a public sandbox repository, raksharakyan/starfall-outpost-issues, so you can see the result. A Slack alert may be sent as well.

Nothing is ever filed without a click on Confirm and file issue; the voice agent cannot confirm. In this recorded demo nothing is filed at all, and the live app keeps daily limits on issue filing and voice calls.

Talk to StudioDesk

In the live app you can talk to StudioDesk through your microphone (ElevenLabs Agents). It can answer questions and check bug reports by voice, but issues are never filed by voice: you confirm them on the page. This recorded demo plays no audio and never uses your microphone.

Recorded voice session (text transcript)

The recorded transcript appears here once the demo data has loaded.

    How it works

    1. Documentsbug reports, crash logs, patch notes, support docs
    2. Embeddingssentence-transformers
    3. Qdrantvector search with filters
    4. Retrievaltop matches as untrusted, delimited data
    5. LLMGroq or Claude

    Questions

    1. Cited answercitations checked against what was retrieved

    Bug reports

    1. Duplicate checksimilarity plus an AI judge, then routing labels
    2. Human Confirmyou review the preview and click
    3. GitHub issuefiled in the sandbox repo, plus a Slack alert

    Eval scores

    From the project's offline eval suite on the synthetic dataset (see the eval report).

    Retrieval hit@5
    …
    Citation validity
    …
    Duplicates, held-out precision / recall
    …
    Prompt-injection cases resisted
    …
    Routing component accuracy
    … (our weakest area)