PRAXIS
Hold a spoken office-hours turn with a simulated CS student.
loading stack…
⚙ Settings
Session settings
Text only (no voice) — generate the reply as text, skip speech synthesis
Show student memory (instructor view) — reveal the split memory: the distilled knowledge base the student carries between turns
Student persona
loading personas…
✨ Generate student
Clear history
📋 Debrief
Pick a student from the dropdown, or
Generate
a fresh one whose misconception is hidden — you diagnose it by asking, the way a real TA does.
Code
the student
brought in
Hold to talk
Connecting…
Send
Conversation
No turns yet. Press and hold the button, speak, then release.
Latency (ms from end of your audio)
STT (transcript ready)
—
LLM time-to-first-token
—
LLM reply complete
—
TTS time-to-first-audio
—
End-to-end to first audio
—
Believability floor
—
Student memory —
—
instructor view
—
Knowledge base
durable distilled facts — older turns, compressed
No facts yet. The knowledge base stays empty until the conversation grows past the verbatim window (the recent turns, highlighted above, are carried word-for-word); older turns are then distilled into facts.
Session debrief —
—
coaching
—
What stood out
Try next time
Log