Lampwright 中文
Back to the bench

Lamp I · In progress Checked

Generalprobe

A rehearsal room for trainee and career-changer teachers. Teach a short lesson to a class of twelve synthetic pupils, go through it afterwards, and rewind to any moment to try another response.

Generalprobe offline demo: a class of synthetic pupils at their desks during a lesson, with the class feed beside them
Screenshot · Generalprobe offline demo, a fractions lesson for grade 7 ·
Room
Trainee and career-changer teachers in Germany
Made of
A single HTML file that runs in the browser, with pre-written pupil lines and keyword rules. An optional local server lets DeepSeek write pupil questions and judge the teacher’s answers; each call has a 4-second timeout, and on a timeout, an error or invalid output the keyword rules decide.
Made for
2nd Accenture AI × Education Hackathon 2026, Germany · Best Idea award
Made by
Hongyu Zhou with two fellow students at TU Berlin
State
Hackathon prototype, September 2026. Local AI pupils added October 2026. Phase I abstract submitted to the Tools Competition 2027 on 5 October 2026. Not a product.
Data
No real pupil data. The offline demo needs no server and records nothing; only the optional voice commands use Chrome’s online speech recognition.
Links
None. The repository is private and the demo is not hosted.
Entry checked

What it does

The teacher picks an activity and talks to the class. Each of the twelve pupils has a temperament, an attainment level and a voice of its own, and the class drifts off task, chats and raises hands without being told to.

After the lesson, a review page goes through what happened. The teacher can rewind to any moment, try another response and see how the class reacts this time.

A rehearsal can be exported as an event log, and replaying it with the same seed and settings reproduces the session. In the local AI mode, every line the model writes for a pupil gets a score from 1 to 5 for how well it fits that pupil.

Assembly drawing

Drawing: Generalprobe taken apart into four numbered layers
  1. 01 The board: slide titles read from a .pptx file in the browser
  2. 02 Twelve synthetic pupils, each with a temperament, an attainment level and a voice
  3. 03 The judge: keyword rules, or an optional local DeepSeek with a 4-second timeout per call
  4. 04 The event log: the same seed and settings replay the same session

A drawing of how Generalprobe is put together, made for this site. It is not a screenshot.

Assembly list

Installed

  • Twelve pupil personas and the rehearsal event log, each with a JSON Schema
  • Export of a rehearsal as an event log, and exact replay from it
  • Optional local AI pupils: model-written questions, model judging and a persona score for every model-written line
  • Slide upload: titles from a .pptx file are read in the browser and shown on the board

Not installed yet

  • A hosted version
  • Model-written speech for every pupil line. In the local AI mode, the other pupil lines are still pre-written.

Where to poke holes

  • The offline demo judges answers by keywords, not by understanding them.
  • When a lesson is rewound, the alternative answer is judged by the keyword rules, not by the model.
  • The first user test had five teachers. There has been no classroom pilot yet.

Backstage

Generalprobe started at the 2nd Accenture AI × Education Hackathon 2026 in Germany, where it won the Best Idea award. The local AI mode was added in October 2026, and a Phase I abstract went to the Tools Competition 2027 on 5 October 2026. The pupils are synthetic: no real children take part.

Back to the bench