The problem it solves
AI can now do a good deal of a student's coursework. That
leaves an instructor with two bad options: ban it and be ignored, or allow it and be
unable to tell what the student actually did.
How does vela do it?
Vela agent limits are enforced operationally, by what the
agent is able to do. When a unit is set to guidance-only, the agent has no tool that
writes the students code. And every interaction is recorded. The work arrives with
its history attached.
That record makes AI-assisted work inspectable and something
an instructor can assess. And it lets the student show how they got there.
Testing access
Access is by invitation while the testing period runs. To ask
for access, or to report anything, email Falk Herwig —
fherwig@uvic.ca
What to expect right now
This is a pilot, not a service.
- The AI model is not always available. Three sources sit behind the
model menu and each has its own limits — see Token source for vela below.
- Not for coursework yet. Volunteer testing only. Nothing here is
approved for graded work.
- For testing only. Work on this server is not backed up and may be
deleted as part of continuous development.
- Data is recorded during testing period. All interactions are
recorded and will be used as feedback to improve vela.
- Sign-in is via GitHub user name. Google, ORCID and UVic NetLink are
not connected yet.
Token source for vela
Every answer in a vela is generated by a language model running
somewhere, and where that is affects speed, availability and where your words go. The model
menu names its source on every turn. There are three today:
- Self-hosted, on the project's own allocation — the default.
Qwen3.6-35B served from a single GPU on UVic's Arbutus cloud, run by us. This is what the
menu opens on. It is also available to you directly, outside vela, if you want to point your
own editor or scripts at the same model:
tokens.velella.ca.
- University research computing, in scheduled windows. The same open-weight
model on shared national research infrastructure. Larger capacity while a window is open,
and simply unavailable when it is not.
- A Canadian metered provider (Nebula Block). A small prepaid token budget,
used as a fallback. Note that the menu can fall through to it automatically if the other two
are unavailable mid-session — the source shown in the menu always tells you which one
answered.
None of this is a dependable service.
All three sources are exploratory: the hardware is borrowed or shared, the budget is small, and
any of them can change or stop. We are finding out what a self-hosted open-weight model can do
for teaching and research — that is the point of the exercise, and it is the reason nothing
here carries an availability promise. Do not build anything that depends on it staying up.
See it working
A short walkthrough is here:
video