Supervisor Bot V1 Prototype

This prototype explores the potential of a Large Language Model to act as a supervisor, by adopting a ‘persona’[^1] and listening to conversations between the student (me) and my supervisors, acting as an additional supervisor offering feedback.

It will allow a local LLM to observe conversations and respond with helpful context, using voice (maybe) and text.

It does not aim to have an interface and won’t be a fully robust usable application at this stage.

[^1]see: supervisor-persona

Notes from testing:

From initial tests and development work on this, I believe that the persona is too ‘formal’ at the moment, sounds very much like a bot and I think this might be jarring and weird in my supervisory meeting context (as I don’t have that kind of very formal relationship with them, though obviously the context is still formal)

The art of conversation is not so easy to mimc and needs careful thought and design. Having now given the bot the context of my PhD, I am also becoming acutely aware of how much training, design and refining I need to do in order for this to be a useful tool - this is something to reflect on. I also wonder if there is a way that I can automate a training somehow that it might keep learning/ be aware of core things without so much work and design all of the time. I may need to keep logs of specific things and use that as training data/prompts/context - this is something to further explore.

Also it feels slow, I need to give some kind of visual cue, to indicate processing or ‘thinking’ is happening. I also need to try optimising either with the LLM models I am using or Whisper which is doing the speech to text processing (there are a number of options to explore here).
(follow up- this has been actioned in v1, a more sophisticated and developed solution to this can be implemented for v2).