The intent behind Gaby is for him to handle the temporal and long-term memory of my main agent, Tom, like a database gatekeeper.
Gaby is a self-maintained data workflow that manages datasets before and after model fine-tuning, plus a small model that ranks data subsets, picks highlights from Tom's observations, and files them into the right storage.
The model
Gaby's ranking model is inspired by Google's PageRank and other Markov-chain models. I'm still experimenting with whether it can beat the usual KV methods for retaining caches during chat sessions, and how it plugs into Tom's decision-making.
The main network I'm exploring is self-/multi-head attention, pulling contextual vectors through transformation functions. So far the only conclusive finding is that these transformations project large text corpora into lower-dimensional spaces and cut compute costs a lot, which makes things like self-organising maps actually feasible.
Self note
The open problem is regularising the cost functions against the observations gathered from Tom's family members. Haven't cracked it.
On a more enthusiastic note, his cloud and API architecture has been much more trivial... and fun.
What he does for people
- Database connections: secured routes into a user's MongoDB, Supabase etc. for wrangling.
- Documentation & modelling: platform integrations plus a fine-tuned model that draws charts (Mermaid coder).
- Cleaning & processing: wrangling messy datasets, the boring Monday-morning fixes.
In production as DataBy AI. Why I'm building him: how has this not been automated?