inspired by a suggestion made by someone from socratica, i decided to add a chat interface to my contact me section. i actually wired it up to openrouter's free model router. due to limited api rates, i don't think it'll be usable if more than one person is trying it out on my site at once but i'm on a student budget so...
i had to spend a bit of time thinking in the shoes of a malicious user, so i set up things like rate limiting, basic guardrails against prompt injection attempts, etc.. i also tried getting cindygpt to sound me. it ended up becoming a more cynical, judgemental version of me that's kind of endearing imo. but if you've talked to me in real life, i'm sure you know that i'm way more bubbly.
anyways, onto my project statuses. my focus is still on agentic ai systems and i'm having lots of fun with it.
building an mcp
this is now done! i created an mcp for querying the ontario sunshine list. i just haven't shipped it yet because i started working on the project below. status: complete
epsilon
ok so this idea replaced my self-improving agent eval framework idea. the motivation is that i have yet to find an llm that explains math well. it either treats you like you're five years old or that you've got a phd. and sometimes it's just plain wrong. the general idea is that i'm fine-tuning an open source llm (one of the qwen ones) using llm-as-a-judge as well as human input to get better at explaining + less hallucinations. still in progress but the rag pipeline is completed. status: in progress
others
fixing my sleep schedule, writing and posting more in public, making more friends at socratica and wat.ai, catching up with lectures, and learning electric guitar.