HackSouthWest Hackathon - Liminal Payroll
2nd place project, combining AI agents with business logic for a crypto banking app
So, me and some friends competed in a hackathon in exeter uni and we combined a few different bits of tech for our project.
The main task was building something to do with AI agents for a blockchain banking startup called Liminal. They gave us an sdk, an api key (that they rotated 3 or so times since it kept getting pushed) and 24 hours.
We took a while to come up with something, just playing around with the sdk and trying to write our own skills for the agent.
We settled on a business logic platform, so employees, payroll and cashflow could be managed and automated. This could be done through buttons to REST APIs, or by asking the chatbot.
The machine learning side of it went through about 3 rewrites, initially being done in python and it ended up being all written in go.
My contribution
My biggest contribution was the employee database, its CRUD functions, and making those funcs accessible to the chatbot.
At first I wasn’t actually that convinced with what I was building. I thought it was a bit of a waste of electricity to get a LLM to perform what just some buttons. When it was all built it was surprisingly useful.
So, since its an LLM, the chatbot can understand natural language and then act on it. This means it is actually able to string together these tools we made for it.
While we had functions to get employees by department and change salaries, there wasn’t a button for changing a department’s employees wages. The chatbot when prompted “reduce engineering’s wage by 10%” called multiple of these tools one after another, almost like a unix pipe.
(I found this when I asked it to drop the table, not realising that this wasn’t a skill, it called all ids, and then remove them all).
I’m still undecided on if its good to implement an LLM, we had claude sonnet 4 which was definitely overkill, when buttons / automated systems have worked fine for this for decades at this point. Although, models have been getting more performant and cheaper to run, so maybe it could be worth it.