Stash

Problem

Most money tools assume a salary: one predictable amount, once a month. A student's money doesn't work like that. There's an allowance, a freelance job, a scholarship you're chasing, a side hustle, a savings goal you'd rather not touch, and all of it is tracked in your head. Budgeting apps want you to file that into forms. A general AI assistant will talk about your money and then forget the conversation. I wanted a tool that held the context instead of leaving it in your head.

Approach

Stash is conversation-first. You tell it what happened, "spent ₦2,000 on transport" or "got paid for the design gig," and it updates your balances, keeps what matters, and raises things when they become relevant, like a scholarship deadline getting close. Under the chat there is a real ledger, savings goals, scholarship tracking, and a memory that persists between sessions.

Key decisions

• Code owns the math. A language model will hand you a confident wrong number, which is not acceptable when it's your money, so the code calculates, validates and writes to the ledger. The model only interprets and explains, working from figures it is given.
• An early version logged phantom expenses: say "I have to buy a car" and it recorded a purchase. Now it separates money that actually moved from money that is owed, planned, or only being talked about.
• Savings goals are earmarks. Setting one marks intent; it doesn't pull the money out of your balance, so what you've set aside stays separate from what you've spent.
• Inference sits behind an OpenAI-compatible interface, so the model can be swapped without touching the agent. The encrypted ledger backs up to decentralized storage, which keeps ownership of the memory with the user.

Outcome

Live at heystash.app, and it reached the Round of 32 of the 0G Zero Cup out of 218 projects. The judges called it a strong concept with nice UX, well aligned with the theme, but it didn't advance. The demo ran on a hosted model. The on-chain compute I'd architected for was router-ready and never wired into anything the judges saw, so the docs described an integration the demo couldn't show. That's the part I'd do differently.