Conversational UX / AI systems / Fintech
I designed and built Nomi around a simple question: can managing money start with a text, instead of another app?
“Where did my paycheck go?” is a real question. A dashboard isn’t always the right answer.
Nomi started as WithBFF, a private-beta financial assistant. I designed the product experience and built its stack: React, Python, Supabase, Plaid, Twilio, and an LLM orchestration layer.
SMS makes the details matter. A confusing answer cannot hide behind a clean dashboard. I have to get the words, timing, and financial logic right together.
The first onboarding flow generated a budget immediately after bank connection. Beta feedback showed a mismatch: a budget could be technically valid without feeling like something the user had agreed to. I redesigned the experience around a seven-day progression, from connection to confirmation to collaborative budgeting.
The tradeoff More steps before a complete budget, in exchange for explicit participation in what the product tracks.
A rigid intent classifier struggled with compound questions and everyday language. I moved toward LLM orchestration: the model selects financial tools, retrieves the relevant data, and composes a response around the user’s question.
The tradeoff More orchestration complexity and model cost. Tool selection and financial accuracy still need careful evaluation.
A transfer to savings should not read like overspending. I separated spending, saving, debt reduction, income, and internal transfers so the product could describe what actually happened to the money.
The tradeoff A more detailed classification layer underneath a simpler explanation to the user.
This interaction reconstruction connects the product decisions above with the uncertainty and correction patterns I’m exploring. It demonstrates design intent, not a verified live capability or a research result.
What this makes inspectable: response structure, classification boundaries, and the person’s control over a correction. Reliability and usability still need evaluation.
Bring relevant spending context into the reply instead of sending someone to a dashboard.
A useful answer explains which spending or budget information it is based on.
When a user disagrees, ask whether the issue is the amount, date, or category before retrying.
The useful feedback came after I thought the design was done.
Onboarding, tool orchestration, and transaction classification are connected design decisions. A clearer message depends on the system underneath it being able to support that promise.
The next work is making the experience more reliable over time: evaluating compound questions, exposing uncertainty in incomplete data, and testing whether the onboarding sequence earns understanding without creating unnecessary delay.
Nomi is an independently founded product in private beta. This case study focuses on the product decisions, rather than measured adoption outcomes.