SmartFinBot

The brief

An academic research project: a personal finance system built around an AI chatbot. The assistant answers money questions in plain language, categorises spending automatically and turns the running balance into budgeting and investment suggestions the user can act on, without manual ledger entry.

Problem

The team started from a research question: can a personal finance app stay useful when it only gets data when someone remembers to enter it? Manual ledgers fail for the same reason gym memberships do. Every budgeting app assumes the user will do the bookkeeping, and the moment they stop, the app becomes a reminder of failure rather than a tool.

My role

Four-person academic team with a shared research write-up. I owned the mobile experience: the interface, the flows, the animation, and the on-device state that keeps screens readable when the network is unavailable.

Process

• Defined the questions a user actually asks about money before designing a single screen
• Designed around a chatbot as the primary interface, with the ledger as a view of what the chat had already done
• Built the categorisation rules for incoming transactions so spending sorted itself instead of asking the user to tag every entry
• Kept reads local-first so the app showed the last known balance instead of an error state when offline
• Tested the flows with real users, then rewrote the parts of the interface the tests exposed
• Kept the chatbot conversation history out of the balance calculation entirely, so a misread question could never move a number the user relies on

Result

A working prototype plus a research write-up documenting the parsing and suggestion rules and where the approach breaks down. There is no live URL to publish because the project was academic, and I am not going to quote a test score I cannot show. The honest result is the architectural one: a chat-first finance app where the user never has to act as the data-entry clerk. What the research also showed is where the idea strains: categorisation is only as good as the transaction descriptions it receives, and Nepali bank exports are far less consistent than the sample data we started with. Any production version of this would need a correction path where the user fixes one category and the rules learn from it.

What it had to do

• AI chatbot that explains every transaction in plain language
• Automated financial tracking with chat-first budgeting suggestions
• Research write-up covering parsing and suggestion rules

Stack

• Flutter
• Dart
• Firebase
• AI chatbot
• NLP

Team credits

• Pushkar JhaKnow Him
• Niranjan BhattaKnow Him
• Sachin SubediKnow Him
• Nabin Ghimire (me)