Inspiration
As a student, sticking to a budget is incredibly important, but the friction of opening an app and manually typing out every expense became tedious. Most of the time, I would simply forget to log my transactions and ultimately lose track of my finances altogether. I realized that if logging expenses felt less like accounting and more like simply talking to a friend, people would actually do it. That’s why we created Buzzy AI.
What it does
Buzzy AI is a voice-native financial assistant tailored for the Nepali language. Users simply tap a microphone button and speak their transactions naturally (e.g., "I spent 400 rupees on groceries"). Buzzy automatically extracts the relevant information—amount, category, merchant—and updates your dashboard. Beyond just logging data, it actively tracks financial goals, learns your spending patterns using an AI memory database, and even generates personalized weekly audio summaries spoken back to you in Nepali.
How we built it
We built the frontend as a React application using Vite and TypeScript, styled with Tailwind CSS for a premium, snappy user interface.
- Backend: We used Supabase for our real-time database and authentication.
- AI Processing: We leveraged Google Gemini as our core brain to take messy, raw voice transcripts and parse them into structured JSON data (categorizations, recurring trends, goals).
- Voice Integration: We utilized ElevenLabs for both native Nepali speech-to-text transcription and dynamic text-to-speech generation for our weekly summary feature.
Challenges we ran into
Handling native Nepali speech recognition and ensuring the AI correctly mapped local slang or casual speech to strict financial categories was tricky. It also took extensive prompting and iteration to get Gemini to consistently return perfectly structured JSON that the frontend could rely on without breaking the app.
Accomplishments that we're proud of
We're incredibly proud of practically eliminating the "friction of logging." The entire budgeting experience requires zero typing. We also managed to implement a totally seamless "Weekly Summary" architectural workflow that queries 7 days of raw data, synthesizes it using Gemini, and reads it back natively via ElevenLabs in just a matter of seconds.
What we learned
We learned that LLMs are vastly more powerful when used as invisible structured parsers rather than just traditional conversational chatbots. We also learned how to effectively chain different AI microservices (Transcription -> LLM Extraction -> Database insertion) without making the user wait through long loading screens.
What's next for Buzzy AI
Next, we want to integrate optical character recognition (OCR) so users can just snap pictures of receipts. We also plan to implement completely proactive voice notifications (e.g., Buzzy speaking when you hit 80% of your weekly food budget) and integrate direct banking APIs so it can cross-verify your voice logs with actual bank activity.
Log in or sign up for Devpost to join the conversation.