ForeSee

Predict the tab. Protect the vibe.

Comment

Inspiration

Budgeting apps treat money as isolated numbers, but real spending is social. You don't just spend $55 on dinner. You spend it with your coworkers after a long week, or on a date night with your partner, or splitting nachos with your friends at trivia night. Every dollar has a social story behind it.

We asked ourselves: what if your finances were organized the way your life actually is? Not by categories in a spreadsheet, but by the people you spend time and money with.

That's ForeSee. A finance platform that organizes your spending into social boards, one for work, one for friends, one for your partner, one for you, each connected to your calendars so it knows what's coming before your wallet does.

What it does

ForeSee is a personal finance platform that maps your spending to the social contexts of your life, transforming the unpredictable nature of a multi-calendar lifestyle into actionable financial insights.

Social Boards. Your finances are split into boards: Work, Friends, Spouse, and Personal. Each board is a living feed of events like dinners, outings, trips, and subscriptions with real transactions linked underneath. You see who you spend with, where, and how much, not just a list of charges. Board members share visibility and can engage with each other's spending patterns.

Calendar-Powered Predictions. Connect your calendars, whether Google Calendar, Outlook, Apple Calendar, or any iCal-compatible source, and ForeSee automatically classifies incoming events into the right board using the Gemini API. It then predicts how much each event will cost based on your spending history at similar events, before you even walk in the door. This turns unpredictable multi-calendar chaos into a clear financial forecast.

AI Financial Coach with Voice. Each board has a voice-enabled AI coach powered by ElevenLabs that streams real-time, data-backed savings tips. Using sentence-level text-to-speech synthesis running in parallel with the language model, the coach starts speaking almost instantly while still generating its response. It knows your patterns and gives specific, actionable advice grounded in your actual data, like telling you how much you could save by bringing lunch twice a week based on your real work dining history.

Community-Driven Savings Challenges. When a new event arrives, the system generates savings suggestions powered by Gemini with deep reasoning. Each suggestion comes with a specific dollar amount, a confidence score, and a detailed explanation. Board members vote on suggestions together, turning saving money into a social, gamified experience. Completed challenges trigger confetti celebrations to reinforce positive behavioral change.

How we built it

Frontend. React with ShadCN components. Three pages: a home page with a board card grid, a board detail view with an event feed and transaction sheet, and an AI powered Coach with emotional bubble response. Skeleton loading states and micro-interactions throughout.

Backend. Express and TypeScript with Supabase (PostgreSQL) as the data layer. Seven tables: users, boards, user_boards, events, transactions, suggestions, and a get_suggestions job queue.

AI Pipeline. Calendar events flow in through automated webhooks that support multiple calendar providers including Google Calendar, Outlook, and iCal feeds. Each event passes through a four-step pipeline: first, Google Gemini 3 Flash classifies the event into the correct social board by analyzing attendees and context, and categorizes the event type. The classified event is stored in Supabase. Then Gemini 3 Flash analyzes all historical same-type events and their linked transactions to predict cost and confidence. Finally, a background job queues suggestion generation using OpenAI.

Voice Coach. The coach chat endpoint uses Server-Sent Events to stream text from GPT-4o-mini and audio from ElevenLabs simultaneously. A sentence splitter buffers tokens and sends each completed sentence to ElevenLabs eleven_turbo_v2_5 for real-time text-to-speech. Audio chunks stream as base64 alongside text tokens, so the user hears the coach speaking the first sentence while the model is still generating the rest.

Suggestion Engine. OpenAI o3 with structured JSON output and high reasoning effort generates detailed savings suggestions. Each suggestion includes a specific dollar savings estimate, a confidence score, and a full explanation. Suggestions are auto-seeded with social votes based on board member count to drive peer engagement.

Challenges we ran into

Getting the calendar webhook pipeline reliable was the biggest hurdle. Events come in all shapes. Some have locations, some don't, some have vague titles with no other context. Tuning the Gemini classification prompt to handle edge cases without over-classifying took several iterations.

Streaming audio and text simultaneously over SSE required careful sentence boundary detection. We built a lightweight sentence splitter that buffers until it is confident a sentence is complete, then fires off the ElevenLabs TTS request immediately while continuing to receive tokens from the language model.

The o3 suggestion engine needed strict JSON schemas to produce consistently structured output. Getting the schema right with nested confidence scores, savings estimates, and explanations took trial and error with the structured output API.

Accomplishments that we're proud of

The four-step AI pipeline feels seamless. A calendar event appears and within seconds it is classified, stored, predicted, and queued for suggestions, all automatically. Users never touch it.

The voice coach experience is genuinely surprising. You click talk to coach and it starts speaking almost instantly, giving you specific dollar-amount advice based on your actual data. It does not feel like a chatbot. It feels like a financial advisor who has been watching your spending.

The social dynamics layer makes saving money feel collaborative instead of isolating. When your board members vote on a suggestion and you complete it together, the confetti hits different.

The design system is clean and intentional. Skeleton loading states, consistent spacing and graphs that don't just display data but communicate every event and action. Everything feels considered and polished.

What we learned

Spending data only becomes meaningful when it has social context. $46 at Cactus Club means nothing. $46 at Cactus Club with your work team, the third time this month, tells a story and motivates change.

Multi-model AI architectures work well when each model plays to its strengths: Gemini for speed in classification, o3 for depth in reasoning, GPT-4o-mini for fluency in conversation, and ElevenLabs for natural voice synthesis.

Real behavioural change in personal finance comes from peer engagement and social accountability, not just showing people numbers on a screen.

What's next for ForeSee

Native integrations with Outlook, Apple Calendar, and other calendar providers for seamless multi-platform support.

Direct banking API integration through Open Banking to auto-link transactions to predicted events.

Shared board notifications so members know when someone completes a savings challenge.

Monthly wrapped summaries per board showing spending trends, biggest saves, and social spending patterns.

Integration with RBC's digital suite to enhance user retention, budgeting capabilities, and to enable users to see the future forecast of their projected spending before it happens.

Built With

  • elevenlabs
  • express.js
  • google-calendar-api
  • google-gemini-3-flash
  • next.js
  • openai-gpt-4o-mini
  • openai-o3
  • shadcn
  • supabase-(postgresql)
  • tailwindcss
  • typescript
Share this project:

Updates