FinanceAI — Personal Finance Assistant

Section 01 · Overview

FinanceAI lets users talk to their money. Ask natural-language questions over your spending data, upload receipts for automatic OCR and categorisation, set budgets, detect anomalies, and get cut-back suggestions — all with strict per-user isolation via Supabase row-level security. The intent router sends 80%+ of queries to Claude Haiku and reserves Sonnet for vision and complex agent tasks — 5× cost savings at scale.

Section 02 · Flow

User signs in with Clerk and gets isolated DB access

Imports transactions via CSV or uploads a receipt

Asks questions in natural language or picks a suggested chip

Intent router classifies: simple_query → Haiku, receipt/lookup → Sonnet

Pre-computed monthly aggregates feed a compact LLM context (~1,500 tokens)

Response renders with charts, alerts and inline action chips

Section 03 · Modules

01

Balance card, spending donut, recent transactions, and a one-tap 'Ask assistant' prompt.

02

Chat UI with suggested questions, receipt upload, streaming-ready answers, and user memory.

03

Paginated, colour-coded by category, with CSV import and deduplication.

04

Set via chat or UI. Progress bars turn red when over budget. Alerts fire at 80% and 100%.

05

Stacked bar chart by category across 12 months, driven by pre-computed aggregates.

06

Z-score (2σ) anomaly detection vs 12-month history, plus recurring-payment pattern matching.