Загрузка...

ClearMoney - On-Device AI Finance App Powered by Gemma 4

A privacy-first personal finance app that runs fine-tuned Gemma 4 entirely on your phone — no cloud, no server, no data ever leaves your dev

ClearMoney is a privacy-first personal finance app built for iOS using React Native and Expo SDK 54. It imports your bank statements — PDF or a photo of a paper statement — and gives you a complete picture of your finances: spending breakdowns, budget tracking, AI-generated insights, month-over-month comparisons, and a conversational AI advisor. Every single operation runs on-device. There is no server, no cloud, no account, and no API call. Your data never leaves your phone.

The AI Stack
Fine-tuning with Unsloth
We fine-tuned Gemma 4 on real-world financial transaction data using Unsloth — a framework that cuts training time in half and reduces GPU memory consumption by 60% compared to standard fine-tuning. The fine-tuned model learns to categorise transactions (Food, Transport, Entertainment, Subscriptions etc.), detect recurring charges, and understand merchant names across all four supported banks. Weights are published publicly on Hugging Face for anyone to audit or build on.

GGUF Quantisation & llama.cpp Inference
After fine-tuning, the model was quantised to GGUF format — the compressed binary standard used by llama.cpp. The resulting file is 4.7 GB and runs real-time inference on an iPhone with no internet connection. On first launch the model is downloaded once to local storage and used permanently from that point forward.

Cactus-Style Intelligent Routing
We built a two-model routing pipeline. A fast rule-based engine (tryRuleCategorize) handles known merchants with zero latency. Unknown or ambiguous merchants escalate to Gemma 4 (batchCategorize) for deep contextual classification. This means the vast majority of transactions are categorised instantly, and Gemma 4 only runs on the subset that actually needs it — maximising speed and battery efficiency.

PDF.js + Tesseract.js Parsing
Digital PDF statements are parsed using PDF.js running inside a sandboxed WebView. Paper statements photographed on camera are processed through Tesseract.js OCR, also in a WebView sandbox. Both run entirely on-device.

Key Features
Dashboard

Financial Health Score (live ring, 0–100, calculated from SQLite data)
Spent This Period with month-over-month trend arrow
Spending Card: remaining budget, daily rate pill, colour-coded progress bar (blue → amber → red)
Quick Actions: Search, Transactions, Subscriptions, Budgets, Calendar, Ask AI
Bill Predictor forecasting upcoming charges
Six-month Spending Trend bar chart
Category Donut chart with segmented colour bar
Budget progress bars and Overview summary table
Filter by bank and by month — all computed locally
Insights

Snapshot strip: Savings Rate, Daily Average, Subscription total
Smart Alerts: spending spikes, low savings rate, subscription bloat, weekend vs weekday anomalies
Monthly Trend chart with month-over-month badge
Spending Breakdown with per-category percentage bars
Habits & Patterns: weekday avg, weekend avg, busiest day, biggest transaction
Recurring Charges list
Income vs Spending net bar
Month picker to analyse any historical month instantly
AI Chat

Custom-named AI advisor powered by fine-tuned Gemma 4
Streaming token-by-token responses with animated cursor
Starter question chips and contextual follow-up suggestions after every answer
Dollar amounts highlighted in responses
All inference runs locally — the AI sees your data because it IS your data
Compare Mode
Side-by-side comparison of any two months with per-category delta badges — green where spending improved, red where it increased.

Privacy & Security

Face ID / Touch ID gate via expo-local-authentication, re-locks after 60 seconds in background
All data stored in expo-sqlite on-device
No account, no login, no tracking, no analytics
Profile page confirms: "All data stays on your device. Zero cloud. Zero tracking."
Factory Reset option permanently deletes all data and the model file
Dark Mode & Haptics

Full dark/light theme toggle
expo-haptics fires tactile feedback on meaningful actions: transaction taps, budget creation, budget warnings (Heavy impact at 90% threshold), successful uploads

Видео ClearMoney - On-Device AI Finance App Powered by Gemma 4 канала Sanjay K
Яндекс.Метрика
Все заметки Новая заметка Страницу в заметки
Страницу в закладки Мои закладки
На информационно-развлекательном портале SALDA.WS применяются cookie-файлы. Нажимая кнопку Принять, вы подтверждаете свое согласие на их использование.
О CookiesНапомнить позжеПринять