Hermes: The Offline, Privacy-First Financial Co-Pilot

Hermes is a radical reimagining of personal finance software: a fully offline, data-sovereign app that treats your financial data as inviolable. It uses on-device machine learning to categorize transactions, forecast income and expenses, and suggest savings goals—all without a cloud server, account, or telemetry. You import bank statements via manual entry or local file imports (CSV/OFX), and all data is encrypted and stored in IndexedDB. The app is a progressive web app (PWA) built with React, using TensorFlow Lite for inference. No data ever leaves your device; even the machine learning models are trained locally from anonymized patterns. Hermes is for people who want intelligent financial guidance without sacrificing privacy.

Imagine opening an app that knows your spending habits, warns you before a bill might overdraft, and suggests a weekly savings target—all without knowing your name, email, or bank credentials. That's Hermes. It's a what-if for a world where financial tools empower without surveillance.

Every popular personal finance app today (Mint, YNAB, Personal Capital) collects user data, often selling it or using it to target products. They require online accounts and bank API integrations, creating a massive surface for data breaches and surveillance. Privacy-conscious users have no alternative that provides intelligent insights without giving up their data. Hermes scratches that itch: a fully local, zero-telemetry financial assistant that never phones home.

• Provide comprehensive budgeting, cash flow forecasting, and savings optimization without any data leaving the device.
• Support manual data entry and local file imports (CSV, OFX) from banks.
• Use on-device machine learning to categorize transactions and predict future income/expenses with >90% accuracy on typical data.
• Achieve <100 MB storage footprint and run smoothly on mid-range phones.
• Fully open-source under AGPL to allow auditing and self-hosting.

• No cloud sync or backup (user manages own backups via encrypted file export).
• No bank API integration (to avoid data leakage).
• No advertisements or data monetization.
• No social features, sharing, or gamification.
• No support for credit scores, loan comparisons, or external financial products.

• Frontend: React (PWA) with IndexedDB for local storage.
• Machine learning: TensorFlow Lite for on-device models (small neural nets for categorization and forecasting).
• Encryption: Web Crypto API for AES-GCM encryption at rest.
• Build tools: Webpack, Babel, Workbox for service worker caching.
• Testing: Jest for unit tests, Cypress for end-to-end.

Hermes is a single-page application that runs entirely in the browser, with no backend server. All data is stored in an IndexedDB database, encrypted with a key derived from a user-chosen passphrase (never stored). When the user imports or enters transactions, they are automatically categorized by a locally-trained TensorFlow Lite model that uses features like merchant name, amount, and date. A recurrent neural network (RNN) forecasts future cash flows based on historical patterns. The app's service worker ensures it works offline and can run background processing. The entire codebase is open-source, and the models are quantized to run on mobile CPUs.

• On-device ML accuracy may be lower than cloud-based models, especially for rare transaction types.
• User adoption may be limited by the lack of automatic bank sync (manual import is a barrier).
• Storage limits on mobile browsers (e.g., Safari's 50 MB cap) could constrain long-term data history.
• Potential performance issues on low-end devices during model inference.

• Should we offer optional encrypted export to a user-controlled cloud (e.g., Nextcloud) for backup?
• How to handle multi-device use without a server? (Maybe encrypted transfer via QR codes?)
• What granularity of ML forecasting is feasible on low-end devices? (e.g., weekly vs. daily predictions)
• Should we include a local 'anomaly detection' feature to flag unusual transactions?

I pitched this to a few privacy-focused groups, but the complexity of building accurate on-device ML from scratch was daunting. I moved on to a simpler budgeting app that at least had manual entry. The idea still feels right, but the engineering effort never materialized.

Consider using WebAssembly for faster ML inference. Also, could integrate with the File System Access API for seamless local backup. A potential UX win: allow users to snap a photo of a receipt for automatic OCR and entry (fully offline using Tesseract.js).