AI Accounting Software is a full-stack, AI-native accounting software built for small-to-medium businesses. It ingests transactions, invoices, receipts, and bank feeds; automatically categorizes and reconciles them; forecasts cash flow; detects anomalies; and generates audit-ready financial reports — all behind a clean, professional interface built with React 18 and TypeScript.
The software is designed for developers, agencies, and technical founders who want a production-ready codebase to launch an accounting SaaS or embed intelligent financial automation into an existing product. The backend is built on FastAPI with async SQLAlchemy and PostgreSQL (pgvector), and the frontend uses Vite, Tailwind CSS, and TypeScript. Docker Compose configurations are included for both development and production, with Nginx reverse proxy and SSL termination ready to go.
Key capabilities include:
- AI Transaction Categorizer — XGBoost classifier trained on sentence-transformer embeddings, with confidence scoring and batch prediction
- Natural Language Financial Assistant — RAG pipeline powered by ChromaDB vector search and the Anthropic Claude API, with streaming responses and chat history
- Invoice AI Parser — uploads PDF or image invoices; Claude vision extracts all fields (vendor, line items, totals, dates, tax) into structured JSON automatically
- Anomaly Detection — flags unusual transactions in real time
- Cash Flow Forecaster — 90-day rolling forecast using Prophet, with AI-generated narrative summaries
- Smart Bank Reconciliation — fuzzy and semantic matching for fast statement reconciliation
- Automated Financial Reports — P&L, Balance Sheet, Cash Flow Statement, and Tax Summary
- Budget vs. Actuals — real-time variance tracking with AI commentary
- Multi-Currency Support — live FX rates via OpenExchangeRates, with gain/loss tracking
- Audit Log — immutable change history
- Role-Based Access Control — Owner, Accountant, and Viewer roles
The codebase includes an ai_training/ directory with training scripts, evaluation tools, and labeled datasets for the transaction categorizer. Synthetic data generation scripts are also included to improve model coverage across transaction categories and to protect real user information during development and testing — meaning you can train and validate the categorizer without ever exposing live financial data.
AI features (the assistant, narrative generator, and invoice parser) require an Anthropic API key set in your environment. Live currency rates optionally use an OpenExchangeRates API key. All other features run locally within your Docker stack. Default demo credentials and a database seed script are included for rapid evaluation.
