AI Bookkeeping Software is a full-stack, AI-powered bookkeeping and financial management solution built for modern businesses. It ships as complete, deployment-ready source code covering a Python/FastAPI backend, a React 18 + TypeScript frontend, and a Docker Compose infrastructure layer — giving you everything you need to run it on your own servers or offer it as a SaaS product to your customers.
The software automates the most time-consuming parts of bookkeeping through eight verified AI and machine learning capabilities:
- Intelligent transaction categorization using FAISS vector search and embedding similarity against a structured Chart of Accounts, with keyword boosting and a per-transaction confidence score
- Natural-language query (NLQ) engine that translates plain-English questions into safe, read-only SQL and returns both raw results and a human-readable summary
- PaddleOCR-powered receipt and invoice processing that extracts vendor, amount, date, and line items from uploaded images and PDFs
- Automated five-stage invoice approval workflow (Uploaded → OCR → Categorizing → Review → Posted) that routes every document through extraction, categorization, and bookkeeper sign-off before creating a ledger entry
- IsolationForest anomaly detection that flags statistical outliers in real time, surfacing them in the dashboard and attaching an anomaly score and explanation to the transaction record
- Cash-flow forecasting using a Prophet + PyTorch LSTM ensemble, producing daily predictions with confidence intervals up to 180 days ahead, rendered as an interactive area chart
- Bank reconciliation with fuzzy string matching and amount proximity scoring, with a two-panel UI for bulk-approving, confirming, or rejecting suggested matches
- Server-Sent Events (SSE) streaming narrative generator that produces a plain-English monthly executive summary of revenue, expenses, variances, and trends, streamed token-by-token to the browser
The included training pipeline and seed scripts use synthetic, programmatically generated transaction data — realistic vendor names, amounts, and categories produced with seeded randomness. This approach improves model coverage across diverse business scenarios while ensuring no real user financial information is embedded in the codebase or training assets.
The software is built for technical buyers: developers, indie hackers, and small software teams who want a production-grade foundation without building the AI and accounting logic from scratch. You get a full Chart of Accounts in JSON, a labelled JSONL categorization dataset, offline training and embedding-index scripts, pytest-asyncio test suites with factory-boy fixtures, Alembic migrations, structured JSON logging via structlog, Prometheus metrics instrumentation, and an Nginx reverse proxy with SSL configuration included.
The backend runs on FastAPI 0.111 with SQLAlchemy 2 async, PostgreSQL 16, Redis 7, and Celery 5 for background ML task queuing. The frontend uses React 18, Vite 5, Tailwind CSS 3, Recharts 2, Zustand, and TanStack React Query. Everything is containerised with Docker Compose, including a production override file and a deployment checklist covering TLS, environment hardening, and managed database configuration. Minimum recommended hardware is 8 GB RAM due to the memory requirements of the bundled ML models.
