AWS architecture for chat-driven bookkeeping with an LLM harness
Finance solutions for small businesses. Book expenses and send invoices by writing a chat message, with an LLM that proposes and deterministic code that decides.
Write what happened, attach the receipt. The harness turns it into a correct booking or asks back.
Numbered, validated and archived before anything leaves the system.
EU region only, encrypted end to end, backed up and fully auditable.
The story
Bookkeeping is the part of running a business nobody wants to do. ale.business turns it into a conversation: write "paid 49 € for software" and attach the receipt, or ask for an invoice to a customer, and the system does the rest.
An LLM on its own is not good enough for finance. It can misread an amount, pick the wrong tax rate or invent an account. So I built a harness around it: Bedrock only returns a structured proposal, and Python code checks the schema, the double-entry balance, the tax rules and duplicates before anything is booked. Unclear cases go back to the user in the chat.
Invoices follow the same principle. Each one gets a gapless number, is validated for mandatory fields and totals, rendered as PDF and e-invoice, sent through SES and archived in S3 with Object Lock, so it can never be changed afterwards.
Because it is financial data, privacy shaped every decision: everything runs in the Frankfurt region, data is encrypted with KMS, backups are encrypted and stay in the EU, and CloudTrail records every access.
Invoicing
An invoice is a legal document, so the pipeline treats it as one. If any check fails, nothing is sent and the user gets a concrete fix in the chat.
Technology
Other case studies worth a look
Working on something similar?
Tell me what you run today and where it hurts. I will come back with how I would build it, and what I would leave alone.
Discuss your project