Bogdan Stoica Get in touch
← back to projects

AI invoice-processing platform

Mailbox ingestion, LLM extraction with a Document Intelligence fallback, approval workflow, automatic booking in Dynamics 365.

Outcome: Built end-to-end as the solo owner: from the mailbox an invoice lands in to the entry booked in Dynamics 365.

LLM extractionAzure Document IntelligenceDynamics 365Next.jsAzure Container AppsKey VaultAzure DevOps

Completed September 2026

An anonymized case study. The client and their numbers are confidential; the architecture and the engineering decisions are mine to talk about.

Context

Vendor invoices arrive by email as PDFs and scans, in every layout a supplier can invent. Someone then reads each one, types the values into the ERP and chases an approval. The goal was to take the retyping out completely and keep a human only where a human adds something: the approval.

What I built

  1. 01 Mailbox ingestion Invoices arrive by email and are picked up on their own.
  2. 02 LLM extraction Fields come out as structured data, not text soup.
  3. if unsure Document Intelligence fallback Low-confidence documents take a second route before validation.
  4. 03 Validation Business rules check the result before anyone trusts it.
  5. 04 Approval workflow The right person signs off, with the source PDF highlighted.
  6. 05 Booked in Dynamics 365 Posted automatically. No copy-paste step left.

The platform is a pipeline with five stages:

  1. Mailbox ingestion. Incoming invoices are picked up from a shared mailbox without anyone forwarding or uploading them.
  2. LLM extraction. A language model reads the document and returns structured fields: vendor, invoice number, dates, line items, totals.
  3. Document Intelligence fallback. When the model is unsure, or the scan is poor, the document takes a second route through Azure Document Intelligence before it goes any further.
  4. Approval workflow. The approver sees the extracted values next to the source PDF, with the regions each value came from highlighted on the page.
  5. Booking in Dynamics 365. Once approved, the invoice is posted automatically.

Engineering notes

  • Every stage is idempotent. A retry never books the same invoice twice.
  • Errors are classified. Transient failures retry with backoff; permanent ones are routed to a person with the context they need.
  • Every document carries a correlation ID through structured logs, so any invoice can be traced from the inbox to the ledger.
  • The infrastructure is part of the product. Container Apps, Key Vault, Blob Storage and Azure DevOps pipelines, all owned and operated by me.