It started with a simple request. A client told us their team was still typing invoices into a spreadsheet by hand, one by one, and asked if there was a better way. At DZ Labs we like that kind of problem: small, real, and something people deal with every single day. So I built Invoice Register. You give it a PDF invoice, and a few seconds later the data is sitting in a Google Sheet: supplier, dates, totals, and every line item in its own row. It works with digital invoices and with scanned ones too. In the video you can see the whole trip: I upload a PDF, the API answers with the structured data, the rows show up in the spreadsheet, and the service is running live in the cloud. Here is what I used to build it:

Python 3.12 and uv to manage the project and its dependencies

FastAPI and Uvicorn for the API

Pydantic to define and validate what an invoice looks like

pdfplumber to read text from digital PDFs

pdf2image, Poppler and Tesseract to read scanned invoices with OCR

LangChain and Claude to turn raw text into structured data

gspread and google-auth to write into Google Sheets

LangSmith to trace every run and see what the model did

pytest for the test suite And here is what I used to deploy it:

Docker to package the app with its system dependencies

Cloud Build and Artifact Registry to build and store the image

Google Cloud Run to run it, scaling down to zero when nobody is using it

Secret Manager to keep the API keys out of the code I will be honest: writing the code was the comfortable part. Getting it to production is where I learned the most. When the data belongs to someone else, security stops being optional. The service runs without any key files, the API is private, and every secret lives outside the code. What I enjoyed most was not the technology. It was knowing that someone will get back hours of their week. Next steps: a simple frontend for uploads, background processing, and automated deployments. If your team is still doing something by hand that a computer should be doing, I would love to hear about it.

#Python #FastAPI #LangChain #GoogleCloud #CloudRun #AI #DZLabs

Since you're tracing runs in LangSmith, I'd separate debugging metadata from invoice contents. Can you follow a failed extraction using a document ID and error details without recording supplier or line-item data in the trace?