The how it works technical thread

The thread, with margin notes

1. Most invoice AI tools only work on clean PDFs. Real supplier invoices are photos of crumpled paper, in 4 languages, with totals that do not add up. Here is how we built an agent that handles the ugly ones:

   The hook names the gap between the easy version and reality. Technical readers love a thread that admits the hard case.

2. Step 1: we do not trust a single pass. One model reads the layout. A second extracts fields. A third checks the math. If the line items do not sum to the total, it goes back.

   Plain words for an architecture. No model names needed in post two. Show the shape of the system.

3. Step 2: the math check catches more than we expected. About 1 in 12 test invoices fail it on the first pass. Most are not model errors. They are supplier errors. Customers now use that as a feature.

   The surprise is the best part of any technical thread. Here a safeguard turned into a selling point.

4. Step 3: confidence per field, not per document. A 98% confident invoice with a wrong IBAN is still a wrong payment. So every field gets its own score and anything below the bar is routed to a human.

   Shows judgement about risk. Buyers in finance care more about this post than any accuracy headline.

5. What we got wrong: We started by fine tuning. Spent 5 weeks. Prompted pipelines with checks beat it in a weekend. Lesson: build the checks first, then decide what needs training.

   A lesson other builders can use. This is the post that gets bookmarked and quoted by AI creators.

6. We are opening the beta to finance teams processing 500+ invoices a month. If that is you, reply or DM. Happy to go deeper on any step above too.

   Qualifies the audience in the ask. Offering to go deeper invites technical replies that keep the thread alive.

Next thread