⏱️ A wage garnishment order used to take a person 30 minutes to type in. Our AI agent has it ready in under a minute. Some context. A court can order that part of a contractor's pay is withheld to cover unpaid child support or a debt. That order arrives as a scanned PDF, and @deel clients used to retype ~30 fields from it by hand. Now they upload the document and the agent does the rest. It runs on Akai, Deel's agent platform, and has been in production end to end since early September. 🚀 Getting the model to read the PDF took days. Making it safe to run in a workflow that moves real money took most of the build, and came down to three unglamorous guardrails: 🛡️ Compare-and-swap writes, so a late callback and a timeout sweep can't both finalize the same record 🔁 Explicit states with a sweeper for runs that never resolve 🧩 A strict JSON schema as the contract, not the prompt None of it is clever. All of it is why this is production-grade instead of a demo. Full write-up on how it was built and what mattered most 👇 ❓ If you've shipped an LLM where the output moves money, which guardrail mattered most to you?