Build log · 2 of 2
Quid · Finance · 2025 – 2026
From LLM demo to agents that do the books
Small finance teams lose days to receipts, invoices and journal entries typed into ERPs by hand. I joined Quid early and turned its first LLM prototypes into agents that do that work.
- Role
- AI Engineer
- Domain
- Finance and accounting
- Stack
- Python · TypeScript · Agno · AWS + GCP
My part
I took Quid's first LLM prototypes to production multi-agent systems: the AI layer of the product, an expense assistant in Slack and agents that help with bookkeeping, plus the full stack around them.
How it works, live
A simplified view. As you scroll, a sample receipt travels from Slack to a finished expense report, with a one-line trace of what the agent is doing at each step.
- 01 · Input · activeReceipt in Slackphoto or PDF
- 02 · Extract · activeFieldsmerchant, amount, VAT, date
- 03 · Map · activeCategorieswhere the expense belongs
- 04 · Report · activeExpense reportready for approval
- 01slack · receipt photo posted (sample)
- 02extract · merchant · total · VAT · date
- 03map · category suggested · asks if unsure
- 04report · drafted · waiting for approval
What makes it hard
Accounting leaves no room for "roughly right". Receipts are blurry, invoices mix languages and tax rules, and every company keeps its books a little differently. A wrong booking costs more than no booking.
So reading one document isn't enough. The agent has to work like an accountant, with the context a person would have.
What I learned
The model is the easy part. What turns a prototype into a product is everything around it: the right tools, the right context, and clear points where a person approves.
Put the agent where people already work, like Slack, and let it ask when it's unsure instead of guessing. And on an early team, most of the job is choosing what not to build.
Outcome
In production
Used by Quid's customers while I was there.