AI Engineer - Accounting Automation

The short version We manage the accounting for SMB — deferred revenue, month-end close, sales commissions, reconciliations We try to automate as much as we can. Today one person builds all of it, working alongside AI coding agents. We’re hiring the person who takes that over. What the work actually looks like Not a list of responsibilities — four real requests from the last few months, close to how they arrived: “In our finance dashboard, if I go to the balance sheet and click the Equity account, it shows nothing. Beginning and ending balances are both zero and there are no transactions. There should be about a dozen.” You’d: reproduce it, trace the drill-down query back through the API into QuickBooks, work out why equity specifically comes back empty, fix it, and write the test that catches it next time. Half a day, and the fix is usually one line somewhere nobody would have guessed. “Our client’s healthcare account should be zero at the end of every month, but it never is. The insurer bills us at the start of the month, and the payroll system is supposed to post entries for each payroll showing what we actually owe. Some of that is company expense and some is withheld from employees.” You’d: pull a year of that account’s activity, work out which side of the mismatch is wrong, and build something that reconciles it every month and flags what a human needs to look at. A few days — and most of the work is figuring out what “correct” even means here. “Our app says this contract is recurring. The only product on it is a fixed-term product. Why?” You’d: find where that classification is derived, discover it’s reading the wrong field or matching on a name that was renamed six months ago, and fix the derivation — not the one record. “For our client that uses a property-management system, this interest expense recurs every month. Can we create a journal entry dated 4/10 and every month after, through December 2028? Confirm the fields with me first.” You’d: build it, show the exact entry before creating anything, and make it idempotent so re-running it doesn’t double-post thirty-two months of interest. Today all of this runs today on one person’s desktop. Over the next few months we hope to move to cloud and multiplayer development. You would have a large hand in that. You are probably right for this if You already drive AI coding agents hard, every day. Our stack, plainly Python 3.12, FastAPI, SQLite, pytest. Plain HTML, CSS and JavaScript — no React, no build step. Claude Agent SDK. QuickBooks Online, Stripe, Airtable, Front, Ramp, Appfolio, Google Sheets. One Ubuntu VM, systemd, nginx, Cloudflare Access. Git worktrees, GitHub PRs, cross-model AI code review. Agencies and dev shops You’re welcome to put someone forward — on the same terms as everyone else. The person who would actually do the work sends their own submission, does their own exercises, and is the only person we evaluate. Do not send any messages from an account manager who is pitching the firm. Have an actual engineer apply. To apply — build a QuickBooks agent No CV, no cover letter, no “tell us about a time.” Build a small thing and show us. About two hours. Build an agent that answers accounting questions and creates records in QuickBooks. Data: we give you the customers — a CSV of 312 accounts with their open balances (attached). Load it into Intuit’s free developer sandbox, or into a mock you write yourself. Your call, and tell us which you chose and why. Interface: up to you. Show us what you’d hand a bookkeeper. Model, framework, language: whatever you’d actually reach for. Make it handle these two requests: “Create an invoice for Acme Consulting, $4,500, for September consulting work.” “Which customers owe us more than $10,000, and what’s the total outstanding across them?” What to send — three things 1. Five short answers. Under 300 words in total — we read short, and this is part of what we’re reading. How long did each request take — before and after any change you made? How do you know the second answer is correct? What did you change along the way, and why? What does the model do in your build, and what does plain code do? Why did you split it that way? What’s your setup — harness, config files, how you drive the agent — and what would you change about it? 2. Ten continuous minutes of Loom from anywhere in the build — your pick of window. Record it while you work and send it unedited; please don’t re-record a tidy summary afterwards, because the thing we want to see is you deciding in real time. Talk while you work: how you decide what to trust, when you check the agent’s output, and what you do when you’re not sure. 3. A zip: the code, your config (CLAUDE.md, AGENTS.md, a rules file — whatever your tool uses), any hooks or settings you run, and the session transcript your tool produced while you built this. Every tool stores that differently and some store nothing at all — send what yours gives you, and just say so if it gives you none. Redact anything sensitive before you send it: keys, local paths, anything from a day job. Spend about two hours. We’d rather see where you got to in two hours than something polished that ate your weekend — and we’ll ask you what you’d do next. Use any AI tools you like; we assume you will.