Kestral Desk

Inspiration

Hyphenate asked: What autonomous system would you actually hire for a finance team? We picked the riskiest corner: everything that leaves the bank account at Kestrel Aerial, a fictional $137M drone maker where 40% of sales are federal. One wrong part in a shipment can cost it those contracts.

What it does

Claude agents fill the AP clerk and payroll specialist seats. They check every invoice and paycheck against real tariff schedules, FX rates, and IRS tables, hold what's wrong, and explain why with evidence. They can only request a person approve every dollar, enforced in code.

Then it learns: it mined 14 months of company history for the controller's unwritten rules, and once she approved them, she got 41% fewer escalations while it still caught 20/20 planted problems with 0 false alarms.

How we built it

Deterministic TypeScript engines do all the math (integer cents, 171 tests) the LLM never does arithmetic. Claude handles judgment: reading emails, deciding what to escalate, explaining why. Convex runs the live approval queue and audit trail. A seeded simulated company with a hidden answer key proves it works.

Challenges we ran into

Making "learning" safe. Kestrel's history includes decoy patterns and a trap: the controller once approved an emailed bank change that was fraud. The desk had to learn "call back first," not "approve bank changes." Some things compliance, bank changes, duplicates — can never be learned away.

Accomplishments that we're proud of

20/20 caught, 0 false findings, $2.79M identified, 0 control violations matching the answer key exactly on two different seeds, so it isn't tuned to one dataset.

What we learned

The hard part isn't catching errors — it's knowing which source to trust when the PO, receipt, invoice and contract disagree. And guardrails make an agent more useful: because it can't move money, it can be trusted to look at everything.

What's next for Kestrel Desk

Switch on the buyer and staff accountant seats, let the controller's overrides retrain rules automatically, and connect to a real ERP, HRIS and bank sandbox.

Built With

• anthropic
• convex
• next.js
• react
• sdk
• tailwind