Close the books
while you sleep ✦

An agent finance team: it reads the inbox, matches the bank, books every entry through a checker that re-performs it, and asks a human exactly once.

Scroll ↓

Drift board

BankLedgerCRMContractsInbox
Initech · short-pay
Invoice INV-1042$12,000.00
Bank received$10,800.00
reason found · CEO email−10%
Wayne Corp−$3,300
no evidence anywhere→ ask once
Globex · wire feeauto-posted
$20 short · compiled policy0 model calls

One month, closed

ARAPBank recRevenueClose
14 entries posted by code · 2 sent to a human
Books tied out✓ to the cent
Wrong entries posted0
01 · one event, every book

A payment lands
$1,200 short. Watch.

The reason is in somebody's inbox

The drift monitor flags the gap. The agent checks the contract, the CRM, then finds the CEO's discount email nobody told finance about. Revenue, forecast, billing and audit all move together, to the same cent.

Nothing posts without proof

Every entry carries a workpaper: the math, the exact quoted evidence, the approval. Deterministic code re-performs every check. Change one letter of the quote and it is rejected.

02 · the loop

Every close,
smarter

Four beats, run live, with a checker that makes the trust provable.

Learns

Replays the quarter humans closed with the answers hidden; compiles their repeated judgment into policies.

wire-fee write-offs → compiled policy

Runs

Routine work clears through rules your team approved. The kernel re-performs every entry before it posts.

14 entries posted by code · zero AI calls · $0

Escalates

A genuine blocker goes to a person's phone, asked once, answered once, remembered with a scope and an expiry.

Vossberg −$2,200 · Haiku, 10 cents · decided once

Improves

Corrections become policy v2, re-tested on all of history. Replay agreement is measured on every learn.

replay agreement: 0 of 9 → 8 of 9
Readerour model reads the mail
Preparerdrafts the entry
Checkercode re-does the math
Humanonly when truly stuck
Auditorre-tests at month end
03 · the model

A small model,
taught by the audit trail

Every kernel-verified decision becomes training data. Overnight, the same open-weight model went from failing this work to leading our table. It trains and runs on the ASUS GX10 sitting at this table.

Full benchmark ↓
field-F1 · held-out · n=120
0.061 to 0.972
same model, before vs after training
cost per 1M generated tokens
$0.40 measured, batched: power + hardware over 3 years
Claude Haiku API: $5 · Sonnet: $15 (published prices). Measured on 614 documents: 50 docs/min, 1.2 s per document, zero errors. A fixed cost we own; API cost scales with volume.
04 · the receipts

NorthwindBench

Every model takes the same exam: 120 financial documents it has never seen. perfect answers means the whole answer was right to the cent. Muse is told the exact answer format on every question. Our model is not, because training taught it the format.

Outside check: on a public invoice benchmark we did not write, graded by their own script, our model got 15 of 18 customers exact to the penny, with every trap document correctly thrown out. And on a brand-new exam generated after training, 614 documents that did not exist before: 0.964 accuracy, matching the original test, so nothing is memorized.

modeldocs 100% right, of 120field accuracynew vendors and formatsspeed per doccost per 1M tokens

Loading results…

05 · one picture

Every model, one exam

Accuracy on the same 120 documents, same grading script. Colors: what kind of help each model had.

fine-tuned by us given a cheat sheet each call no help, as-is

And what a million generated tokens costs. API rows are published prices; ours is measured all-in cost on our own hardware (electricity plus the machine over 3 years).

Our model on the GX10 $0.40 measured Claude Haiku 4.5 $5 Claude Sonnet 5 $15 Claude Opus 5 $25
05 · live

Decisions,
as they happen

Real cases from the seeded month: what arrived, what the system did, what it cost. Green ran with no human. Coral waited for one, exactly once.

Initech Corpcredit memo · CEO email cited
−$1,200.00
kernel ✓
Globex Systemswire fee · compiled policy
−$20.00
0 model calls
Wayne Corpno evidence · asked once
−$3,300.00
awaiting you

You asked how much of a finance team
an agentic system can run.

In our test month: 8 of 11 bank payments were booked by pure code, 3 went to a human exactly once, and 0 wrong entries were posted. The books tie to the cent.

See the results
Money Maxer
An agent finance team for the office of the CFO. Built in 24 hours at HackMIT 2026 by Preet, Karan & Atharv.
© 2026 Money Maxer