Numeric Hosts The First Finance Engineer Cup

The Builds

The finalists were invited to bring anything they'd built for their own close process to the championship. Here are those final builds.

Alexandria Ramirez (Betterment): The Four-Eye Check

"Putting AI into the close is a trust problem, not a tech problem." That was Alexandria's opening line, and her whole build follows from it. "We don't build faster reconciliation. We build governance for AI reconciliation, a trust architecture that a non-engineer can actually audit and deliver to auditors."

The Four-Eye Check runs two independent AI agents, each with different instructions, against the same reconciliation feed. A deterministic controls layer built from traditional code auto-clears items where the agents agree and flags disagreements for human review. It's segregation of duties, translated to agents. The output is a tamper-evident evidence package, and the whole thing maps to COSO's guidance on internal control over generative AI, down to a cadence drift detector and multi-party approval for rulebook changes.

The repo is public, and Alexandria wants it that way: "Please iterate, fork, improve. I'm a huge fan of the open source idea."

Leslie Cunnane (Jasper AI): Cash Intelligence Dashboard

"The issue was never understanding what changed with the numbers throughout the month, but it was really being able to explain why."

Leslie's Cash Intelligence dashboard connects through MCPs to Numeric, NetSuite, and Ramp into one live view of cash: burn, implied runway, forward projections, and AI-generated flux commentary for every month-over-month and quarter-over-quarter move. Everything drills down to the transaction, controllers set their own variance thresholds, and sign-offs flow into a running audit tab.

The clever part is how she solved hallucination. Early versions pointed the model straight at the GL, and "the data hallucination became a real problem. I was just getting wild responses that were inconsistent month over month." So she anchored the commentary to something deterministic instead. Leslie's team has been a Numeric customer since last year, and every line item that runs through their bank is already tagged against one of Numeric's cash-match rules. When Numeric added the ability to export that rule data through the MCP, she fed the model each line's existing classification, instead of pointing it at raw GL memos. The model stopped guessing, and the commentary settled into what she wanted: "very concise one-sentence liners about the flux."

Leslie also took home the audience vote, making her the People's Champ of the inaugural Cup.

Casey Barry (Pulumi): The 18-Day Ramp

"On day one, my head of finance gave me the most finance team welcome possible: 'Great to have you. We need to close April yesterday. Go figure it out.'"

Casey joined Pulumi mid-close, inherited books that didn't tie, and treated onboarding like an engineering problem with his build, 18 Day Ramp. He used AI as a reconstruction layer to rebuild a drifted subsidiary's trail transaction by transaction. He recorded process walkthroughs and turned them into operating maps: who starts the process, which system is the source of truth, where the control points sit. Then he rebuilt the close workbooks to be agent ready, each with an instruction tab telling the agent what data to use, what to flag, and what a human reviews before anything posts.

Casey’s philosophy on building with AI: "Move fast, but only as fast as the explanation can follow. I use AI not to skip the hard parts of finance, but to make the hard parts visible, structured, repeatable, and reviewable.”

Matt Atieh (ClickUp): Argus

Matt built Argus in Replit to eat ClickUp's AP inbox. Invoices land, Argus extracts, classifies, and matches them to POs, then pushes everything to Coupa in pending-approval status, so approvals and controls stay in the system of record. Instead of matching rules, it takes natural-language "memories" ("AWS marketplace invoices always go to this vendor") that get injected into context every time an invoice processes. It even runs its own MCP server, so ClickUp agents can query AP data and draft vendor emails without anyone exporting a thing.

Matt put a number on the time it saves his team: "Just implementing this, we're saving the team like 60 to 80 hours a month from processing invoices."

Crowning The Champion

Alexandria Ramirez took home the $5,000 price with Four-Eye Check. Her build’s central idea, two AI agents checking each other's work under a deterministic controls layer, sits squarely inside two of the five scoring categories: CONTROLS and ACCT'G INTEL, the ones built specifically to reward a build that treats auditor scrutiny as a design constraint rather than an afterthought.

None of that scrutiny was lost on Alexandria herself. Looking back on the win later, she put it simply: "Honestly, both very exciting and surprising, and really cool to see a lot of other people respond in a positive manner and see what I was seeing." For more on Alexandria’s win and approach to building, read our full profile with her after the competition ended.

That same bar applied to the whole field. Chris Gaetano summed up the four finalists: "Every single one of these solutions was responding to real problems that I have heard a lot of people moan or complain about in their day-to-day. I've seen some incredibly janky vibe-coded apps before that leaked like a sieve. None of these were like that. I was incredibly pleasantly surprised."

Beyond The Cup

The scores were in, the winner announced, and half the room was already talking about what they'd build next. Steve had already made plans: "I can't wait to get back to my desk and start building, because this is just so motivating."

Finance engineers are increasingly choosing to build on platforms that structure the close in a way that gives an agent real ground to stand on, instead of forcing it to guess at raw, disconnected data that lives in your ERP. Leslie's build is a great example of this, connecting to Numeric’s existing deterministic cash matching rules through the MCP allowed her to build a reliable cash dashboard. It's the same reason we've seen models connected to both an ERP and Numeric reach for Numeric on their own when asked a question about the close: it's simply where the context already lives.

One take away form the competition that stands true: If you're a finance engineer with your own workflow in mind, consider building with something that gives your close context instead of disparate data to parse through.