How Alexandria Ramirez Won Numeric's First Finance Engineer Cup

Alexandria's build that won the Cup is a governance tool. Two independent AI agents reconcile the same feed. A deterministic layer clears everything the agents agree on, and anything they disagree on gets flagged for a human to review. The design borrows from a control auditors already trust: it mirrors the four-eyes principle rather than replacing it.

The instinct came from her seat. Betterment is a late-stage fintech where the controls mindset is, in her words, very forefront. "I don't have an audit background, but I question things the same way an auditor would," she says. So while she's been running at AI as fast as anyone, she's carried a healthy skepticism the whole way: "It just spun this out really cool, really quickly. But how do I know?"

She's deliberate about framing Four-Eye Check as a starting point for exploration rather than a full-fledged product. "I wanted to show a v1 of a concept  where you could run this agent-led reconciliation and feel really good about its output."

The thesis behind the build

Alexandria's building philosophy rests on two convictions, and a third opinion she'll happily label a hot take.

Every accountant becomes an agent manager

"I firmly believe we're moving to a state where everyone will be managing a sub-fleet of their own agents," says Ramirez. In her picture of the near future, the people on her team don't all move into people management. They move into agent management, overseeing sub-agents that each perform a slice of their old day job. The work doesn't disappear; it shifts into maintenance, error review, system checks, and sync reviews.

She won't commit to a date. "I don't have a timeline, but that's where we're running." What she will commit to is preparing for it: the habits she's building on her team today (more on those below) are shaped around that destination.

Trust is a build problem, not a tech problem

The second conviction follows from the first. If everyone in AP and AR is overseeing agents, how does a manager know the output is right? How do you trust it enough to close the books on it?

Her answer : "Other industries have proven that this is actually not a technology limitation. It's a build limitation from how we as accountants have thought about this problem. The technology makes it pretty easy to lock this down."

The pieces for governance exist in other industries, she argues; it's just that nobody has assembled the right governance layer for AI in accounting yet. She's floated ideas as far out as a tamper-evident record for agent actions, "a blockchain or something even more secure," while being upfront that "we're in the infant stage of even conceptualizing it." What worries her is the industry stopping short: "My concern is that we plateau at 'an agent can run it' and don't actually finish getting through to 'how is it truly secure.'"

Auditors, she thinks, will sort along the same line. "Some teams will get really lucky with forward-thinking auditors at the forefront," she says, pointing to the very public AI investments the Big Four are making in their own upskilling. "Other auditors will be more grounded in what they know, trying to apply the current mindset to new tooling."

Why data ownership belongs in finance

Ask Ramirez what she's chewing on outside of AI and you get her most polarizing opinion (her word, not ours). It's about who owns the definitions.

"The way it's always been, your data engineer sets up the query, sets up the warehouse, and defines the metric: the universal definition across the whole system. That doesn't really work in finance," she says. Finance historically kept its data siloed, and when it needed something from the shared warehouse, someone in data analytics handed over the metric and the query. "I don't think anyone in finance felt empowered to ask, 'well, how did we define that?'"

That mismatch, in her view, is the root of a familiar misery: the endless manual cleanup to wrestle a number into matching finance's definition. Her prediction:

I foresee, and it could be a finance engineer a true data person who sits in the finance org whose sole responsibility is managing and writing that metric and those definitions for teams to use.

The habits that made it stick

If you're reading this wondering what her team actually does differently, here's the machinery. Everyone on the team has a DataCamp subscription. There's a dedicated Slack channel for AI learnings where people showcase course accomplishments. Every two weeks, the entire accounting team checks in on where they are with AI learning.

But Ramirez says her real lever is smaller and quieter: her one-on-ones with her direct reports.  With them, she asks, "what are we building, what are we working on, how can I help unblock?" The payoff shows up unprompted. One team member recently asked for an API key so they could pull data from a system themselves. Another told her, "I built this dashboard, it's in production, this is what I'm using it for."

"That's what I've gotten most excited about. It's a culture of building," she says.  When a tenured team member recently left and the workload got heavier, her message was that the learning doesn't stop; it just stays incremental. "We need consistent incremental progress: 15 minutes some days, longer on others, but you just have to do something and check in to keep building on that learning."

The pause before the sprint

The habit that made everything else work was, oddly, stopping. When Claude first landed, her team met weekly in a rush of stand-it-up energy. Then, around early May, Ramirez hit pause and asked herself two questions: "Do I have enough fluency to review everybody else's work without re-performing everything myself to verify the output?" And: does everyone on the team actually understand the limitations of this technology?

The honest answer to both was not yet. That pause is what turned enthusiasm into a practice. The DataCamp subscriptions, the shared channel, the bi-weekly check-ins all came out of it. "It's fantastic technology and we were never afraid of it," she says, "but there are limitations we need to be aware of."

What still keeps her up at night

Ramirez has plenty of momentum, but two worries keep pulling at her:

A two-speed industry

Her biggest fear is arithmetic: token costs rising faster than accounting's willingness to adopt. "There's a risk that AI moves so quickly, and token costs increase so dramatically, that other teams don't get to catch this wave and miss it entirely," she says. People like her, builders inside tech-forward teams, are still a small subset of the profession. She used to think of this as a three-to-five-year concern. "At this rate it might be by year-end."

The divide is already visible outside the major hubs. One round-one competitor from Atlanta told us that in his circles, people might know Claude or ChatGPT, but the name Anthropic draws blank stares. He feels like he's in a bubble that San Francisco and New York take for granted.

And Ramirez takes the worry one step further, into the CFO's chair. As model costs climb, AI spend becomes an allocation decision. "Is that an area where people say, 'actually, let's not spend on AI here; let's allocate more to something revenue-generating like marketing leads, sales, engineering, and core products'? It's a very real possibility as model costs keep increasing."

Betting on the right vendors

For someone who just won a building competition, Ramirez is unsentimental about the limits of building. "If a company is allocating millions of dollars toward a dedicated platform for something you're trying to solve, you'll never keep up with it as a side project you give an hour or two a day," she says. "There are subdomain experts and best-in-class vendors for a reason."

The AI boom hasn't changed her build-vs-buy math so much as made vendor diligence harder. The question she's iterating on is which vendors are real and which are riding the moment: "which are built and investing to take it to the next level, which have the right framework to keep being inquisitive and won't stagnate, versus which are just trying to seize the moment with something flashy that might not work."

Her rule of thumb splits the difference. Personal builds, anything that hands you back your own time, are always worth doing yourself. "You're not delivering the end product to anybody else; it's not going anywhere. But if it gives you back time, that's the best possible use case of AI, in my opinion." Anything higher stakes, with real maintenance behind it, she'd rather buy from someone whose whole job is maintaining it.

Why she really showed up

"Doing all of this in the vibe-coding environment can be very isolating. It's given me a lot of empathy for software engineers," she says. Weekends of solo building left her thinking, "I haven't talked to anybody, I've been sitting here attacking this problem for a while." The Cup's live format gave her what solo building can't: presenting to people who get it, live feedback, and a window into how other builders attack the same problems. It's the same environment she has brought to her own team.

As for where she goes next, Ramirez isn't pretending to know, and that's her favorite part. "In January I did not think I'd be here in July, and I don't want to predict where I'll be in December." She has one request for the rest of the org chart, about the team she runs: "We're not just bookkeepers. We're the foundational piece for where all the rest of the data goes. We're step one."