Inside a Mathematician’s AI Harness for Fortune 500

Set theory as a first question

Host Enzo Cavalie opens with an anecdote: when Nazlı Tan first met the person who is now her chief of staff, her first question was whether she had studied set theory. Nazlı explains that people who have taken courses in foundational logic, such as set theory or proof theory, tend to think differently and ask different questions when they face a problem.

Why the company is called Hilbert

The name connects to the same idea. Around the early 1900s, during what is known as the foundational crisis of mathematics, logicians began finding paradoxes and contradictions in the math everyone relied on. One of the first attempts to restore trust was David Hilbert’s program, which was also the company’s original name when it raised its pre-seed.

Nazlı sees growth the same way: as foundational. Left alone, a company naturally declines. Retention falls, markets saturate, competition grows and ad inventory gets more expensive, so even holding steady takes effort. Her favorite question for executives is whether they would bet their company’s one-year revenue on knowing what works and what does not. She says she never gets a yes. Instead, modern businesses run on meetings where teams compare different numbers, and companies hit targets without growing sustainably. With companies about to deploy a large agentic workforce, she argues, the problem had better be well defined, or mistakes will multiply.

Why she hires for math

Nazlı describes growth as users behaving in a space called a company. A mind trained in mathematics, she says, is set up to understand such systems, to ask the right questions and to be uncomfortable when something is not rigorous. The skill she uses most in business is logic: proof by contradiction, induction and, above all, counterexamples.

Her own path started at Citibank in credit analytics, then PwC in valuation and M&A, work she found boring. In Turkey, she watched Alex Schultz’s video on how Facebook got users and grew, which drew her into growth and, years later, led to a Hilbert Space episode with him. She went on to train everyone joining her teams with his videos, quizzes and cohort-reading exercises, aiming for a real “I understand” moment.

What frustrated her about growth practice

As the person colleagues sent hard questions to, Nazlı saw a lot of wrong analysis and reasoning. What bothered her most was how unaligned people were, and how much time and money went into meetings to align them. That, she says, comes back to the data foundations, which is why she believes you have to start with infrastructure.

She also rejects the split between creative and numerical work. If you can measure something, measure it fully and correctly. If you cannot, as with some top-of-funnel TV campaigns, use logic. She adds that she thinks acquiring a customer is more serious than raising money: that customer lives in your cohorts and retention rates for as long as the company exists. Because customers behave differently in month three, year one and year five, teams are forced to make growth decisions faster than the data matures. Hilbert’s idea is a metric system that gives full visibility into that long journey without waiting six months or a year.

Landing a very large customer early

Asked how Hilbert closed a very large enterprise customer so early, Nazlı credits starting bold, a strong team of four co-founders and her co-founder’s talent for outreach. The breakthrough came from finding an executive who was as frustrated with the problem as she was. Early on, her sales style was education-heavy; after a sales course, she learned to ask about pain: what is the biggest problem you cannot solve, and “give us an expensive one.”

Deep learning as the backbone

Hilbert started from deep learning and machine learning models and built the platform on top, so some customers used the metric system before the platform existed. Customers are mapped onto Hilbert’s schema rather than receiving a custom solution. No single schema fits every business, she explains, but a finite set of vertical schemas can. Hilbert typically asks for three to four years of history, cleans it thoroughly and says it has always found issues in the data.

Enzo notes that SaaS traditionally had no opinion, and AI lets software have one. Nazlı clarifies that what Hilbert changes is the customer’s metric system, including how they define value, risk and inactivity, but its opinion comes from the company’s own history in a closed learning loop. The answer is already in the data; deep learning can connect dots across long journeys that humans cannot.

Deterministic versus probabilistic

Her first differentiator is that Hilbert is not a pure LLM product. Deep learning predictions are deterministic: the same question about the same customer state gets the same answer. LLMs are probabilistic, and she asks whether anyone could deploy hundreds of millions of dollars on a system that changes its mind every day.

Hilbert therefore uses a specific harness in which agents are significantly limited. Early versions let agents define “risky” customers, and the definitions changed constantly: no order in three months, then six, then shrinking baskets or falling contribution profit. She is not persuaded that more context fixes this, and points to an agent declaring month-over-month decline on day 20 of a month, a mistake no junior analyst would make. Hilbert optimizes for trust, with extensive tests, and agents are not allowed to do math beyond rounding; even simple arithmetic runs on the database.

Proving and pricing value

Hilbert does not start an engagement without an A/B test, and Nazlı stresses causality over correlation: if third-order customers retain, does pushing customers to a third order actually create retention? On pricing, she avoids per-seat pricing because everyone in a company should use Hilbert, and offers fixed fees tied to company size and complexity, or success-based arrangements for those who prefer them. Asked whether success raises the bar each year, she says that is growth: the goal is a compounding machine rather than linear or diminishing returns.

What Silicon Valley gets wrong

Nazlı argues that most growth literature comes from Silicon Valley and digital products, and often says paid marketing does not work. For groceries, retailers and QSRs, she says, “all of us have castles built on Metas, Googles,” and you cannot walk away. The real issue is not knowing early what you are acquiring. On creativity, she agrees with Alex Schultz that it is not about pixels but about ideas like a referral mechanic that goes viral, and says Hilbert can act as a thinking partner for those ideas.