A tenth of a second is scarcely enough time to regret hitting Send. For Nick Ganju, it is enough time to begin guessing what a caller means. When a person pauses while speaking to Zocdoc's automated phone assistant, the system can start preparing a response after 100 milliseconds of silence. If the person continues, the software throws that attempt away and begins again. It is impatient in precisely the way a good listener is patient: always ready, never committed too soon.
This small trick is called greedy inference, a phrase that sounds like a vice at a hedge fund. In practice, it is an effort to remove the dead air that makes a conversation with software feel like a long-distance call to the moon. A conventional system might wait two seconds to be certain the speaker is finished, then spend more time forming and voicing an answer. Ganju's approach begins work at the first hint of a pause. The discarded guesses cost computing power. The surviving one buys time.
The trade is characteristic. Ganju has spent much of his life finding the mechanics inside things that other people experience as mysterious: computer science, probability, billiards, pop songs, a marketplace, now conversation itself. His career is full of complicated systems, but his preferred unit of thought is agreeably small. A number. A rule. A repeatable move. A pause.
The price of a pause
Illustrative timing from Ganju's technical account: begin processing after 100ms of silence, then cancel if the caller continues.The child at the keyboard
Ganju grew up in the western suburbs of Chicago and started programming games on an Apple IIc when he was about nine or ten. In high school he moved the habit to programmable calculators. The social dividend was limited. Years later he supplied the dry verdict: “It turns out that writing Tetris on your calculator doesn't actually win you the cheerleaders.”
The line is funny because it admits the earnestness beneath it. He liked computers before they could confer status, and he liked making things before “builder” became a flattering noun. A guidance counselor suggested computer science. At the University of Illinois at Urbana-Champaign, he found a discipline capacious enough for that appetite and completed his degree in 1998.
His route after college traced the late-1990s technology map. He joined Trilogy Software in Austin, then moved west during the dot-com rush. He worked on eBoodle, which was later acquired by BizRate, and in 2002 founded Mixonic, serving as architect, manager and lead programmer. He later became CTO of cfX, where the software served large Wall Street firms. The work moved between consumer products and financial systems. The recurring role was translator between a messy real-world process and a piece of software expected to behave.
At Trilogy he shared a desk with Cyrus Massoumi. The furniture turned out to be consequential. In 2007, Ganju, Massoumi and physician-turned-consultant Oliver Kharraz founded Zocdoc in New York. The first version was narrow by design: a booking service for dentists in Manhattan. Ganju served as the first CTO, personally built early iterations and oversaw development through many of the company's formative years.
He was simultaneously earning an MBA at Columbia Business School, completed in 2008. It made for a tidy pairing: the engineer who could write the product was formally studying the machinery around it. Zocdoc expanded beyond its original specialty and city into a national marketplace. Ganju's work widened with it, from site architecture and databases to the less glamorous discipline of making a growing operation cohere.
“I've been into computers since I was young and I like creating new products that make people's lives better.”Nick Ganju
A founder changes the weather
By 2018, Ganju was studying neural networks and their mathematical foundations. He saw promise, but not yet a dependable business tool for the work he had in mind. Large language models changed his judgment in 2022. He shifted his attention toward what Zocdoc could build with them and gathered a team around the question.
In July 2023 he became Chief AI Officer. The new title was not merely decorative brass for the office door. Ganju has argued that a founder brings political capital to a technological priority. If a newly hired executive arrives saying that everything must change, colleagues may reasonably wonder whether the slogan came with the luggage. When a founder changes his own remit, the signal travels farther.
He used that authority in an unfashionably literal fashion: he wrote the first cut of Zocdoc's voice agent himself. He also used early versions of the coding tools he wanted the organization to adopt. The point was not that an executive could out-type an engineering department. It was that advocacy is sturdier when the advocate knows where the sharp edges are. His summary was six words long: “I'm not an ivory tower guy.”
Computer science degree, then Trilogy Software in Austin.
Founded Mixonic after working in Silicon Valley.
Co-founded Zocdoc and became its first CTO.
Became Chief AI Officer and built the first voice-agent prototype.
Zocdoc launched Zo; Ganju published its architectural lessons.
The useful art of changing your mind
The product that emerged, Zo, contains a pleasing historical loop. Zocdoc originally helped people escape the phone by moving appointment booking online. Its AI project returned to the phone because many people still use it, and because a technology company does not get to abolish a habit merely by finding it inconvenient. Progress sometimes means going back for the people who did not follow the first route.
The prototype began with a question at an internal hackathon: what if a caller could book at any hour without waiting on hold? Early models proved the conversation was possible and also exposed the danger of trusting a probabilistic system with rigid operational rules. The team's answer was a separation of duties. The language model interprets speech, identifies intent and extracts information. Deterministic services enforce scheduling rules and consult the systems of record.
Speech becomes text and the system identifies what the caller wants.
The model extracts intent and details without owning the rules.
Deterministic services check availability and commit the action.
Another architectural choice made the system rewindable. Rather than repeatedly mutating one conversation record, it stores a sequence of modifications. The current state can be rebuilt by replaying them. An engineer can return to a particular conversational turn, inspect what the agent knew, and test what it did. Partial guesses remain isolated; only the final interpretation persists.
That design unlocked the impatient 100-millisecond guess. It also created an economic problem. In 2024, launching several inference calls for a single spoken thought made each conversation expensive. Ganju's team bet that model costs would fall and continued. By the summer of 2025, he wrote, lower prices and internal optimizations had reduced inference costs tenfold. The indulgent architecture had become practical.
Pool angles and four chords
To understand Ganju's engineering style, it helps to look away from engineering. He is a former Bay Area amateur billiards champion. He took up table tennis lessons as an adult and talked publicly about grip, spin and the value of drilling a skill correctly. He plays guitar and has fronted a semi-professional cover band. In a long conversation with his friend Tim Ferriss, he moved easily among probability, poker, cognitive bias, movie plots and the chord progressions beneath pop songs.
This is not eclecticism for its own sake. Each hobby offers a system that yields to observation. A pool shot is geometry with consequences. A pop song can sound inevitable while resting on a small harmonic frame. Probability turns a hunch into a claim that can be examined. Ganju's gift, as Ferriss has described it, is making difficult subjects seem simple. Simplicity here does not mean removing the hard part. It means locating it.
The same instinct shows up in his management examples. He has used the fictional power company in Monsters, Inc. to explain why a goal needs a number and a route rather than a cheerful command to improve. Even the cuddly monster must eventually face a spreadsheet. It is a very Ganju joke: whimsy invited into the room, then asked to quantify itself.
His public persona contains little of the futurist oracle. He is more comfortable with the useful mechanism, the measurable lag, the first working version. That temperament suits the present AI moment, which has no shortage of declarations and a chronic need for debugging. The interesting question is not whether software sounds impressive in a demonstration. It is what happens when a caller hesitates, changes a detail, or adds three decisive words after everyone thought the sentence was over.
The hard part of a voice is not producing words. It is living correctly in the fractions of silence between them.
The next unfinished sentence
Ganju's remit now extends beyond one product. He is responsible for finding AI opportunities across Zocdoc and helping departments adopt tools that did not exist a few years earlier. Yet the voice agent remains the clearest expression of his method. It begins with a familiar irritation, descends into architecture, and ends with a human measure: does the exchange feel prompt, clear and complete?
There is also a quiet continuity in his return to code. In 2007, Zocdoc needed a founding engineer willing to turn an awkward transaction into software. In 2023, it needed somebody to discover whether a new class of software could handle the awkwardness that remained. Ganju did not solve the second problem by admiring the first one. He opened the tools and made a version.
The builder who once fitted Tetris into a calculator is now fitting context into a conversation. The scale is different. The pleasure seems much the same. Break the system apart. Find the rule. Test the assumption. Listen for the pause. Then, before the silence becomes uncomfortable, begin.