The most useful thing James Zammit built may have begun as the least dignified job in artificial intelligence: calling his own bot. He and co-founder Daniel Gauci Mizzi would change a prompt, alter a model, or nudge a call flow, then pick up the phone. Again. And again. Sometimes hundreds of times. The agent might improve. It might loop, interrupt, hesitate, or discover a new and peculiar way to obey the words while missing the point. The founders were not debugging a clean function. They were auditioning a personality with a telephone line.
That grind became Roark, the San Francisco company Zammit leads as co-founder and CEO. Its software simulates conversations before voice agents launch, monitors calls after they meet actual humans, and evaluates the bits a transcript flattens: timing, turn-taking, tone, accents, long pauses, and whether the exchange achieved its purpose without wandering into a ditch. In Zammit's compact account, “That pain became Roark.” Most origin stories benefit from embroidery. This one benefits from a phone charger.
Before the voices, there were levels
Zammit's route to voice AI began in Malta with the cheerful promiscuity of a young engineer. In 2014, while studying, he helped develop a mobile guide for Majjistral Nature and History Park. The four-student team worked from the MITA Innovation Hub with university and industry mentors, producing an app for both physical and virtual visitors. It was software with mud on its shoes: maps, place, public use.
That same period produced Neon Flux, a 48-level arcade game he made with a fellow math and physics student during the month before finals. Zammit's own description jokes that they wanted to fail their exams. They did manage to make a cross-platform game, and its trailer passed 15,000 views. The anecdote reveals no mystical founder gene. It shows something more dependable: he liked making things enough to choose the deadline with consequences.
There were external signals too. Apple awarded him a WWDC scholarship in 2013. Later came a Facebook F8 scholarship and a Google Cloud architecture certification. A 2015 interview found him in his second year of software development at the University of Malta, already offered a Google internship and arguing that students should experiment beyond the curriculum. His public record from those years reads less like a master plan than a workbench.
Two companies arrived before their moment
Kwaver, a music collaboration app started around 2015, earned backing from Firebase founder James Tamplin and appeared in Firebase's orbit at Google I/O in 2016. Zammit was its engineering lead, building real-time collaboration when the mobile ecosystem still made such work feel faintly illicit. Next came Napo, a chatbot platform that received a Y Combinator Startup School grant in 2019. Chatbots then carried the social charm of an automated parking fine.
Zammit now describes both ventures with useful lack of melodrama. The timing was wrong. The founders were early in their careers. They lacked domain knowledge. One product took roughly a year and a half to reach its first version because the team wanted it polished. By launch, either the market had moved or the perfected answer belonged to a question nobody urgently needed solved.
The lesson he borrows from Michael Seibel is plumbing. You can inspect every inch of a pipe for a leak, or turn on the water and let pressure reveal it. Shipping creates pressure. It replaces the founder's private theory with somebody else's inconvenience. Zammit also discarded a second romantic belief: good products do not sell themselves. Distribution and sales deserve engineering-grade attention, even if neither looks handsome in a code diff.
Finance taught the decimal point to bite
At AngelList, Zammit moved from the loose possibility of early startups into software whose mistakes had surnames, documents, and dollar signs. As a senior engineer, he worked on core portfolio infrastructure while assets managed on the platform grew from $10 billion to $124 billion. He then led Relay, an AI-powered portfolio analyzer that processed thousands of financial documents each month.
Relay turned inbox debris into structured portfolio information. It organized legal documents and company updates, extracted investment terms, generated investor updates, and made the unstructured habits of private markets more legible. The work gave Zammit a close view of large language models in a consequential setting. A model misreading a decimal point was not a quirky screenshot. It could become a compliance problem.
This is the quiet bridge to Roark. Both products take a mess humans can inspect but cannot comfortably inspect at scale, then convert it into structure. In finance, the mess is documents and email. In voice AI, it is conversation: an unruly stream of words, silences, overlap, intention, and tone. Zammit did not leap from one fashionable acronym to another. He kept the shape of the problem and changed the medium.
A conversation refuses to be a unit test
A conventional API test asks whether an input returns the expected value. A voice agent can reach the correct result and still produce a dreadful exchange. It may interrupt the caller, misunderstand intent, ignore a policy, or wait ten seconds before answering. The transaction completes; the person decides never to call again. Text can report what was said while concealing how thoroughly the moment sagged.
Zammit calls conversation-level testing Roark's breakthrough. The platform runs complete simulated calls across hundreds or thousands of scenarios, with varied personas, accents, languages, backgrounds, and behavioral profiles. In production, it monitors and evaluates calls on built-in and custom metrics. The important conceptual move is from checking isolated outputs to examining the whole exchange. Software has always enjoyed pretending the user is an edge case. Voice makes the pretense audible.
Roark joined Y Combinator's Winter 2025 batch. Zammit says the company entered the three-month program processing no calls and ended it analyzing more than 10,000 each week. The distance between those numbers contained hundreds of customer conversations, constant iteration, and the ordinary emotional weather of a young company. The founders had finally turned on the water.
A national-park app and a 48-level arcade game put student code into public hands.
Kwaver and Napo supplied two early lessons in timing, domain knowledge, and distribution.
AngelList Relay made financial inboxes structured and AI systems accountable.
Roark moved from call analytics into simulation, evaluation, and a continuous quality loop.
The six-person orchestra
By 2026, Roark was described as a six-person company serving large organizations and processing millions of call minutes a month. Its sales recipe was stubbornly terrestrial: targeted outreach, referrals, useful published material, inbound interest, demos, pilots, contracts. No enchanted growth lever, just a product surviving contact with buyers.
Inside the company, the team uses more than twenty agents of its own. One reads the week's pull requests, drafts a changelog, and takes screenshots. A human reviews the package on Monday and sends it. Several hours of product-manager labor become about thirty seconds of attention. This is automation without costume jewelry: give the machine the repetitive draft and keep a person at the point of judgment.
The small-team model fits Zammit's career. His projects repeatedly turn unruly material into something a compact group can act on: park information into a mobile guide, music into a collaboration layer, financial paperwork into portfolio data, conversations into tests. Each system is a modest machine for recovering attention.
Thousands of voices, one old question
Zammit expects voice AI to move from demos into ordinary infrastructure. Specialized agents will answer support calls, qualify leads, make bookings, collect payments, and perform jobs whose failures affect real businesses. His public vision is equally operational: every interaction, simulated or live, should be tested, measured, and improved automatically. Teams should know when a release creates a regression and what changed without listening to thousands of calls.
He describes the ambition as making Roark the quality layer behind the voice AI industry. The grand phrase becomes credible only when returned to the founders' original scene: two engineers phoning their software yet again because there was no easier way to know whether Tuesday's tweak had spoiled Monday's work.
There is a decade of founder education inside that call. Ship before polish grows sentimental. Go where the customer is. Treat sales as part of the product. Measure the experience, not merely its transcript. Turn each failure into a test. The future may indeed contain thousands of voices. Zammit's contribution is the unfashionably adult suggestion that somebody should listen to them first.