Jacky Koh and the work hiding between the apps

The Pokémon were the easy part. In 2016, Jacky Koh and Daniel Vassilev had an app called PokeWhere that showed players where to look. Open a map, find a creature, head in the right direction. There was something pleasantly direct about the arrangement. Software supplied the information; a human did the walking.

Years later, Koh would help build a company whose pitch involves handing more of the walking to software. Relevance AI lets businesses create agents, give them tools and arrange them into teams. The journey between those products is less tidy than a founding legend, and more interesting. It passes through statistics, corporate machine learning and the discovery that a business can possess plenty of information while remaining remarkably bad at using it.

Koh’s projects keep returning to that awkward distance between having something useful and being able to act on it. A location on a map. A pattern in customer comments. A task that everyone knows how to do, but nobody wants to spend another Monday doing. His career has changed the machinery around that distance.

A Pokémon finder with a statistics habit

PokeWhere belonged to an earlier chapter of Koh’s partnership with Vassilev. Their consumer apps attracted millions of users, and PokeWhere reached number one on the App Store. A Pokémon-location app may sound like a light apprenticeship for enterprise software. It does, however, introduce the builder to a fairly stern judge: someone who downloads a product because they expect it to work immediately.

There was another kind of scrutiny at university. In May 2018, Koh and Saksham Yadav took the top prize in the EY Data Science Challenge. The contest drew 374 participants across Australia. The two students were the only undergraduates among the finalists, and each received a cash prize and an EY internship.

Koh was studying actuarial studies, majoring in statistics. The challenge involved analysing global quality-of-life measures and predicting relationships with broader trends and education. It was a problem of extracting a useful relationship from a collection of numbers, rather than finding a creature around the corner. Both require a decent answer to the question of where to look.

Dan Brockwell, who met Koh at university, later recalled a student already building apps and advising companies on machine learning while others concentrated on securing corporate internships. Koh also served as president of the UNSW Data Science Society. The picture is of someone learning the mathematics and finding occasions to put it to work.

The information left in the comment box

Before Relevance AI, Koh worked in data science at EY and led machine learning at IAG Satellite. Corporate experience gave him a different view of the information problem. Businesses were producing text and other unstructured material that their existing analysis often left behind.

A spreadsheet welcomes a number. A customer’s paragraph is a less obedient guest. It contains useful meaning, but it rarely arrives in the tidy shape that a database would prefer. Koh’s early account of Relevance AI centred on making that material usable alongside conventional business data.

The company began in 2020. By December 2021, Koh and Vassilev were explaining a vector experimentation platform in an AusBiz interview after an A$4 million funding round. The interviewer needed the vocabulary unpacked. Fair enough: “vector” is a word with an unusually good ability to make an ordinary business problem sound like homework.

The practical idea was to compare information through numerical representations of meaning. By 2022, Koh was explaining how differently worded feedback could be grouped together, then connected with business measures. He also stressed the next step: people with knowledge of the business had to examine the groups and decide whether the analysis made sense. Finding a pattern was only part of making it useful.

Three founders, several ways of working

Relevance AI brought Koh together with Vassilev and Daniel Palmer. The link with Vassilev went back to a Year 8 maths classroom. That is a long time to know a collaborator before a company starts asking you both to make expensive decisions.

In May 2025, Koh described the partnership as complementary. Both could code, sell and speak to investors. Their differences lay more in how they worked: Vassilev supplied greater day-to-day consistency, while Koh felt suited to concentrated bursts and difficult moments. He described himself as calm in an emergency. Shared capabilities did not require identical temperaments.

There is a useful modesty in that account. A shared CEO title can invite elaborate theories about organisational design. Here, part of the explanation is simply that two people know when the other is better placed to take the lead. Their history gives that judgement something firmer to stand on than a diagram of responsibilities.

When the diagrams became the product

The company’s move into agents was visible by its December 2023 Series A. It raised A$15 million in a round led by King River Capital, with participation from Peak XV’s Surge, Galileo Ventures and Insight Partners. At the time, it reported 6,000 clients signed up over the preceding three months, running more than 250,000 tasks.

The examples were recognisable office jobs: customer inquiries, outbound sales and market research. The ambition had expanded from analysing information to completing work with it. For a small business, the appeal was the chance to build a team of agents without first assembling a large engineering department.

As customers built more elaborate systems, they began drawing diagrams of their own AI workforces. That behaviour suggested an interface. The Relevance AI team developed Workforce, a visual canvas where users could see their agents and manage how they collaborated.

It is a revealing product detail. People had invented a way to understand what they were building before the software gave them a suitable place to understand it. A sketch became a clue. The canvas brought agents, tools and relationships into the same view, including decisions about delegation and escalation to humans.

In May 2025, Relevance AI announced a US$24 million Series B led by Bessemer Venture Partners. Its products included Workforce and Invent, which generated agents from descriptions in ordinary language. For Koh, the intended builder remained the person who understood the job, even when that person did not write software.

A little mischief in the pixels

The agent business has a visual problem. An autonomous system is difficult to put in a photograph. Relevance AI’s pixel characters give its software inhabitants a face without requiring them to resemble a colleague who has accidentally wandered into a video game.

Koh has discussed those characters and their Pokémon connection. The earlier app chapter makes the association feel rather natural. A business product can borrow a little of the pleasure of collecting and arranging characters while asking its users to do quite serious work.

The same year, Koh, Vassilev and Palmer appeared on the Forbes 30 Under 30 Asia list. Koh responded on LinkedIn with a joke that the publication should have used their agent photos. The honour could apparently survive a less solemn headshot.

The joke supplies a small but welcome clue to the person. Public founder narratives often flatten into a succession of funding announcements. Koh’s post makes room for play, even at a moment designed for a polished announcement. The pixel workforce had, at least briefly, applied for the founders’ own publicity job.

The cost of going faster

By August 2026, Koh was discussing another difficulty at Growth Summit Sydney. AI could make people faster while leaving a team pointed in several directions. He acknowledged that his own team had shipped projects that drifted from its customer mission. The admission adds some friction to an industry accustomed to celebrating speed.

“AI is making everyone super fast.”

Jacky Koh · Growth Summit Sydney, covered August 2026

His point concerned alignment: the goals, judgement and shared understanding that make quick execution useful. More output does not automatically settle the question of what should be produced. A team still needs to agree on the job it is trying to finish, and how that job serves its customers.

That concern also reaches back to his early data work. Someone had to interpret the clusters. Someone now has to decide what an agent should do. Koh’s tools offer ways to move faster; his accounts of using them keep a place for the expert who knows whether the result belongs in the business at all.

Keeping the next job in mind

Koh’s public profile now also identifies him as the founder of Arg Labs. At Arg, he describes a shared workspace for people and agents. The focus is the material that gets scattered between tools: briefs, corrections, recordings and finished work.

The proposition is that these things should accumulate. A recurring job can leave an artifact that someone can inspect; a correction can inform its next run. Arg describes boundaries that stop autonomous work and ask for a person’s decision. These are statements of product intent, with a familiar concern behind them: useful context should reach the next task.

From a Pokémon map to business data to agent teams, Koh has spent years working on the passage from information to action. Now the question includes what action leaves behind. If the next Monday begins with the last Monday’s lessons still available, the software has done something an office worker can appreciate. It has saved them from explaining the same thing again.