David Kang and the art of taking boring seriously

David Kang’s company began with a whiteboard in his one-bedroom apartment. That is an awkwardly small stage for a business concerned with the American tax system, but perhaps an appropriate one. Taxes have a habit of spreading across whatever surface is available. Kang and his co-founder, Paul Koullick, set out to make that burden easier for independent workers to carry.

The apartment belongs to Keeper’s 2018 beginnings, as Koullick later recalled them. Kang now leads the company as CEO, following years as its chief operating officer. The changing title gives his career a convenient milestone. The more interesting continuity is the work underneath: taking a complicated financial process, examining its troublesome details, and finding out which parts can be made easier.

There are founders whose businesses require a long explanation of why the problem matters. Kang chose a subject that needs little introduction. Even people who enjoy earning money can become remarkably busy when asked to account for it. Keeper has built its business in that gap between having financial obligations and knowing how to handle them.

A trader changes the subject

Before Keeper, Kang studied mathematics and economics at the University of Chicago and worked as an equity options trader. His career included PEAK6 Investments, where he also held an analytics and execution leadership role. The background puts him in familiar territory when the conversation turns to models, probabilities and the quality of a decision’s inputs.

Trading and tax preparation serve different customers. Still, both require care about what a number means. A transaction can be recorded perfectly and understood badly. A model can produce an answer without settling whether it was asked the right question. Kang’s subsequent technical writing makes those distinctions unusually tangible, down to the obscure language on a bank statement.

Kang and Koullick were childhood friends and debate partners. Their company brought finance and product experience into the same room. The business they chose concerned workers whose earnings arrived outside the familiar simplicity of a salary. The ambition was practical: make it easier to identify business expenses and get through the work of filing.

It is an appealing reversal of the usual relationship between customers and financial complexity. Often the customer is expected to become more knowledgeable before the service becomes useful. Keeper’s premise asks how much of that work the service itself can absorb. That question has followed Kang from the company’s early operations into its current product.

The backend had cells

Keeper joined Y Combinator’s Winter 2019 batch. For a young company, an accelerator provides a setting in which an idea must start surviving contact with customers. Keeper’s early implementation was modest: Kang later described a backend built largely from Google spreadsheets. The machinery behind a financial technology company can look surprisingly like the machinery behind a school fundraiser.

His early customer outreach included Craigslist posts and conversations with rideshare drivers during rides. Those encounters had a useful economy. A driver could understand the proposed benefit without sitting through a theory of the future of work. The company had to explain itself in terms that fit the customer’s day.

“Overengineering is a common trap.”

David Kang, 2020

Kang’s recollection gives the spreadsheet an editorial function as well as a technical one. It removes the comforting idea that sophistication must precede usefulness. Early service delivery can teach a founder which problems actually need sophisticated machinery. A beautiful system built around the wrong assumptions merely gives those assumptions better accommodation.

The distinction matters particularly here. The obligations facing a tax customer do not become provisional because the company serving them is young. The goal is to learn quickly while taking the task seriously. Kang’s account of those early experiments is valuable because it keeps the learning process visible, rather than polishing it out of the company’s history.

Money for the rest of the job

In April 2021, Keeper announced a $13 million Series A. Participants included e.ventures, Matrix Partners and Foundation Capital, alongside individual investors including Jack Altman, Jeff Seibert and Ilya Volodarsky. Mathias Schilling and Jake Jolis were to join the board.

The stated purpose of the round was to build a more complete tax platform for gig workers and creators, combining advice, expense management and filing. That is a revealing progression. Finding an expense helps with one part of the customer’s problem. Getting a return prepared and filed carries the service further toward the outcome the customer came for.

It also expands the responsibility. Each additional step introduces another place where information can be missing, a question can be misunderstood or an exception can require attention. Capital gives a company room to build. It also gives it more promises to keep. For an operations leader such as Kang, those promises eventually become processes, staffing decisions and work that somebody has to finish.

The bank statement fights back

Kang’s explanation of automatic write-off detection begins with an uncooperative object: transaction data. Merchant names can arrive truncated or crowded with banking information. Purchases can appear as pending and later finalized. A familiar brand name can refer to different businesses. The system must establish what happened before it can sensibly ask what that purchase means.

His account describes a combination of search, rules, machine learning and user feedback. The engineering has several jobs, including identifying merchants, categorizing purchases and handling duplicates. Knowing something about a user’s work adds context. A category alone cannot supply everything the system needs to understand a transaction.

This is where Kang’s story becomes more specific than a general claim about AI doing paperwork. The product has to work through ordinary ambiguity. Its problems emerge from the way people spend, banks record and descriptions change. Much of the achievement lies in resolving enough of that confusion that the customer can stop thinking about it.

Eight points, many small corrections

By December 2021, Keeper’s use of customized GPT-3 had become a public example of model fine-tuning. The reported transaction-classification accuracy moved from 85 percent to 93 percent. About 500 new training examples were being added weekly. These were historical results for a particular classification task, rather than a measure of the correctness of an entire tax return.

There is something refreshingly workmanlike about that account. The improvement arrives through repeated additions and corrections. Five hundred examples become a week’s work. A week becomes another opportunity to examine the result. The apparent magic acquires a schedule.

In his 2024 discussion of AI implementation, Kang emphasized beginning with available models, testing useful applications and recognizing where people still need to be involved. The degree of accuracy required changes the design of the service. A tax question calls for particular care. His approach makes room for experimentation while preserving a reason to examine the answer.

That is a useful way to understand his place in Keeper’s development. His public explanations spend time on implementation: what to try, where it falls short, and how to improve it. For a customer, the payoff is much less abstract. The relevant question is whether the service can help with their situation and carry the work through.

The operating job becomes the CEO job

Keeper publicly celebrated Kang’s succession to the CEO role in May 2025. Koullick pointed to his co-founder’s work in operations and AI as preparation for the next stage. It was a handover within a company they had built together, years after the apartment and whiteboard.

The present-day service pairs AI preparation with human tax expertise. Its recruiting materials describe a goal of giving customers access to an accountant who understands their particular situation. They also describe a remote-first organization with a San Francisco office. Those details suggest the shape of Kang’s operating problem: combine software, specialist work and a distributed team around a service customers experience individually.

His public conversations have followed that wider remit. An August 2025 podcast appearance addressed tax preparation for high earners and small business owners. A Blinkcode episode announced for March 2026 put seasonal staffing, technical recruitment and a vision of more automated filing on the agenda. That announcement described team expansion of five times during tax season and an ambition for fully AI-driven filing within two to three years.

A timetable for an ambition remains a timetable. Delivering it will require the quieter work that has occupied Kang all along. His official biography supplies two diversions, tennis and chess, but the professional pattern is already clear enough: look closely at the details, build a system, then keep improving what it does for the person using it.

For a company handling taxes, the desired ending is wonderfully undramatic. The questions get answered. The documents get checked. The return gets filed. The customer goes back to the rest of their life, with no particular wish to admire the machinery. Kang has chosen a business in which a successful performance can end with everyone happily changing the subject.