Live-streaming platform

What was at stake

The company set out to build an interactive consumer live-streaming product. Creators go live from a webcam and a PC; an audience watches, chats, reacts and sends virtual gifts bought with real money, all in real time.

A product like this has three ways to fail in public. If the video stalls or lags, viewers leave. If the platform goes down, it is not degraded, it is off, in front of everyone at once. And if the gift economy gets a balance or a payout wrong, it is someone's earnings. On top of that, the company had to grow from a three-person codebase to a team that could keep shipping without breaking any of the above.

What was done

Roy co-founded the company in 2019, wrote the first line of code, and led engineering as CTO for six years.

In the early years that meant building most of the system directly: the API, the real-time layer, the first player and the streaming pipeline. As the company grew, the role moved to setting the architecture and engineering standards, and to hiring and leading the team that ran the product day to day.

Built to stay up and to get money right

Two decisions carried most of the weight, because they are where a live platform loses users or loses trust.

Fast where it can be, exact where it must be. Chat, presence and reactions run on a real-time layer (Socket.IO with a Redis adapter) that scales by adding instances, so a room can go from quiet to crowded the moment a popular creator goes live. Anything that moves money (top-ups, gift balances, creator payouts) runs on separate, durable, idempotent paths, so a retry or a spike cannot double-charge a viewer or short a creator. Drawing that line early is what let the product stay both responsive and correct under load.

Close to a global audience. The platform ran in two AWS regions, one in the US and one in Asia, with adaptive-bitrate video so each viewer's player drops quality gracefully rather than stalling. Viewers far from one region were near the other.

Underneath, the platform is one TypeScript codebase with a typed API shared end to end, so a change to the API surfaces in every client before it ships. Errors and behaviour were monitored from early on, and moderation, rate limits and abuse controls were built in alongside chat and streaming rather than added later. In its later years the platform added AI-powered features on AWS Bedrock.

Growing the team through COVID

Engineering grew from three people to around twelve while the company as a whole grew to around fifty. The standards put in place along the way (end-to-end types, clear module boundaries, testing, CI/CD, and a preview deployment for every feature branch) were what let twelve engineers change the system as safely as three had.

Most of that hiring happened between 2020 and 2022, when very few people were willing to change jobs or relocate. The team was built where the talent was: first in China, then in Taiwan, and from 2023 in the Netherlands. Keeping a distributed, multi-timezone team together and shipping through that period was as much of the job as any architecture decision. The repository holds more than eleven thousand commits from dozens of contributors over those years.

What changed

A prototype written in 2019 became a global live-streaming platform with a company of around fifty behind it, running across the US and Asia. In 2024 the company was acquired by an international group. Roy continued to lead engineering through 2025, across the integration that follows an acquisition, with the platform and the team both holding up through the change.