the intelligent layer for global voice.

English TTS is mature; every other language degrades on translation, and buyers aren't multilingual enough to catch it. OnePin knows which model wins for each language, so businesses can scale to markets they couldn't serve before.

A football match in English, headed for a global audience. Before vs. after OnePin.

The bet behind it. The obvious lever is faster generation, produce the ~9,000 files quicker. But the bottleneck wasn't producing audio, it was judging it: deciding which voice was right was the slow, human part. So OnePin isn't another generator, it's a decision layer that picks and evaluates, and the defensive bulk runs disappear.

Straight to the people living the pain, then prototype fast and lock direction in a month.

OnePin is, at its core, an opinion about AI output. My job was to turn "is this voice good?", a gut call, into a flow a team and an agent could run the same way every time. It's node-based with a chat agent on top, so you can wire the steps by hand or just describe what you want, the same judgment runs underneath either way.

Instead of one hidden, one-size-fits-all pipeline, I exposed every correction and validator as a node the user, or an agent, can switch on or off. An AITuber and an enterprise broadcaster don't share the same bar for "good", so the judgment had to be tunable, not baked in.

One surface where a human, or an agent, drives a production end to end: describe the job in chat, watch it resolve as a live node graph, script in, corrected, validated, and ready to ship. Four design decisions made that judgment trustworthy:

I learned the problem by living in it. Three weeks on-site in Seoul, whiteboarding the pipeline with the founder, a voice researcher, I learned to hear what "good" actually means: where a model's accent drifts, where a player's name is mispronounced, where a take is technically clean but lands wrong. Those sessions are why the validators are the ones a real localization lead would trust, not a generic quality score.

For me the Figma file was the starting point, not the deliverable. I designed in clean components and standard layout patterns on purpose, so the interface could be built straight from the design, then I took it the rest of the way myself.

From there I moved into Figma Make and Claude Code to turn the chat-agent and node interface into working code, and wrote the pipeline, phoneme injection, normalization, and the validators as Python skills a person or an agent can run. AI removed the execution bottleneck; the judgment and the taste stayed mine.