Abstract
Most AI companies build services: APIs, endpoints, models-as-a-product. At Klang Research, we have taken a different path. We build interfaces — tools that put AI directly in front of people, in contexts where the interaction itself is the experiment. This paper explains why, and what we have learned from doing it.
The interface thesis
There is a fundamental difference between an AI service and an AI interface. A service processes inputs and returns outputs. An interface creates a relationship between a person and a capability. The service asks "how accurate is the model?" The interface asks "does this actually help?"
We believe the most important lessons in applied AI come from the second question. And you can only answer it by building things people actually use.
This is not an argument against benchmarks or API-first products. It is an argument that they are insufficient — that the interesting problems only emerge when real people interact with real systems in real contexts.
Voice: what dictation teaches us
Klang Dictate started as a straightforward engineering challenge: build a Nordic-first dictation system that runs on-device. The technical requirements were clear — low latency, high accuracy, system-wide integration on macOS.
But the moment we put it in front of users, the interesting questions shifted. Accuracy matters, but so does feel. A system that is 95% accurate but pauses awkwardly at clause boundaries feels worse than one that is 92% accurate but flows naturally. Latency is not just a number — it is the difference between "talking to your computer" and "waiting for your computer."
These are not insights you get from a benchmark. They emerge from the interface.
We learned that vocabulary customisation is not a feature — it is a necessity. Every user has a handful of domain-specific terms that trip up generic models. The technical challenge of on-device custom vocabularies is real, but the user insight that drove it only came from watching people actually dictate.
Video: what live meetings teach us about real-time AI
Klang Video pushed us further. Dictation is one person, one microphone, one language (mostly). Video meetings are multiple people, overlapping speech, language switching, and a context that evolves minute by minute.
The first thing we learned: post-meeting summaries are table stakes, not the product. Every tool does this. The hard problem is making AI useful during the meeting — and the definition of "useful" changes throughout the conversation.
At minute five, a live transcript is helpful. At minute thirty, nobody is reading the transcript anymore — they want a summary of what they missed while they were thinking about their own point. At minute sixty, they want action items, not a summary.
This temporal dimension of usefulness is invisible if you are just building a transcription API. It only becomes apparent when you build the interface and watch people use it.
We also discovered that multilingual handling in Nordic contexts is not an edge case. It is the default. A meeting between Swedes, Norwegians, and Danes will drift between languages constantly, sometimes mid-sentence. Models that treat language as a per-utterance classification fail here. The interface taught us that language in Nordic meetings is a spectrum, not a switch.
Community: what an open challenge teaches us
Knack Klang is our most unconventional project — a public challenge where anyone can test their transcription skills against difficult Nordic audio clips. It is not a product. It is not a service. It is an interface between our research questions and a community of people who care about the same problems.
What we learned surprised us. The clips that humans find hardest are not the same ones that machines find hardest. Certain dialectal features that trip up automated systems are trivially easy for human listeners, and vice versa. This asymmetry is a goldmine for model improvement — it tells us exactly where our models are failing in ways that aggregate metrics cannot.
The challenge format also surfaces edge cases we would never have designed test sets for. When hundreds of people attempt the same clip, patterns emerge: systematic failures on specific phonetic transitions, unexpected difficulty with common words in uncommon contexts, consistent confusion between dialectal variants.
These are not the kinds of insights you get from running a benchmark. They come from building a community interface and letting people surprise you.
The common thread
Across all three projects, the pattern is the same: the interface is the experiment. We did not build Dictate, Video, and Knack because we had finished our AI research and needed products to ship. We built them because they are the research.
Every user interaction is a data point. Every frustrated pause, every delighted correction, every unexpected usage pattern teaches us something that no benchmark captures. The interfaces are not downstream of the models — they are upstream. They tell us what to build next.
What comes next
We are doubling down on this approach. The next phase of Klang Research is about closing the loop tighter — making the insights from our interfaces feed back into our models faster, and making the improved models create better interfaces sooner.
The gap between "AI that performs well on benchmarks" and "AI that helps people" is still enormous. We believe the fastest way to close it is to keep building interfaces — keep putting AI in front of people and learning from what happens.
If you want to be part of this, try Knack Klang, test Klang Video, or sign up for the Klang Dictate waitlist. Every interaction makes the research better.