Curious enoughto uncover the real problem. Technical enoughto build the response. Pragmatic enoughto know how much is enough.
I'm Kai, a product engineer with a full stack background. Over the years, my work has increasingly moved beyond implementation into understanding product problems, shaping solutions, and taking them through delivery.
Featured work
I turned repeated lending builds into a configurable platform
Used by a bank for several years, generating several thousand leads and becoming a foundation for later projects.
I turned repeated lending builds into a configurable platform
Used by a bank for several years, generating several thousand leads and becoming a foundation for later projects.
After three to four similar lending applications, the repeated journey was clear: keep the core stable while making branding, wording, credit calculations, and client requirements configurable.
I planned the modules and system boundaries and implemented most of the infrastructure, backend, and frontend. A state machine modeled each lending journey and generated its frontend wizard.
One bank used the platform for several years and generated several thousand leads. Although no second client launched, the code became the foundation for more than three later projects.
The system was stable, but we abstracted ahead of market validation. Today I would deepen modularity only after observing several committed customers.
I added voice to a grounded AI assistant without sacrificing trust
A cross-platform beta for several hundred users, with grounded responses beginning in under ten seconds.
I added voice to a grounded AI assistant without sacrificing trust
A cross-platform beta for several hundred users, with grounded responses beginning in under ten seconds.
Voice needed to feel immediate, but it still had to use the product's slower retrieval and grounding pipeline so answers remained trustworthy.
I owned the push-to-talk flow across React Native and FastAPI, integrating transcription, retrieval, and streamed speech. Users can interrupt at any time; a new recording cancels generation and queued playback.
The closed beta serves several hundred users on iOS and Android. Grounded responses begin playing in under ten seconds, and adoption is being measured before further investment.
Latency was not the only goal. Preserving trust and making recording, waiting, playback, and cancellation predictable produced the better product trade-off.
I built an AI workspace that turned research into editable artifacts
An internal product that teams used long enough to commission a second development phase.
I built an AI workspace that turned research into editable artifacts
An internal product that teams used long enough to commission a second development phase.
Teams manually turned research and planning material into structured artifacts. Stakeholder interviews helped define how those inputs and outputs should connect in one workflow.
As lead developer in a two-developer team, I worked across the React app, Python backend, authentication, file processing, transcription, and rich-text editing. Users could generate, refine, edit, and export artifacts.
The product saw extended internal use, and feedback led the client to commission a follow-up. It is now being phased out as general-purpose AI platforms improve faster than its budget allows.
A standalone model interface is not a durable advantage. The stronger product would connect the tools teams already use and treat models as replaceable infrastructure.
There’s more.
These are only a few of the things I’ve built. The rest is better shared in conversation.