Since July 2025, I have been working as an MX Manager at Soundmind. On a single child safety product, I built all four layers myself, from the Android native code on the device to the collection server, the parent app, and zero-downtime deployment.
The problem I wrestled with longest was not losing location data on devices where a management policy keeps mobile data off and logs cannot be read. After funneling data opening and release through a single gateway and correcting for the power-saving timer delays I measured in the field, the transmission success rate reached 100%. On the server, I changed the pipeline to drain and store incoming locations atomically and quarantine failures separately, so a single broken record no longer halts the entire load.
In the parent app, I built the logic that interprets noisy coordinates into places stayed and travel routes, eliminating the problem of indoor stays splitting apart and baseless movement guesses appearing on screen.
At WIGTN, a five-person AI development and research crew, we build AI systems that run in real environments. On WIGVO, which performs two-way real-time interpretation over the ordinary phone network, I handled the relay server and deployment infrastructure, recording a median latency of 555 milliseconds on real calls and zero echo loops across 147 calls. The paper was accepted to ACL 2026 System Demonstrations.
With the same team, we released two open-source plugins covering Claude Code's parallel agent orchestration and Codex's Evidence Contract, and took second place at the Snowflake AI & Data Hackathon Korea 2026.
At the company, I have built one product end to end; on the team, I have explored how to put AI into the development process itself. Connecting the two, I want to work on redesigning how development and operations are done, not just using AI as a tool.