I am drawn to problems that look technical on the surface but fail in real life unless you connect engineering with context: infrastructure, adoption, policy, and human behavior.
I treat AI models as one layer in a larger system that must be measurable, usable, and trusted. If something fails, the root cause is often the surrounding workflow, not the architecture.
I start with reality: the data we actually have, how users behave, and the edge cases that break systems. Then I build the simplest design that stays reliable under those constraints.
I care about interpretability, reliability, and what happens after the demo: maintenance, failure modes, and whether the output genuinely helps someone decide.
A single question changed how I think about technology:
if AI is advancing so fast, why do critical problems still
hurt people in predictable ways?
During severe flooding in Vietnam, I started from a naive assumption, “we just need a better predictor”, and discovered the harder truth: forecasting is not only computation.
That spark pushed me toward work where the “last mile” matters: sensing, UX, reliability, and making outputs actionable for real stakeholders.
HFA – Health for All is where my mindset changed:
from optimize the model to stabilize the system
I joined HFA as one of three builders, initially approaching it like many technical projects: define the problem, write the code, optimize the output. I took on the pieces I could execute directly—turning ideas into working prototypes, testing early pipelines, and iterating fast when the system didn't behave as expected.
Once we moved beyond prototypes, the work became less clean and more honest. Medical data didn't fit neat assumptions. Edge cases appeared everywhere. Small errors created outsized consequences.
In our team check-ins, progress didn't always look impressive—sometimes it was hours spent tracing a single bug, validating signals, or rewriting a component because it wasn't dependable enough.
Our turning point came when we spoke with a doctor at the National Hospital of Endocrinology and learned about a patient in Thai Nguyen who couldn't travel regularly for monitoring. The doctor's message was direct: in healthcare, it's safer to consult experts than to rely entirely on AI.
As a three-person team, we made a hard decision: pivot from “AI diagnosis” as the headline to real-time alerts that help connect patients and doctors. That choice forced us to prioritize responsibility over flashiness—and to build for trust, not just performance.
After the pivot, I stopped thinking of AI as the product and started thinking of it as one component in a larger system: user behavior, device reliability, clinical workflow, and communication under pressure.
My work became less about chasing perfect metrics and more about making the system stable, interpretable, and usable in real life—because a health tool that isn't dependable is worse than no tool at all.
