Healthcare AI
Clinical prediction, EHR analytics, medical NLP, calibration, fairness, and safe paths from retrospective evidence toward real clinical collaboration.
Research portfolio
Research that survives scrutiny.
My research approach emphasizes falsifiable questions, honest baselines, reproducible pipelines, explicit failure modes, and a clean separation between what the evidence demonstrates and what remains unproven.
Trustworthy AI
Model evaluation, leakage checks, data-quality contracts, reproducibility, governance, uncertainty, and fail-closed systems.
Real-time AI
Streaming data, online inference, event detection, telemetry, distributed services, and high-consequence decision pipelines.
Standard research template
Every flagship case study is organized so an employer, hospital research team, professor, or grant reviewer can understand what was tested and how strong the evidence is.
| Stage | Question answered | Evidence expected |
|---|---|---|
| Problem | What real-world problem matters? | Stakeholder need, clinical or technical significance |
| Research question | What exactly are we trying to learn? | Specific falsifiable statement |
| Data | What observations support the test? | Provenance, cohort, inclusion/exclusion, quality |
| Method | How is the hypothesis tested? | Features, models, controls, baselines |
| Validation | Can the result survive leakage and overfitting checks? | Holdout design, temporal/OOS testing, reproducibility |
| Results | What happened? | Metrics, confidence, visual evidence |
| Negative results | What failed? | Rejected hypotheses and failure analysis |
| Limitations | What can this work not prove? | Explicit boundaries and external-validity limits |
| Governance | What safety/privacy constraints matter? | Privacy, fairness, access, authority, review |
| Next experiment | What would reduce uncertainty next? | Prioritized research plan |
Research evidence cards
A common schema makes experiments easier to audit, compare, reproduce, and eventually convert into technical reports, manuscripts, grant preliminary data, or product decisions.
Experiment ID
→
Hypothesis
→
Dataset + cohort
→
Baseline
→
Candidate method
→
Validation
→
Decision
Research principle: a negative result is useful when it is measured honestly. I do not treat failed hypotheses as something to hide; they help constrain the next experiment.