Uncertainty
I model volatility first
Before training anything, I define the uncertainty structure so decisions stay valid when inputs drift.
3 uncertainty geometries tested per high-stakes model
I design quantitative systems that remain stable when assumptions fail.
Decisions must survive deviation, not just expectation.
Formalize the decision. Then build the model.
Different vocabularies. Same mathematical backbone.
Before training anything, I define the uncertainty structure so decisions stay valid when inputs drift.
I use explicit cost curves and threshold rules to choose operating points, instead of chasing one benchmark score.
Every recommendation is paired with assumptions, limits, and validation evidence so teams can trust the output.
Uncertainty is not noise. It is structure waiting to be modeled.
Drag risk tolerance and watch the recommended policy update in real time.
Estimate applicant income for digital onboarding where classic documentation is unavailable.
Data is anonymized and noisy, with high-cardinality behavioral features and proxy leakage risks.
Compared linear and tree families with interaction analysis, feature-importance checks, and residual diagnostics.
Selected the top-ranked gradient boosting configuration and applied bias correction for TL-scale projections.
Built a deployable decision system that keeps interpretability while improving prediction quality in onboarding flows.
From masked behavioral data to validated model selection and calibrated income estimates.
Designed a decision system, not only a predictor: model choice, calibration, and governance were integrated together.
Masked applicant + behavioral attributes with 1 merged contextual rent signal.
Log-income transformation and engineered interactions improve linear separability.
0 direct salary fields used; external validation is still required for production drift.
10-fold CV, test-set check, residual diagnostics, and smearing correction.
Generate radiotherapy plans that remain clinically valid when biological parameters shift from nominal assumptions.
Tumor coverage and organ-at-risk thresholds must hold even under adversarial uncertainty realizations.
Formulated robust counterparts with geometric uncertainty sets and compared conservatism profiles.
Selected the operating point at the knee of the conservatism curve to balance reliability and plan quality.
Produced robust plans with controlled trade-off against objective value while preserving clinical safety.
Visual comparison of uncertainty geometries and robustness cost under clinical safety constraints.
Translated mathematical rigor into a clinical decision policy where safety is a hard constraint, not an afterthought.
Voxel-level planning outputs evaluated under geometric uncertainty perturbations.
Uncertainty can be represented by Box, Budgeted, and Spatial sets.
Higher conservatism increases objective values; trade-off must be clinically accepted.
DVH overlays and 6-point gamma trade-off stress test.
Daily branch performance reporting relied on repetitive Excel refresh operations and manual formatting.
Manual exports and pivot refreshes introduced delay and inconsistency before management review.
Mapped the process as an IE time-study and decomposed reporting tasks into refresh, validation, and delivery stages.
Built a SQL-connected Power BI dashboard with scheduled refresh and standardized KPI definitions.
Reduced reporting latency and improved consistency for branch-level decisions in daily management routines.
Operational reporting transformed into a live monitoring surface for leadership decisions.
Converted reporting from a manual routine into a decision infrastructure with measurable time savings.
Daily deposits, loans, margins, income-expense, and branch profitability signals.
KPI definitions remain standardized across branches and reporting periods.
Bank confidentiality required anonymized screenshots and no external data export.
Parallel checks against legacy Excel process before full operational handover.
Translate sustainability ambitions into operational targets that can be managed like business KPIs.
Reporting burden and fragmented indicators create execution noise across departments.
Built a structured framework using SWOT, SMART goals, benchmark practices, and change-management mapping.
Prioritized energy efficiency, focused ESG data architecture, training, and circular card transition.
Defined a bank-ready sustainability operating model with explicit milestones and accountability channels.
Strategy translated into measurable ESG execution metrics with time-bounded milestones.
Reframed sustainability from narrative reporting into a target-driven operating system for banking execution.
Integrated annual report evidence, interview findings, and benchmark company practices.
Regulatory direction and stakeholder pressure continue to increase ESG reporting demand.
Not all ESG indicators have equal signal quality; prioritization is mandatory.
SMART goals mapped to KPI table and cross-checked against external best practices.
SQL data extraction. Automated Excel workflows. Daily and monthly KPI reporting for interest rates, income-expense, and volume.
Solar installation design via SketchUp and PVCase. Performance simulation with PVsyst.
Production tracking via SAP and Logo Tiger 3. Data analysis and visual reporting.
Financial analysis, reporting, account management.
Distributionally robust approaches, adaptive optimization. Applications in financial risk, supply chains, healthcare.
How does the prediction change the decision? What happens when it's wrong? ML within institutional systems.
Financial systems as optimization under asymmetric loss. Portfolio construction, credit risk, risk architecture.
Healthcare planning and portfolio allocation share mathematical DNA. Same structure, different vocabulary.
I am building toward environments where mathematical precision meets institutional consequence. Advanced study in operations research or financial engineering is the next structural step.
Formulate well.
Model precisely.
Decide responsibly.