Hi, I'm Olivia.

Independent project · 284,807 card transactions · 2026

Where should the alert threshold sit when a missed fraud costs $200 and a false alarm $5?

Compared logistic regression, random forest and XGBoost on data with 0.17% fraud, then set the alert threshold at 0.06 instead of the default 0.5 from the cost of each error. Sliced performance by transaction amount and found recall drops to about 0.50 on $500+ transactions, where each miss costs most, and logged it as an open risk with PSI drift monitoring on top.

WHY IT MATTERSEvery fraud team lives on this trade-off: an alert costs an analyst's time, a miss costs money. The threshold is a business decision, not a default, and the $500+ gap is the kind of blind spot a monitoring plan has to own out loud.

Independent project · 307,511 loan applicants · 2026

Where should a lender set its approval cutoff when a missed default costs 10× a wrong decline?

Built XGBoost on 96 features from 10 credit-history tables and set the cutoff at 0.10, close to the analytical optimum of 0.091. A fair lending audit then found age fails the four-fifths rule (DI 0.65); dropping the gender feature cost 0.4% PR-AUC and closed most of the gender gap.

WHY IT MATTERSLenders and model risk teams ask the same two questions: where should the cutoff sit, and who does the model treat unfairly. This answers both with numbers a validator can check.

Carnegie Mellon · Time Series Forecasting with Python · 2026

Does a forecast beat the simplest alternative, and keep winning on periods it has never seen?

Two case studies. On JPMorgan Chase monthly prices, the best smoothing model turned out to be the naive forecast, and only a combination of models beat it; macro indicators that looked correlated on levels lost their link once trends were removed. On drug sales, a rolling backtest over four 12-month windows ranked multiplicative Triple ETS first in every window.

WHY IT MATTERSModel validation starts with one question: does it beat the naive benchmark on data it has never seen? Here the honest answer was sometimes no, and saying so is the job.

Independent project · AI governance · 2026

Does a content-safety model catch what the policy says it should?

Built a content-safety classifier, mapped its outputs to policy severity tiers, and wrote a 69-case red-team suite to test it. The baseline passed 53 cases and missed every coded-hate case, which is exactly the kind of gap a review needs to see before a model goes live.

WHY IT MATTERSAI governance needs tests, not just policies. A red-team suite turns "is this model safe?" into a list of cases it passes and fails.

Top-five Korean life insurer · Basel III operational risk system build · 2023

Built the indicator and control layer of a Basel III operational risk system

Owned a third of the technical workstream on a four-month build. Defined 9 key risk indicators across 9 departments, each specified down to formula, source system, interface ID and target table, with statistical thresholds set on a 24-month moving average at 95% and 99% confidence. Wrote 13 of roughly 40 screen specifications with role-based permission matrices and approval workflows, and validated a decade of loss data across two legacy systems left over from a merger. Detection lag went from quarterly to monthly.

WHY IT MATTERSSame problem as tuning a fraud rule: catch real breaches without flooding people with alerts. The thresholds came from 24 months of data, not from guesswork.

Deloitte · IT audit, ICFR / SOX, SOC review · 2023 to 2026

Tested the controls around transactions at 23 companies, from lenders and insurers to e-commerce and gaming

IT general and application controls testing for financial reporting audits: approval limits, input validation, calculation logic, and interface reconciliation, plus user access, segregation of duties, disaster recovery drills and backup restore tests. Reviewed SOC reports to evaluate outsourced service providers. On a 722-control joint audit, wrote the test plan and Python scripts that cut manual sampling by 40%.

WHY IT MATTERSPayments and fraud teams run on these controls: approval limits, input checks, reconciliations, access rights. I know where they break because I tested them at 23 companies.

Client names and findings remain confidential.

Deloitte · practice methodology · AI governance

Co-built the practice's first AI audit method: 77 controls across the model lifecycle

Built on a team of three. Four risk domains, data, AI model, AI system and enterprise, crossed with a plan, build, operate and retire lifecycle. Where an AI risk resembled an existing general IT control I mapped it back to that control instead of duplicating it, and every control was classified by whether it needs specialist AI expertise or a generalist auditor can run it. That classification was the point, because without it "who can audit this model?" had no answer and engagement leads had no basis for staffing AI work.

WHY IT MATTERSThis is model risk management from the other side of the table: what to test across a model's life, and who is qualified to test it.

Control contents remain client and firm confidential.

Leading Korean securities firm · first external IT audit

Turned 40+ interviews into an audit method that didn't exist yet

Co-owned the scope that had to outlive us: a risk pool across five audit domains, a quantitative scoring model grading every area from priority to monitoring, and a security architecture for external LLM use. Codified into a seven-chapter manual written for non-specialist auditors, plus Tableau dashboards embedded on the firm's own platform with drill-down for the audit team and aggregate trends for executives.

WHY IT MATTERSA scoring model decided where the audit looked first, and the LLM rules set limits before anyone used the tools. That is model governance applied to a real audit plan.

A large Korean securities firm · joint IT and information security audit · 2023

Designed controls for MyData, cloud, and robo-advisory with no audit precedent

Led the emerging-technology review on a seven-week joint audit. MyData, cloud and robo-advisory were where fintech partnerships were being built on infrastructure nobody had audited yet, so every control had to be reasoned from first principles rather than pulled from a library.

WHY IT MATTERSThis is where fintech risk lives: new products built on partner infrastructure with no playbook yet. Every control had to be reasoned out from what could actually go wrong.

Top-three Korean financial group · 7-month engagement

Mapped statutory accountability for 6 executives under a brand-new regime

Korea's senior managers regime had just landed with no domestic precedent, so I researched the UK original, then worked across business and control functions to break executive duties (covering technology, AI, and asset management) into department-level controls. The manuals department heads use today came out of that.

WHY IT MATTERSModel risk and AI governance come down to one question: who is accountable when it goes wrong. I mapped that answer for executives covering technology and AI.

Bank & securities group · post-regulatory-change remediation

Sequenced 10 remediation initiatives, then shipped 18 compliance screens on schedule

After regulators tightened conflict-of-interest rules for dual-role executives, I ran 30 interviews to score and sequence the response, then owned functional design through to live screens. Analysis and delivery in the same pair of hands.

WHY IT MATTERSRules only work once they are built into the tools people use every day. I scored which fixes mattered most, then designed the screens that made them stick.

Financial institution clients · cross-team process design

Traced repeated requirement rework to one cause: six teams, six vocabularies

Compliance, risk, technology, security, strategy, and vendor teams were using the same terms to mean different things, so requirements kept getting rebuilt. I built a shared glossary and cross-check process, adopted on two implementations.

WHY IT MATTERSEvery metric starts with a definition. Getting six teams to use the same words the same way is the unglamorous first step before anything can be measured.

Korean insurance subsidiary · first-time lead consultant

Benchmarked 12 sections of a global IT asset standard against local reality

First engagement as lead: benchmarked the subsidiary against its group standard across 12 sections and reconciled ServiceNow CMDB records against physical assets. The gap analysis and a 7-process ownership redesign became the local team's remediation plan.

WHY IT MATTERSYou can't monitor what you haven't inventoried. Checking the records against what actually exists is the habit behind every good reconciliation.

Findings and remediation detail remain client confidential.

Suwon City Hall · Smart City Division · Team Leader · 2021

Built a composite risk scoring model on a 250m grid, and six of its recommendations became city policy

Merged 30+ datasets (40,000+ police call records, floating population, CCTV and lighting infrastructure, demographics) onto a standardized 250m × 250m grid. Multi-criteria decision analysis with consistency-ratio validation set the weights, K-means clustering classified the grids, geospatial analysis in QGIS mapped them. Honorable Mention at the city's public safety big data competition.

WHY IT MATTERSA composite risk score with validated weights, used to decide where money and attention go. Fraud risk scoring asks the same thing of every transaction.

Suwon City Hall · Smart City Division · led a 4-person team · 2021

Built a text classification pipeline that sorted 10,600 complaints at 87% accuracy, cutting manual effort 85%

An end-to-end NLP pipeline: morphological parsing, TF-IDF vectorization, LDA topic modeling that surfaced 23 complaint themes, then a Naive Bayes classifier trained on 2,000 labeled samples and applied to the full corpus. Average resolution time fell 23%, and three neighboring cities reused the framework. Top prize at the Ministry of the Interior and Safety's national public big data competition.

WHY IT MATTERSTriage at scale: sort thousands of cases so people spend their time on the right ones. A fraud operations queue is the same problem.

Suwon City Hall · Smart City Division · Data Analyst · 2021

Designed the metric framework a city allocated relief budget against, then replaced monthly PDFs with self-serve dashboards

Credit-card transactions, mobile GPS foot traffic, and demographics sat in incompatible geographic units, so I built a crosswalk in GIS to join them. Then defined the KPIs the city would actually decide on: revenue decline, foot-traffic reduction, a weekly recovery index, a composite vulnerability score, with segmentation by business type, district, and time slot, and shipped a three-tier self-serve Tableau dashboard that fed executive economic briefings.

WHY IT MATTERSBuilt on credit-card transaction data. Turning raw transactions into the few numbers a decision rests on is daily work in payments risk.

Seoul Metropolitan Government delegate · CES 2022, Las Vegas

Generated 258 leads and closed 3 contracts for a Korean AI startup's US entry

Segmented buyers out of the exhibitor database, then ran 69 meetings delivering English pitches and demos. The contracts and letters of intent that came out of it reshaped the startup's roadmap.

Kyung Hee University · Project Lead · 2022

Ran regression, ANOVA, and time-series decomposition on thirty years of global energy data

Enerdata and World Bank data, 1990–2020. Regression controlling for GDP and population as confounders, ANOVA across blocs, and ARIMA time-series decomposition to isolate the 2008 and COVID shocks. Delivered as a reproducible R Markdown report plus an interactive Shiny dashboard. Global consumption rose 73% over the period; 2020 was the steepest single-year drop in thirty years.

Korea Employment Information Service competition · 2022

Placed top four by public vote with a user-centered employment analytics dashboard

A job-market dashboard built for people actually looking for work rather than for the agency publishing the data. Pipeline through interface, built solo.

Trusty · Co-Founder & Product Lead · 2019 to 2020

Founded a women's health platform on 300 user interviews and hit 120% of its funding goal

A 2017 scandal found harmful substances in most Korean menstrual products, and no app told anyone which ones were safe. We partnered with a testing institute and a university lab to turn chemical analysis into consumer-readable safety scores, then built a recommendation engine on top. Coordinated engineers, medical advisors, and three research institutions; secured ₩50M+ across three programs.

KPMG · Digital Group · Digital Consulting Intern · 2022

Turned 25 departments' requirements into one documented cloud evaluation model

Consolidated cloud HR platform requirements from 25 departments into a central requirements database covering functional needs, integration points, compliance, and UX. Assessed vendors within a 27-vendor review on modules, ERP integration feasibility, and data-migration risk. The framework replaced subjective debate with documented criteria for leadership's decision.