Trilingual HR and recruiting operations professional, working across high volume screening, interview coordination, ATS data hygiene and pipeline tracking.
I keep recruiting records accurate, build the trackers and dashboards a hiring team actually runs on, and coordinate stakeholders through fast moving cycles. The result is pipeline visibility, quicker follow up and hiring readiness that rests on clean data rather than on chasing people.
Greenhouse · Workday Recruiting · Airtable · Excel · Looker Studio · Power BI · Chinese, English, Spanish
Three projects
SQL · Excel · Looker Studio
A recruiting operations dashboard that keeps itself current: SQL into Excel into Looker Studio, so nobody rebuilds it every month. It scores data hygiene and splits each SLA breach by the person who holds the step.
Read the case studyExcel · Power BI
Survey and attendance data cleaned in Excel, then modelled in Power BI for LIDERA, the UN's internal leadership platform. Built so senior leadership can see which events land and where the strategy needs to change.
Read the case studySPSS · R Studio
A time-series study of whether a falling coal share actually lowers CO2 intensity. Built in R over 31 years: stationarity tests, ARIMA selection, then ARIMAX with GDP per capita and a human rights index.
Read the case studyIndex
Case study 01
A recruiting operations dashboard that keeps itself current, built around the question a talent acquisition team is actually asked every week: what is broken right now, and who has to fix it.
Most recruiting dashboards report outcomes, and get read once a month. But the daily work of a TA operations role is data hygiene: requisitions nobody closed, candidates nobody moved, scorecards nobody submitted. So outcomes stay on page one, and the rest of the dashboard answers a different question: where is the data wrong, and who has to fix it.
Page one, executive summary. About thirty five KPIs, each with a target and a plain reading of what the number means.
Page two, ops console. SLA breach by stage, breach by owner role, a hygiene score per recruiter, and live lists of stale candidates and outstanding scorecards.
Page three, pipeline deep dive. Funnel, pass rates, applicants per requisition, source volume against source quality.
Page four, findings. Ten findings, each with the current metric, the target, an owner and an effort estimate.
Marketing hiring managers take 9.3 days to return interview feedback against a five day target, close to twice the next slowest department. Marketing's overall time to fill is 68 days against a 45 day target.
Hiring managers breach SLA more often than recruiters do. Splitting breach rate by the role that actually holds the step changes the conversation from "recruiting is slow" to a named owner.
Eighteen requisitions carry at least one open data issue, and five filled requisitions were never closed in the ATS. The open requisition count leadership sees is wrong.
1,145 applications produced 19 hires. Data and Analytics gets 9 applicants per requisition against a healthy floor of 20, which is a sourcing problem rather than a process problem.
Forty candidates have not moved stage in over two weeks, and forty four scorecards are past the 48 hour SLA.
Pipeline funnel, candidates reaching each stage
SLA breach rate by stage, against a 15% target
Dashed line: 15% target · Scale 0 to 50%
Ten findings in all, each with an owner and an effort estimate. Seven are low effort.
Built on a synthetic dataset for a fictional school network, generated and then reconciled independently so every summary figure ties back to the detail rows. Candidate records use anonymous references, not invented names.
Next: LIDERA EngagementCase study 02
LIDERA is the United Nations' internal leadership platform for staff. Our team built an engagement plan for it. My part was the dashboard that shows senior leadership which parts of the programme are working and which part of the strategy has to change.
The programme was already running events and collecting feedback, but attendance sat in one export and survey responses in another, so nobody could see the two together. The reported numbers looked healthy: 4.53 out of 5, 87.5% saying they learned something new, 91% saying they would recommend it. The number next to them was not: only about 15% of attendees leave any feedback at all.
Reading the dashboard against the member list and the event reports made the shape of the gap clear: the programme could see what happened inside a session, but very little about who was in the room or what happened to them afterwards. So the second half of the work was designing the intake that would fill it.
Shown as rates and averages only. Headcount figures are withheld, and the member reach tile is redacted in the screenshot above.
Next: Determinants of CO2 EmissionsCase study 03
A time series study with a narrow question: if the coal share of the United States power grid falls, does CO2 intensity fall with it, and how quickly?
Three hypotheses: that a lower coal share means lower CO2 intensity in the same year and up to two years later; that the drop is larger when coal is replaced by renewables than by gas; and that the effect survives controls for demand, development and the policy shocks of 2010 and 2020.
The dependent variable is roughly normal, with skewness of -0.54 and excess kurtosis of -1.16, so no transformation was warranted on distributional grounds. The series drifts downward with level shifts, the ACF decays slowly and the PACF cuts off at lag one, which points to differencing once rather than to an autoregressive term.
The independent variables are heavily collinear. Fossil and low carbon shares both correlate with CO2 above 0.98 in absolute terms, and the coal fraction at t, t-1 and t-2 all correlate above 0.96. Putting them in together makes the coefficients unstable, so the specifications were kept parsimonious: one fuel mix proxy, one structural control, plus the policy dummies. Models were then selected on out of sample RMSE and MAE with residual diagnostics, not on raw correlations.
| Model | Transformation | AIC | BIC | Remark |
|---|---|---|---|---|
| ARIMA(0,1,0) with drift | None | 2593.4 | 2597.6 | Random walk with drift, best on the raw series |
| ARIMA(0,2,1) | Box-Cox, λ = 0.353 | -0.63 | 3.53 | Selected. Better fit after variance stabilisation |
The ARIMAX model tracks the emissions trend on a held out test set with a mean absolute percentage error of about 4.4%. RMSE 2.37 billion and MAE 1.92 billion in CO2 units.
GDP per capita has a strong positive effect on emissions across every specification. Over this period, growth has not been decoupled from emissions.
The human rights index is negatively associated with emissions. Stronger institutions line up with lower emissions, although the design supports association rather than cause.
The honest reading is narrower than the headline: growth still drives emissions and governance tracks against them, so the lever is not the fuel mix alone.
Coursework in quantitative analysis, presented with four classmates. The analysis was mine: stationarity testing, model selection, the ARIMAX specifications and the residual diagnostics, in R with forecast, tseries, lmtest and car. Original dataset HRIDATA-2.sav.
Next: HR Operations Dashboard