Approach

Pre-COVID was a six-year flat plateau at ~15,800 K-5 students. COVID broke it to ~12,400. With only two post-break data points, a Prophet-style ML answer pretends to know more than the data can say.

• →Stitched 12 years of NYC DOE and NYSED enrollment data and fit a piecewise linear model with a known structural break at 2020. Bounded the future with three analyst scenarios (decline, stabilization, recovery) instead of a single point forecast.
• →Back-tested when NYSED published the 2022–25 actuals. The Base scenario landed at 8.71% MAPE. Prophet on the same data landed at 19.20%. The analyst-bounded approach beat the ML default by more than 2x.
• →Tested the popular "wealthy catchments left first" story and refused to ship it. Pearson r = +0.14 between catchment income and enrollment decline. Lowest-income schools in Chinatown and the Lower East Side declined more than the high-income downtown catchments; the resilient outlier is Roosevelt Island (+13%), suggesting geographic friction of exit, not income, is what matters.
• →Shipped a four-page Power BI dashboard: system-wide fan chart, scenario/model operational matrix, single-school drill-through, and a backtest validation page.

Approach

Event detection in cities usually leans on calendars and permits. That leaves out the events nobody publishes a permit for and biases the data toward the venues that are easy to list. I wanted to know how much of the signal lives in the ridership itself.

• →Built day-of-week and season-aware ridership baselines per station-hour, so a sold-out Sunday MSG game gets scored against the right Sundays, not against a generic weekday median. Without that, the model flags every Friday night.
• →Quantified each event with five fingerprint dimensions (peak intensity, lead time, lag time, decay half-life, and pre/post asymmetry) instead of just an excess-ridership number. The shape is what tells you Knicks from Rangers from a U2 show.
• →Clustered the fingerprints and reported the result honestly: the venue dominates the signature, not the sport. Knicks, Rangers, and MSG concerts land in the same cluster. Yankees day games and night games sit ~3x apart in intensity. Parades stay asymmetric: people arrive early and exit fast.
• →Validated against a held-out 2024 calendar of 513 events: 96.5% recall, without ever giving the model the calendar.

Approach

Fare debates usually live in spreadsheets and op-eds. Riders can’t see how a policy change would actually affect their own trip.

• →Built the analysis in pandas and NumPy off the open MTA OD pairings, fare structures, and station metadata. The whole thing is reproducible, so the model can be re-run when the data changes.
• →Uncovered roughly $913M in MTA revenue gap between current flat-fare receipts and a distance-based structure, then proposed a pricing optimization that recovers the gap without raising consumer fares.
• →Reported the loss as well as the gain. About 15% of riders in the model pay less, and the rest pay more. Hiding that would have been a worse choice than reporting a worse number.
• →Framed the page around a single comparison, your trip today versus the proposed model, so the reader is the one driving the page and not the agency.

Approach

A new CPG brand needs a site that reads as confident and shoppable before there’s actually anything to ship.

• →Built a small token-driven system first, so the flavor pages, the locator, and the marketing pages could share the same rhythm instead of reinventing components for each.
• →Split the flavor pages out of the catalog. Each SKU has its own URL, image set, and ingredient story, instead of a single overloaded grid.
• →Used motion to reinforce hierarchy on entry, then quieted it on scroll so the secondary content reads without anything competing for attention.

Approach

Long-form reading on the web is hostile to focus. Most "speed reader" UIs end up adding controls until the controls become the distraction.

• →Stripped the interface down to one fixation point and the controls you need to recover from a misread. Everything else is keyboard.
• →Tuned the ORP highlight contrast so the focal letter reads as the anchor and doesn’t flicker between words.
• →Sized the reading column for what’s actually readable at high WPM, not for what fills the screen.