I build machine learning systems people actually use.

The school planned rosters and boat allocation by feel, and 37.7% of private sails ended in a last-minute cancellation. They needed to see ten days ahead.

Four models on 37,000+ historical events collected by a scraper I keep running around the clock: weather-aware LightGBM demand forecasting, a calibrated cancellation-risk classifier, deterministic A* safety routing over a graph of the marina waters, and a LambdaRank recommender with 103 features across 18,626 ranking groups that sorts 1,400+ members into 7 archetypes. It all lands on one bilingual Streamlit dashboard, with around 400 automated tests behind it.

The demand model reached MAE 4.56, a 49% improvement over the seasonal baseline, and won all 7 rolling-origin validation windows.

Israelis kept asking one small question with no data behind it: if I lie down for forty minutes, what are the odds a rocket alert interrupts me?

A Poisson risk engine over live alert data from Israel's Home Front Command, blended with trend, recency decay, and time-of-day bias. The interesting bit tells Iranian and Lebanese attacks apart with zero manual labeling: one alert category, the early warning, always precedes Iranian ballistic missiles and is physically impossible for rockets fired from Lebanon, so it gives away the source on its own. Next.js and TypeScript with 78 tests, self-hosted on a server in Tel Aviv because the government alert API only answers requests from inside Israel.

In production, covering 1,450+ cities in Hebrew and English, and it has processed 6,500+ alerts across 996 affected cities since launch. Rewriting the core lookup from a nested scan to a single indexed pass took it from O(n²) to O(n), which keeps the page interactive during an alert storm. The whole front end ships in 137 KB.

Most flights leave on time. A model that always predicts on time scores high and helps nobody. The useful skill is flagging the late ones without crying wolf.

Temporal features engineered from the schedule data, SMOTE to rebalance the rare delay class, and the trained classifier served as a Dockerized FastAPI endpoint.

82% precision on flagged delays, packaged as a service another system could call rather than a notebook. Supervising real departures taught me which false alarms would annoy the people who'd rely on it.

Which students are heading toward failure, early enough for a university to step in? The raw material: 10.6 million rows of clicks from the university's online course system.

A PySpark pipeline. Aggregating before joining cut the working data 300x before any model saw it. We compared four models on the reduced set.

Gradient Boosted Trees reached 92.9% accuracy. The finding that matters is simpler: assessment participation alone carries 43% of the feature importance, a cheap early-warning signal any university could track.