HetangPatel

Three layers, one question. A Difference-in-Differences model with team fixed effects and HC3 robust errors isolates the regulation's true impact from pre-existing trends — the DiD coefficient lands at +0.068 (p = 0.050), a real but modest convergence, exactly consistent with Red Bull re-establishing dominance by 2023.

On top of that sits a walk-forward Ridge forecaster that learns regulation-response patterns as each post-2022 season closes, then projects the 2026 reset. The whole thing is wired through a production cloud stack — raw parquet into Snowflake, dbt staging-to-mart transformations, and a three-view Power BI dashboard connected live to the warehouse.

The F1 framing is the hook. The method — separating a true policy effect from noise, then tracking a rolling forecast against live actuals — is the same one used for pricing shocks, market entries, and demand planning.

Halifax Transit's GTFS feed went into a normalized PostgreSQL + PostGIS schema with GIST spatial indexing, automatic geometry triggers, and a validation layer that caught and filtered broken foreign-key references before they corrupted any join.

From there, SQL analytical views, CTEs and window functions surface the network's truth: 41% of stops are served by a single route, 973 stops sit isolated beyond 500 metres of a neighbour, and the 5 PM peak pushes 33,635 departures. A stop-connectivity query that took four seconds was driven under one with the right B-tree and spatial indexes.

It ships as an interactive Streamlit dashboard with Folium maps over all 2,380 stops — built so an operations stakeholder can see coverage inequity, not just admire a chart.

An XGBoost lap-time predictor (MAE 2.6s, trained on 86,000+ laps across three seasons) drives a modular race-simulation engine that runs a full field over an entire Grand Prix. Feed it Strategy A versus Strategy B and it returns predicted finishing positions, times, and the winner — side by side.

The engineering discipline is in what was left out: I stripped outcome-leaking features like average speed and throttle, restricting the model to signals genuinely available before each lap. XGBoost beat an LSTM by more than 2x on this tabular data — the right tool over the fashionable one.

Validated against the 2024 Abu Dhabi Grand Prix: the model called Norris's conservative one-stop over a two-stop alternative, which is exactly how the real race played out. Deployed as a Flask API on Railway behind Docker, React front end on Vercel, CI/CD through GitHub Actions.

As Development Director I led a 10-person cross-functional team rebuilding a faculty client's legacy desktop genomics tool as a modern web application. I translated dense bioinformatics requirements into a clean architecture: a React + Firebase front end talking to an R Plumber API that exposes native DESeq2 analysis without rewriting the science in another language.

The proof of concept the client cared about most — one complete analysis method, executed end to end with verified-correct results — we delivered. A POST hits the API, temporary count and metadata matrices are reconstructed in isolated workspaces, DESeq2 runs, and the resulting volcano plot, MA plot and heatmap come back as rendered figures.

My job was as much communication as code: weekly client engagement, eliciting analytical requirements, and steering a team across the full skill range. Frontend reached ~80% completion and every milestone landed on schedule — the client confirmed the project viable for continued development.