Work

AI systems don’t fail quietly — they fail in the gaps between prediction and human action. I wanted a practical lab to test “trust + autonomy + clarity” in real workflows, not just benchmark accuracy.

• What I built: A set of small, repeatable experiments (prompt variants + scoring rubrics) that turn messy LLM behaviour into something you can actually compare and improve
• ML / DS angle: Converted qualitative UX/behaviour signals into measurable metrics (e.g., autonomy threat, clarity, reactance risk) so interventions can be evaluated like models
• Pipeline & craft: End‑to‑end run artefacts (generate → score → summarise tables/plots → write‑up) so results are reproducible and stakeholder‑friendly

I learned that “good answers” aren’t enough — the winning systems reduce friction, preserve autonomy, and make the next step obvious.

Forecasts are nice, but retail teams need decisions. I wanted a “decisioning loop” that turns demand modelling into a concrete, constraint-aware price/promo recommendation.

• What I built: An end‑to‑end prototype that extracts the dunnhumby Excel to parquet, builds features (lags, rolling means, price/promo signals, seasonality), then trains a next‑week units model
• ML / DS angle: Simulated candidate actions (price grid × observed promo combinations) and compared policies (naïve revenue baseline vs a penalised, constraint-aware policy)
• Pipeline & craft: Config‑driven scripts + generated summary artefacts so you can rerun scenarios, compare policies, and review outputs without chasing notebooks

I learned that the hard part isn’t modelling — it’s counterfactual thinking and constraints. A “best” action in a sandbox can be a terrible action in a real store.

People don’t “use” AI advice — they negotiate with it. I wanted to model when explainability helps versus when it backfires, especially under disagreement and autonomy threat.

• What I built: A behavioural simulation engine plus a Streamlit dashboard to explore outcomes across reactance and explainability, and to reproduce paper-style figures/summary metrics
• ML / DS angle: Monte‑Carlo simulation with interpretable primitives (trust, signal strength, disagreement, regret, completion) to stress-test “AI utility” under realistic human responses
• Pipeline & craft: Batch runs + run-by-run archives so experiments can be repeated, compared, and extended without manual tweaking

I learned that explainability is not a checkbox — it’s a behavioural intervention. You can increase uptake, but you can also trigger resistance if you ignore autonomy.

EV adoption turns charging into an infrastructure problem. I wanted to prototype a lightweight predictive workflow that connects grid load signals to charging demand.

• What I built: A small Python app and prediction workflow (app.py + prediction.py) with a bundled dataset and deployable artefacts
• ML / DS angle: Packaged the modelling stack as loadable objects (preprocessor.pkl, poly.pkl, dt_model.pkl) so inference stays consistent and reproducible
• Pipeline & craft: Made the “train → save → predict” path explicit with saved artefacts + requirements so the project is runnable without guesswork

I learned that simple, well-packaged ML often beats “fancier” models — shipping a reliable prediction path is the real win.