// Data · Machine Learning · Analytics
Interactive, real-world demos of ML, Data Analytics & Science — built for exploration. No code, no setup. Just curiosity.
// Data Analytics & Science
Real-world analytics and applied intelligence tools spanning pharma regulatory affairs, clinical trials, pharmacovigilance, public policy, and quantitative finance. Built on live external data sources — CMS, FDA, EMA, CT.gov, Eurostat, UK Police — with AI-assisted analysis layered in where it genuinely adds value, not for its own sake. No code, no setup — just input and explore.
Demand forecasting and spend-variance analytics on real CMS Medicare Part D drug spending — two forecast models, validated MAPE, budget tracking across 20 drugs.
FDA label intelligence & EMA regulatory analytics — version diff, NLP change classification, and AI Q&A with hybrid retrieval + reranking, measured on a public evaluation page.
Ask questions of 13 public trial protocols or your own PDF — layout-aware parsing, PII redaction, injection-tested, with a public evaluation page. Plus trial analysis, site maps & protocol drafting.
Clinical trial data analysis using ADaM built on CDISCpilot01 — Xanomeline TTS in Alzheimer's Disease.
Patient cohort analysis on SyntheticMass-Synthea synthetic healthcare data.
FDA FAERS PV Signal Explorer, post-market drug safety surveillance analyses.
Clinical trial management platform targeting small biotech sponsors, academic research.
Independent geospatial and descriptive analysis of street crime and stop & search data across five English police forces — 2023–2025.
Official Eurostat data on asylum flows, decisions, returns, and crime by citizenship across EU-27 — 2013 to 2025.
Vectorized backtesting engine for quant trading strategies — SMA crossover,RSI mean reversion, and momentum breakout, with equity curves and performance metrics.
A new demo is in development. Stay tuned.
// Machine Learning
A full spectrum of supervised ML — imbalanced, multiclass, ordinal & multilabel classification; regularized, quantile, count & survival regression; transfer-learning computer vision and transformer-based NLP. Every app is built end-to-end in Python — Scikit-Learn, XGBoost, LightGBM, statsmodels, lifelines, TensorFlow — on real public datasets, deployed via FastAPI + React with an honest diagnostics panel showing exactly how each model was validated, not just its headline number. No code, no setup — just input and explore.
// Classification
Predicts Class I (most severe) drug recalls from openFDA enforcement data. XGBoost vs. LightGBM, SMOTE vs. class-weighting compared head-to-head, PR-AUC as the honest metric under severe imbalance.
Multiclass hazard-type classification on CPSC consumer product recalls — 17 hazard categories, SHAP explanations, and a documented multi-hazard ambiguity failure case.
Ordinal prediction of UK Food Hygiene Rating (0–5) from FSA inspection data — QWK-optimized thresholds, with an honest look at where the model genuinely can't separate Poor from Fair.
Multilabel prediction of UK Companies House SIC industry codes from structured company profile data — binary-relevance modeling across hundreds of possible codes.
// Regression
Ridge vs. Lasso vs. ElasticNet on 4.2M UK Land Registry sales — caught a mis-specified alpha grid before reporting, per-property-type outlier winsorization, and a real accuracy gain from finer-grained geography.
Quantile regression on 25 years of UK electricity demand — caught a silent date-parsing bug corrupting every 2025 record, then tuned regularization against genuine held-out coverage, not in-sample fit.
Poisson vs. Negative Binomial regression on 5 years of FDA FAERS reports — a 25GB ETL crash fixed, a join-key bug that was silently dropping amended cases, and a numerical blowup caught before it shipped.
Cox Proportional Hazards survival analysis on a breast cancer clinical trial dataset — full diagnostic arc from raw fit through stratification, Kaplan-Meier curves, and a Random Survival Forest benchmark.
Parametric survival extrapolation into a QALY cost-effectiveness model on real breast cancer trial arms — AIC-ranked curve fits, probabilistic sensitivity analysis with a CEAC, and a closed-form threshold-price explorer.
// Computer Vision & NLP
Transfer-learning image classifier (ResNet50) for plant disease detection — frozen vs. partial vs. full fine-tuning compared head-to-head, with Grad-CAM showing exactly what the model is looking at.
Sentiment classification on IMDB reviews — TF-IDF baseline vs. frozen vs. fine-tuned DistilBERT, with a genuinely surprising finding: fine-tuning, not the transformer architecture alone, is what wins.
A new demo is in development. Stay tuned.
// Power BI Analytics
Interactive Power BI reports built on real public datasets — ClinicalTrials.gov, NHS prescribing data, and the Bank of England's own time-series database — each built around a distinct advanced technique: row-level security modeling, AI-powered key influencers and decomposition analysis, and DAX time intelligence with What-If forecasting. No login required — click through and explore.
Trial status, sponsor mix, and enrollment geography across 868 ClinicalTrials.gov studies, with a Gradient Boosting trial-success model and a full architecture & methodology page.
£11bn+ in NHS prescribing spend across 48 ICBs and 5 BNF categories, built to demonstrate dynamic Row-Level Security by role. Shown here as an unrestricted public demo — the RLS mechanism is documented in-report.
Bank Rate and GBP/EUR, GBP/USD trends sourced directly from the Bank of England's own database, with DAX time-intelligence measures and an interactive What-If tool for exploring hypothetical rate paths.
A new demo is in development. Stay tuned.