Dual-mode book discovery across a corpus of 10,000+ titles — natural language search via a custom RAG pipeline with semantic embeddings, plus OCR-based cover recognition from photo input.

EDA and hypothesis testing on a 4-table e-commerce dataset (10K customers, 45K transactions, 70K sessions) — 10 statistical tests (t-test, chi-squared, Pearson r), Cohen's d effect sizes, and business insights on discount impact, acquisition channels, and revenue trends.

Three-model NLP comparison on WELFake — 72,134 news articles across 4 benchmark sources. TF-IDF feature extraction with Logistic Regression (96.22%), Random Forest (95.09%), and ANN (96.12%). Live prediction demo flags fabricated content at 99.97% confidence.

Weighted least squares regression on UK Living Costs & Food Survey 2013 (~4,000 households) — models household expenditure against occupational class, tenure type, and household composition using survey weights and HC3 robust standard errors. Validated across 9 robustness appendices.

Time-series analysis of 5 years of daily ADANIPORTS equity data — built ARIMA and LSTM models, compared forecast accuracy (RMSE), and surfaced 3 actionable market trend signals.

End-to-end NLP classification pipeline across 15+ threat categories — cut analyst triage time by 40% (~3h/day per analyst). Deployed as a production REST API with FastAPI.

Interactive policy dashboard across 27 EU member states — visualising €110B+ in annual fossil fuel subsidies with country-level drill-down. Built for policy researchers and journalists.

Custom CNN trained on 54,000 images across 38 classes — achieved 98% test accuracy, then pruned and quantised for mobile deployment at <10MB model size.

Real-time attendance system recognising 50+ registered faces simultaneously from a live camera feed — eliminated 100% of manual roll-call entry for a 200-student cohort.