Justin Babu

Assistant Manager

Goods And Services Tax Network • Temps plein

Jun 2025 - Present • 1 yr 4 mos

•Built and benchmarked three Text-to-SQL architectures (LLM baseline, domain-specific RAG, and multi-agent framework) with query planning and schema grounding over a GST analytics database — RAG improved lenient accuracy from 56% to 64% and reduced the silent-wrong rate from 42% to 32%; evaluated across 39 gold queries on execution rate and result correctness. •Measured the incremental revenue effect of national policy changes using a Difference-in-Differences (DID) framework — benchmarking firm-level fixed-effects regression against an XGBoost counterfactual trained on pre-intervention data across 20,000+ accounts, isolating a statistically significant 10.6% uplift (p = 0.05) in one category against a null result in another. •Modelled the link between tax collections and macroeconomic indicators (GDP and GVA) using SARIMAX, with two-quarter-ahead GDP forecasts landing within 2.96% of published actuals. •Built scenario and sensitivity models tracing how rate changes flow through an 11,283 × 11,283 Input-Output matrix, quantifying revenue impact across 1,399 product-rate combinations. •Built an end-to-end semantic search pipeline (Product Search Engine) using sentence transformers (multilingual-e5-small) mapping free-text product descriptions to tax codes across 150 million records — achieving 72% top-5 accuracy against a 700-query benchmark, outperforming the existing keyword-search tool.

Analytics Intern

KPMG • Temps partiel

Jun 2025 - Aug 2025 • 2 mos

Built six-week-ahead electricity demand forecasts for a Central Electricity Authority engagement, comparing SARIMA, ARFIMA, Holt-Winters and Ridge/Lasso specifications on forecast accuracy. •Benchmarked machine learning alternatives: gradient boosting, random forest, SVR, GAMs and neural networks, against the econometric baseline over the same horizon rather than committing to a single method. •Reviewed published forecasting research across short, medium and long horizons to establish which approach suited each, informing model selection for the engagement