Data Scientist
MATRIX Analytics Corporation (Domain of Arimax Global Corporation) · Aug 2023 – Present
- Built hybrid SARIMAX, XgBoost and Ridge Regression models for monthly forecasting across 40+ products, reducing MAPE by ~23% versus legacy baselines.
- Engineered preprocessing with MAD-based outlier correction, 6,000+ FRED/Census indicators, lag features, month dummies and feature standardization.
- Improved inventory planning, reducing stockouts by ~18% and over-ordering by ~12%, delivering approximately $0.5M annual cost savings.
- Built an automated validation template across 6 countries and 40+ products, monitoring Actual-vs-Forecast variance, buffer breaches, PSI and historical trends; delivered weekly Monday exception reports for stakeholder action.
- Replicated and audited auto-loan credit risk scorecards, creating 7-module Excel audit packs covering performance, PSI, characteristics and multicollinearity.
- Validated a 24-month PD scorecard for used-auto loans, achieving stronger OOT discrimination with KS 57.3% vs. 47.7% and AUC 84.3% vs. 78.1%.
- Flagged calibration and population drift with ATM of 131.5% vs. 104.3% and PSI of 0.141 vs. 0.054; max correlation of 0.215 showed no multicollinearity concern.
- Built an on-premises financial RAG platform combining fine-tuned FinBERT, hybrid retrieval and a local LLM, converting 5,000+ news articles into source-cited reports and 24-month historical timelines.
- Engineered a 4-agent LangGraph pipeline for retrieval, materiality, verification and reporting, achieving 0.88 RAGAS faithfulness, 0.82 context precision and 93% citation correctness.
- Developed confidence scoring, contradiction and citation checks, flagging 18% low-confidence reports for review and reducing estimated monitoring effort by ~60%.
- Developed Python web scrapers using Selenium, Requests and BeautifulSoup to extract and standardize business data from 5+ public sources (SBA, SEC, GSA, DOL and OpenCorporates), enriching enterprise records for downstream credit-risk scorecard development.
- Engineered 60+ model-ready features from 30+ merchant and external attributes, including growth, ratio, rolling-window, seasonal and risk indicators to support downstream credit-risk scorecard modelling.

