Sreekaran Reddy Ramasahayam
Professional Summary
Data Scientist with an M.S. in Computer Science and experience applying predictive models to pricing and scholarship decisions. Uses Python and SQL for feature engineering, baseline comparisons, and model evaluation, with project work in causal inference and time-series forecasting. Deployed a bidding-support tool that reduced analysis time by 50%.
Professional Experience
Data Scientist
- Engineered features from construction costs, labor, location, and project duration for bid-pricing models using 20+ predictors.
- Benchmarked Ridge and Lasso against linear regression to evaluate held-out bid-pricing performance.
- Deployed pricing models in an internal bidding-support tool, reducing bid analysis time by 50%.
- Standardized 50K+ construction records using SQL and Pandas, improving reporting reliability by 30%.
- Presented Tableau reporting to the CEO and client board bi-weekly, supporting an operational efficiency gain of 15%.
Data Science Intern
- Transformed graduated-student histories into per-grade snapshots to align model inputs with partially observed academic records.
- Selected XGBoost over mean and running-average baselines under cohort-grouped cross-validation, achieving 0.8 MAE on a 10-point scale.
- Fit quantile regression models to estimate lower and upper bounds for nominal 95% prediction intervals.
- Defined scholarship tiers with borderline-case review, reducing the required reviewers from four to one.
- Analyzed departmental expenses to identify cost-saving opportunities of 10–15%.
Information Services Lab Assistant
- Guided exploratory analysis, statistical validation, and assumption checks for 20+ student research projects.
- Resolved lab hardware, login, and workstation issues alongside Python and SQL support, closing 75+ tickets per semester.
Projects
FanHouse: Membership Impact Study (Simulated Data)
- Matched 12,153 of 40,000 synthetic customers using propensity scores, reducing standardized mean difference from 0.396 to 0.001.
- Estimated purchase-value impact using matched difference-in-differences, finding a statistically significant lift of $0.14/customer/day.
- Found no statistically significant lift in net customer payments after discounts and rewards, distinguishing purchase value from cash paid.
Loan Delinquency Forecasting
- Compared ARIMA, XGBoost, and LightGBM under a common walk-forward backtest across five FRED loan-delinquency series.
- Corrected target-construction leakage and added a regression test to GitHub Actions to prevent recurrence.
- Evaluated forecast-error differences with Diebold-Mariano tests, finding results favoring ARIMA in 27 of 40 comparisons at the chosen significance threshold.
- Analyzed tree-model predictions with SHAP, attributing 47–59% of feature importance to the most recent lag.
MedPredict-X: Readmission Risk Modeling with RAG
- Benchmarked five model types on 100K+ patient encounters, selecting XGBoost with an ROC-AUC of 0.93.
- Designed prototype outreach tiers using predicted risk: doctor follow-up at p ≥ 0.6, nurse calls at 0.3 ≤ p < 0.6, and automated messages.
- Built a FAISS and Qwen-32B RAG layer to generate contextual explanations alongside predicted risk tiers.