Dell Technologies - Data Scientist Case Study

Samridh Srivastava | March 2026

---

Overview

This repository contains my solution to the Dell Global Operations manufacturing analytics case study. The objective is to predict kit end date and kit cycle time (the interval in minutes between kit assembly start and completion) for manufacturing orders, evaluated primarily on MAE, with a secondary lens on on-time delivery classification.

The work is structured as a four-iteration, cumulative Jupyter notebook that progressively builds from data cleaning through to business-ready recommendations.

---

Repository Structure

├── SamridhSrivastava_DS_Case_Notebook_2026.ipynb   # Main analysis notebook
├── SamridhSrivastava_DS_Case_Slides_2026.pdf      # Executive summary slides pdf
└── README.md

Note: Raw data files (train.parquet, calendar.parquet, capacity.parquet, inference.parquet) are not included in this repository per submission guidelines.

---

Notebook Structure

Iteration	Focus	Key Output
1. Data Cleaning & Baseline	Negative durations, sentinel dates, outlier removal, Linear Regression baseline	Clean dataset, baseline MAE
2. Calendar Feature Engineering	Fiscal week, EOQ flag, weekend indicator, day-of-week	Best holdout MAE (~0.88 days); LR + Calendar outperforms tree models on holiday-dense test window
3. Capacity Feature Engineering	OT utilisation, site congestion (weighted by remaining expected_kit_minutes), lag/rolling features	Time-series CV with reduced variance vs. Iteration 2
4. Insights & Classification	SHAP attribution, OT/congestion what-if simulations, on-time classification (Logistic Regression + Random Forest), operational recommendations	ROC/PR curves, 5 strategic recommendations

---

Key Modelling Decisions

• Time-based validation throughout: TimeSeriesSplit and a chronological holdout set — no random splits that would leak future information.
• Calendar features over model complexity: Fiscal calendar flags (is_eoq, fiscal_qtr_weeknum, is_weekend) outperformed tree-based ensembles on the holiday-dense test window, reinforcing that domain context can matter more than algorithmic sophistication.
• EOQ defined by Dell's fiscal quarters, not calendar quarters.
• Site congestion weighted by remaining expected_kit_minutes rather than raw concurrent order count.
• Causal caveats on simulations: What-if OT analyses are framed as model-based simulations, not causal claims.

---

---

---

Performance Summary

Best Test MAE: 0.70 days (XGBoost + Calendar + Capacity Features)

The model achieved a 24% improvement over the baseline (0.92 days → 0.70 days) through progressive feature engineering. Calendar features alone contributed a 18% gain (0.92 → 0.75 days), while capacity features added the final 7% (0.75 → 0.70 days). Critically, hyperparameter tuning via TimeSeriesSplit grid search produced <0.005 day improvement, confirming that feature engineering, not model complexity, drove the gains.

Progression across iterations:

• Baseline (Order + Expected Minutes): 0.92 days (22 hrs)
• • Calendar Features: 0.75 days (18 hrs)
• • Capacity Features: 0.70 days (17 hrs)

The model is stress-tested on the Dec 2025 – Jan 2026 holdout period (EOQ + holidays), making the 0.70-day MAE especially robust for operational deployment.

---

Strategic Recommendations (Summary)

1. Weekend staffing adjustments based on observed cycle time differentials
2. Congestion-based scheduling to smooth order intake during high-utilisation periods
3. OT utilisation as a leading indicator for proactive intervention
4. pre-EOQ preparedness planning aligned to Dell's fiscal calendar
5. Shipping promise calibration informed by predicted cycle time distributions

---

Tools & Libraries

Python · pandas · scikit-learn · XGBoost · SHAP · matplotlib · seaborn