A production-level warehouse analytics system that identifies operational bottlenecks, predicts order delays with machine learning, and generates quantified business recommendations — built to demonstrate the full data engineering and analytics stack required for manufacturing operations roles.
This project simulates 18 months of warehouse operations data (5,000 orders) and delivers:
- SQL-driven KPI engine — 12 analysis functions querying a structured SQLite database
- Bottleneck detection — composite scoring across zones, shifts, congestion, and equipment
- ML delay prediction — RandomForest classifier (ROC-AUC: 0.876, Accuracy: 79.6%)
- Automated recommendations — 7 data-driven business insights with quantified impact
- 4-page Streamlit dashboard — interactive, production-quality visualisations
| Finding | Metric |
|---|---|
| Zone B (bottleneck) vs warehouse average | +9pp delay rate, only zone exceeding SLA (time ratio 1.02) |
| High vs Low congestion | 1.44× pick time, 1.37× delay probability |
| Junior (1–3 yrs) vs Senior pickers (7–10 yrs) | +20pp delay rate, 58% slower per item |
| Night vs Morning shift | +8pp delay rate with similar avg experience |
| Urgent order SLA compliance | 64.8% breach rate (28-min target) |
| Equipment downtime impact | 1.32× packing time, affects 19% of orders |
Four interactive pages built with Streamlit + Plotly:
| Page | Contents |
|---|---|
| Operations Overview | KPI cards, order volume by shift, delay rate by zone, monthly trend, priority SLA |
| Bottleneck Analysis | Composite bottleneck scores, zone × shift heatmap, congestion impact, experience curve |
| Delay Prediction | Live ML inference form, risk gauge, feature importance, confusion matrix |
| Recommendations | 7 severity-filtered business insights with findings, actions, and quantified metrics |
Model: RandomForest Classifier (200 trees, balanced class weights)
Design principle: trained exclusively on pre-order features — no picking or packing times — to ensure the model is genuinely predictive, not retrospective.
| Metric | Score |
|---|---|
| Accuracy | 79.6% |
| Precision | 71.4% |
| Recall | 80.5% |
| F1 | 75.7% |
| ROC-AUC | 0.876 |
Top predictive features: workload index, items count, order priority, SLA target time, picker experience.
warehouse-operations-analytics/
│
├── src/
│ ├── generate_data.py # Synthetic dataset generation (5,000 orders, 18 months)
│ ├── preprocess.py # Cleaning, feature engineering, SQLite ingestion
│ ├── analysis.py # 12 SQL-driven KPI and bottleneck functions
│ ├── ml_model.py # RandomForest training, evaluation, persistence
│ └── recommendations.py # Automated insight generation from live data
│
├── dashboard/
│ ├── app.py # Home page
│ ├── utils.py # Shared loaders, CSS, colour palette
│ └── pages/
│ ├── 1_Operations_Overview.py
│ ├── 2_Bottleneck_Analysis.py
│ ├── 3_Delay_Prediction.py
│ └── 4_Recommendations.py
│
├── data/
│ ├── warehouse_orders.csv # Raw generated dataset
│ ├── warehouse_orders_clean.csv # Enriched dataset (26 features)
│ └── warehouse.db # SQLite database (orders + ml_features tables)
│
└── models/
└── model_metadata.json # Accuracy, confusion matrix, feature importance
| Layer | Tools |
|---|---|
| Data generation & processing | Python, NumPy, Pandas |
| Database | SQLite (via sqlite3 + pandas.read_sql) |
| Machine learning | Scikit-learn (RandomForest, train/test split, metrics) |
| Visualisation | Plotly Express, Plotly Graph Objects |
| Dashboard | Streamlit (multi-page, @st.cache_data, @st.cache_resource) |
| Model persistence | Joblib |
Prerequisites: Python 3.12+
# Clone and enter the project
git clone https://github.com/SepehrKalantariSol/warehouse-operations-analytics.git
cd warehouse-operations-analytics
# Create virtual environment
python3 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt
# Run the full pipeline
python src/generate_data.py # 1. Generate dataset
python src/preprocess.py # 2. Clean + load into SQLite
python src/ml_model.py # 3. Train and save the ML model
# Launch the dashboard
streamlit run dashboard/app.pyOpen http://localhost:8501 in your browser.
| Column | Type | Description |
|---|---|---|
order_id |
TEXT | Unique order identifier |
order_date |
DATETIME | Order timestamp (Jan 2024 – Jun 2025) |
shift |
TEXT | Morning / Evening / Night |
zone |
TEXT | Warehouse zone (A–E) |
order_priority |
TEXT | Low / Medium / High / Urgent |
items_count |
INT | Number of items in the order |
picker_experience |
INT | Picker's years of experience (1–10) |
distance_travelled |
FLOAT | Estimated pick path distance (metres) |
congestion_level |
TEXT | Low / Medium / High |
equipment_available |
BOOL | Whether handling equipment was available |
picking_time |
FLOAT | Time to pick all items (minutes) |
packing_time |
FLOAT | Time to pack the order (minutes) |
total_time |
FLOAT | picking + packing time |
target_time |
INT | SLA target in minutes (by priority) |
delayed |
INT | Target variable — 1 if total_time > target_time |
This project is modelled on the analytics requirements of production steering and logistics optimisation roles in manufacturing environments (e.g., BMW Group, automotive Tier 1 suppliers). The analysis framework mirrors real operational challenges:
- Zone-level performance monitoring against SLA targets
- Shift and staffing impact on throughput
- Congestion and equipment utilisation as controllable delay drivers
- Experience-based workforce planning
MIT