Separate generation and analysis
A modular generate → analyze → visualize workflow makes the simulated inputs and transformations clear.
Data Analytics / BI · Synthetic data · Case study
From orders
From orders
to operating decisions.
A five-page Streamlit platform over synthetic Sri Lankan food-delivery data, built to examine marketplace health, pricing, restaurants, courier operations and restaurant churn signals.
UEUber Eats Sri Lanka Analytics cover
70,000+Synthetic 2024 orders
200Simulated restaurants
5 pagesInteractive dashboard
From generated data to decisions
Each stage feeds the next, from simulation to dashboard.
1
Generate
Synthetic 2024 marketplace data
2
Store
Orders and partners in SQLite
3
Query
SQL for demand, pricing, operations
4
Model
Random Forest flags churn risk
5
Explain
A five-page Streamlit dashboard
The problem
Marketplace metrics need operational context.
A food-delivery marketplace needs more than a revenue chart. Teams need to see demand by hour, cancellations by cause, zone performance, courier activity and restaurants that may need attention. This portfolio project simulated those operational questions for Sri Lanka using synthetic data rather than real Uber Eats records.
Architecture
Generate, query, model, explain.
The diagram shows how data moves through the project.
FakerSynthetic records
SQLiteFour tables
SQL + PythonQueries and analysis
Random ForestChurn-risk score
StreamlitFive dashboard pages
Python modules generate the synthetic marketplace, run ten business SQL queries and perform deeper analysis. Plotly presents the results in Streamlit.
Key decisions
Why I built it this way.
Design by business question
Five pages group marketplace, pricing, restaurants, driver operations and churn rather than mixing every chart on one screen.
Use a risk score for outreach
A Random Forest model scores restaurant partners so a partner team can prioritize investigation, not treat a score as confirmed future churn.
Results
Findings in the simulated marketplace.
Scope and measurement matter as much as the headline numbers.
~82% conversion
Platform-wide conversion in the synthetic order data; the remaining fallout is the gap worth investigating.
~45% of demand
Orders between 12–14h and 19–21h account for roughly 45% of simulated daily orders.
~15% high risk
The churn model marks about 15% of simulated restaurant partners as high-risk signals.
Measurement: These are descriptive results from generated 2024 marketplace data: 70,000+ orders, 200 restaurants and 300 couriers. The churn target is derived from low volume and low conversion; held-out predictive accuracy is not reported here. The elasticity coefficient is approximately −0.5 in this simulated analysis.
What broke
What I learned while building it.
Synthetic data limits
Because the marketplace is simulated, pricing and churn findings demonstrate the method and the modelling approach, not real Sri Lankan market behaviour.
Model target design
The churn label is based on low order volume and conversion, so the risk score is a demonstration of prioritization rather than externally validated churn prediction.
Synthetic operating context
Data generation enables reproducible scenarios, but the resulting relationships may not represent a live marketplace.
Limitations
What this project does not claim.
All marketplace records are synthetic. The model
target is a constructed churn label, and no held-out model accuracy or causal pricing estimate is claimed. Findings are
portfolio demonstrations and should not be described as real Uber
Eats Sri Lanka performance.
Explore the work
Code, report and next steps.
Explore the five-page dashboard or inspect the data generator, SQL queries, analysis modules and Streamlit application in the repository.