How I Replicated Uber's Core Marketplace in Python: A Technical Deep Dive

Bikash Kh — Fri, 19 Jun 2026 23:38:35 +0000

I Built UberSim v2.0: A Production-Grade Urban Mobility Intelligence Platform 🚗🧠

Every time you open Uber and see a 2.1× surge multiplier, a complex system has already predicted demand, optimized prices, matched drivers, and logged events for future learning — all within milliseconds.

I wanted to understand how those systems work.

So I built UberSim v2.0.

A Python-based urban mobility intelligence platform that simulates the core engineering challenges behind modern ride-sharing marketplaces.

Instead of building another dashboard project, I wanted to recreate the intelligence layer behind a ride-sharing platform from scratch.

🚀 What's Inside?

🧠 Demand Forecasting

Spatio-temporal demand prediction (R² = 0.89)
Weather effects, seasonality, lag features, and neighboring zone influence
Predicts ride demand across multiple city zones

🕸️ Graph Neural Networks

Models the city as a graph
Nodes = city zones
Edges = historical trip flows
Captures spatial mobility patterns that traditional models miss

🤖 Reinforcement Learning Pricing

Built a PPO-based surge pricing engine that learns pricing policies instead of relying on hand-crafted rules.

Optimizes multiple objectives simultaneously:

📈 Platform revenue
🚕 Driver earnings
😊 Rider welfare
⏱️ Wait times
⚖️ Fairness constraints

One interesting finding:

The RL agent learned to gradually increase surge prices instead of aggressively reacting to demand spikes. This behavior wasn't explicitly programmed.

⚡ Kafka-Style Real-Time Streaming

Implemented an event-driven architecture with:

Ride request streams
Driver status updates
Pricing events
Match results

Supports historical replay and live marketplace metrics.

🧠 Driver State LSTM

Predicts four operational driver states:

online_idle
online_busy
relocating
offline

Built entirely in NumPy with Backpropagation Through Time and Adam optimization.

🧪 Counterfactual A/B Testing

Implemented production-style experimentation techniques:

IPS (Inverse Propensity Scoring)
Doubly Robust Estimation
CUPED variance reduction
Bootstrap confidence intervals

This allows evaluating policies without deploying every experiment in production.

🗺️ Multi-Modal Transit Planning

Journey planning across six transportation modes:

🚗 Rideshare
🚌 Bus
🚇 Subway
🚲 Bike
🛴 Scooter
🚶 Walking

Uses A*/Dijkstra optimization to balance:

Travel time
Cost
CO₂ emissions
Number of transfers

💡 What I Learned

The hardest problem isn't maximizing revenue.

It's maximizing revenue while remaining fair.

Without constraints, optimization naturally prioritizes high-demand areas and disadvantages low-supply neighborhoods.

Adding fairness fundamentally changes the optimization landscape.

Some other takeaways:

RL discovers strategies humans don't explicitly program.
GNNs capture spatial relationships that tabular models miss.
Causal inference is essential for policy evaluation.
Pure NumPy is more powerful than people think.

🛠️ Tech Stack

Python · Streamlit · Plotly · Stable-Baselines3 · NetworkX · NumPy · Scikit-Learn · Gymnasium

🔮 What's Next?

[ ] Graph Attention Networks (GAT)
[ ] Multi-Agent Reinforcement Learning
[ ] Real Kafka Broker Integration
[ ] WebGL City Visualization
[ ] Real-World Dataset Integration (NYC TLC, Chicago Divvy)

🔗 GitHub

https://github.com/kh-bikash/ubersim

Feedback, ideas, and contributions are welcome 🚀

DEV Community: Bikash Kh