Creating Transformative Impact with AI in Food Delivery 🤖

Introduction

• Speaker's Greeting: Encourages questions and interaction.
• Topic: Transformative impact of AI by sticking to fundamentals.

Delivery Business Model

• Value Proposition: Fast delivery from favorite restaurants to your door.
• Types of Users:
  • Customers (Eaters): Want fast food delivery.
  • Riders (Couriers): Aim to maximize earnings per time.
  • Restaurants: Seek reliable and prompt delivery partners.

The Optimization Challenge

• Complex Problem: Requires optimizing delivery in real-time with many variables.
• Uncertainty: Orders and rider availability are unpredictable.
• Approximation: The key is finding better approximations continuously.

Initial Approach

• Manual System: Initially used McKinsey analysts with Google Sheets (Do things that don’t scale - Y Combinator advice).
• Growth: Operated manually through four rounds of funding, multiple country launches, and significant growth.

Transition to AI

• AI Definition: Intelligent Automation - algorithms to automate human decisions.
• Initial Manual System:
  • Restaurant and delivery workers made key logistical decisions.
  • Algorithms now automate all time estimates and critical operations.

Impact of Automation

• Business Growth: 150% growth, elimination of tedious tasks.
• Improvement Metrics:
  • 40% reduction in late orders.
  • 12% reduction in order duration.
  • 13% reduction in rider-to-restaurant time.
  • 48% reduction in late pickups.
• Happier Users: All three user types show increased satisfaction.

Fundamentals for Success

• Incrementalism: Break down the automation journey into manageable themes.
  • Travel-time model, rider delay model, food prep model, objective function, and solver.
• Bias to Simplicity: Prefer simple methods that deliver value quickly.
  • Use generalized linear models with careful sampling and feature engineering.
  • Avoid over-complication; simplicity in approach to validation and iteration.
• Rigorous Experimentation: Measure live metrics and network effects.
  • Use randomized block design for experiments to maintain statistical power and causality.

Incrementalism in Practice

• Sequencing: Build and deploy models progressively, ensuring each step is a live experiment.
• Live Experimentation: Ensures real-time feedback and quick adjustments.

Bias to Simplicity in Practice

• Machine Learning Models: Prefer classic GLMs over complex methods due to ease of implementation and good performance.
• Feature Engineering: Focus on detailed domain knowledge and regularization.
• Iterative Testing: Validate with empirical data rather than theoretical principles.

Optimization Approach

• Solver Simplicity: Use basic deterministic hill climbing, intense domain knowledge.
• Objective Function Approximation: Employ mathematical expectations for non-linear combinations.
• Neighborhood Operators: Analogous to feature engineering in ML; essential for success.

Conclusion

• Key Takeaways: Combination of incrementalism, simplicity, and rigorous empirical validation leads to substantial impact.
• Open for Questions: Encourages further discussion and answers specific queries about the methodologies.

Final Note

• Experimentation and empirical validation are crucial in fast-paced, uncertain business environments.
• The approach blends fundamental scientific principles with practical engineering challenges.