Thomas Step

← Blog

If you still need help building out AWS infrastructure, send me an email and I'll see if I can help out. thomas@thomasstep.com

re:Invent: How Amazon.com Transforms Customer Experiences Through AI/ML

This is an overview of a session that I went to during re:Invent 2021. I start by providing the notes I took during the session, and then I will give my take and comments if I have any at the end.

Wednesday 16:00

INO202

Choong Lee and Laura Squier

How Amazon approaches the idea of applying AI/ML to everything

Agenda:

• Key use cases
• Lessons learned

AI/ML can be used now because we have enough data volume and velocity

• What powers Prime one-day shipping
  • 400 million products are forecasted everyday down to the zip code level
  • Fulfillment centers, robots, trailers, and delivery vans all come into play
  • Well-placed fulfillment centers
  • Know how many trailers and delivery vans are needed to get products out
• Some product have obvious demand
  • Sun lotion
  • Cleaning products
  • Etc.
• Some products are not obvious at all
  • New product or a demand spike
  • Seasonal products
• Goal is to anticipate what products will be needed at what time in what locations
• Forecasting models
  • Traditional approach has input and output with a bunch of if-else conditions
    • Bunch of rules and are manually coded
  • AI/ML models are trained by data points and provide output in probability
• Models used to be decision trees and are now neural networks
  • Nonlinear relationships
  • Automated feature engineering
  • There is a lot of interconnected information that was not used by decision trees that can be with neural networks
• After the forecast comes the fulfillment center
• Items are randomly stored in cubbies which are moved by robots
• Robots/computers know which cubbies have which items
• Robots move to pickers (humans) based on the probability that they will need a certain item they hold
• Robots present the correct cubby to the picker

What other customer experiences are transformed through AI/ML

• Moving from a reactive approach to a proactive approach
• Food allergen or health and safety risks on items
  • They used to gather feedback from customers, identify and tag items with concerns, then suppress items and notify regulatory bodies
  • Had to wait for customer feedback or recalls
  • Goal was to predict health and safety risks
  • Built model around key parameters and assign a probability of concern before anything is shipped
• Shipping packaging size reduction
  • ML model based on business rules and visual inspections
  • That model also knows about “collectibles” and can predict if an item is collectible
  • If collectible, that model knows whether or not the item needs additional protective packaging
• Teams at Amazon use AWS tech to build and develop new models in weeks

What did Amazon learn was most important to get right in the ML journey?

• The hallmark of Amazon is speed and scale
• At that speed and scale, traditional algorithms won’t work
• There are hurdles for businesses that are not already using AI/ML
• Crossing those hurdles
  • ML-first mindset and culture
    • Start with one question: how will you use ML?
  • Enabling teams for the mission
    • Break down the silos between business domain experts and technical experts
    • Evolve the teams as you grow and scale
      • Start small and build within the team before scaling across teams
    • Apply the right tool for the job
      • Simple, off-the-shelf, pretrained AI service; ML platform ready to be trained; or individual algorithms like tensorflow
  • Powerful data platform
    • Data fuels ML-driven innovations
      • Teams closest to the customer will be able to apply ML with the largest impact
      • Amazon uses a big data marketplace for AI/ML teams to have self-service discovery and subscription to data
  • Choosing the right first project
    • Does it solve a real and significant prob for your customer?
    • Where/how does ML unlock new capabilities?
      • If traditional BI tools do a majority of what’s needed, ML might not be a good first project
    • Are there places where we already have a lot of untapped data?
    • Can we achieve success in the first 6-10 months?
      • Get something to prod and prove value to customers
    • Is is important enough to get sustained attention and support?
  • Scaling
    • Remove the blockers to the path to deploying solutions to production

My notes:

I am new to AI/ML. This session made me think that AI/ML was more of a business-related technology than a developer-related one. I am sure that there are applications for both, but I found it strange that they presented much more of the business side of things especially given the venue.

Categories: aws