Zalando successfully applied AI to improve warehouse logistics. They started by developing an algorithm called OCaPi to optimize picker routing and order batching, which increased efficiency by 8.3%. A neural network was then used to estimate route lengths, accurately predicting travel times. This optimized storage location allocation in the warehouse. The presentation emphasized starting small by focusing on clear, measurable goals and using readily available data to optimize existing processes, which builds trust and skills for greater AI applications.

1. AI, erfolgreich angewendet
Lagerlogistik bei Zalando
Calvin Seward • Christian Bracher
26 April 2018
Minds Mastering Machines – Köln

2. OVERVIEW
Introduction to Zalando
Data Driven Organizations
AI Projects: A Fraught Landscape
AI Success Story: Warehouse Logistics at Zalando
Start Small to Win Big
2/47

3. COMPANY INFORMATION (2017)
• Europe’s leading online fashion platform
• Operating in 15 countries
• e4.5 billion net sales
• 15,000 employees from 100+ countries
• 300,000 articles offered at any time
• 23 million active customers
Interested? Visit https://tech.zalando.de and https://jobs.zalando.de
3/47

4. ZALANDO RESEARCH LAB
• Machine learning & artificial intelligence
• Understand fashion & style
• Revolutionize online shopping experience
• Product impact, papers, patents, peer education, prestige
• Autonomous researchers
• Big Data, NVIDIA GPUs, Tensorflow. . .
Web: http://research.zalando.com
Contact: research@zalando.de
4/47

5. WHO WE ARE & WHAT WE (REALLY) DO
Christian: Computer vision for fashion
Fashion DNA: The geometry of style
christian.bracher@zalando.de
Calvin: Generative-adversarial networks
Fashion design by hallucination
calvin.seward@zalando.de
5/47

6. FASHION DNA: READING FASHION IMAGES
−→ SHOE!
6/47

7. FASHION DNA: READING FASHION IMAGES
−→
Gender–Male 79.9%
Gender–Female 29.5%
Age Group–Adult 99.5%
Domain–Sports 98.6%
Brand–Nike Performance 99.7%
Silhouette–Low Shoe 99.8%
Men’s Natural Running Shoe 40.6%
Men’s Lightweight Running Shoe 35.7%
Color–Green 56.1%
Upper Material–Textile/Synthetics 60.2%
Size–Runs Small 79.5%
. . . . . .
7/47

8. FASHION GAN: IMAGINE A SWEATER. . .
8/47

9. 9/47

10. DATA DRIVEN ORGANIZATIONS
1. All business relevant data is curated:
Collected, stored, and accessible
2. Management decisions are made based on data,
not intuition
3. User Experience is A/B tested and continuously improved
4. Business processes are automated and optimized
(buying, pricing, predictive maintenance)
5. Data unlocks new customer facing products
But how do we get there?
10/47

11. 11/47

12. NICCOLÒ MACHIAVELLI (1469–1527) — STRATEGIC CONSULTANT
It must be considered that there is nothing
more difficult to carry out nor more doubtful
of success nor more dangerous to handle
than to initiate a new order of things; for the
reformer has enemies in all those who profit
by the old order, and only lukewarm
defenders in all those who would profit by
the new order. . .
12/47

13. A PANOPLY OF POTENTIAL ADVERSARIES
Encounter. . .
• Reactionaries
• Bean Counters
• Revolutionaries
But consider. . .
• Some will be your colleagues
• They have good arguments, too
13/47

14. REACTIONARIES: THE LORD OF THE REALM
Owns and controls a business process
• Tightly managed
• Impenetrable to others
• Most reluctant member controls
speed of organizational innovation
14/47

15. REACTIONARIES: THE BUSINESS DINOSAUR
Kodak, Xerox, Nokia, Blackberry, . . .
• Reliable short term profit
• Lagging know-how
• Innovation may take company by
surprise
15/47

16. REACTIONARIES: THE LEADER FROM BEHIND
Outsources decisions, avoids
responsibility in uncertain circumstances
• Infuses external ideas
• Ignores internal company knowledge
• Not specific to product
16/47

17. REACTIONARIES: THE PESSIMIST
Suspicious of innovation
• Prevent ill-advised adventures
• Delays adoption of technology
• Detrimental to can-do attitude
17/47

18. BEAN COUNTERS: THE ACCOUNTANT
Emphasizes planning and controlling
mindset
• Manages cost and resources
• Stifles creativity and serendipity
• Ill-suited for rapidly evolving field
18/47

19. BEAN COUNTERS: THE BLACK BOX BELIEVER
Introduces AI "on the cheap" with
third-party services
• Lightweight solution
• Does not build competence
• Second-rate/failing performance
19/47

20. REVOLUTIONARIES: EVANGELISTS AND VISIONARIES
Excitable, optimistic, technophile
• Promotes innovation, thinks
beyond quarterly results
• Arms race of promises
• The future is hard to predict
20/47

21. REVOLUTIONARIES: THE PURIST
Always looking for the latest &
fanciest software and algorithm
• Keeps knowledge current
• Neglects business needs
• Diminishing returns
21/47

22. THE EAGLE HAS LANDED . . . IN COLOGNE
22/47

23. 23/47

24. ROME WAS NOT BUILT IN A DAY
If a company is going to commit to AI being a part of their future
business plans, they better commit to a long-term development
plan that might include several periods of rebuilding and
disruption. AI is going to go through several periods of both
slow and rapid change. -—
Todd Thibodeaux, CEO and President, Computing Technology
Industry Association (CompTIA)
24/47

25. THE ZALANDO SUCCESS STORY
WHEN YOU THINK OF HOW ZALANDO STARTED, YOU MAY THINK OF THIS
25/47

26. THE ZALANDO SUCCESS STORY
BUT IT STARTED MORE LIKE THIS
26/47

27. THE ZALANDO SUCCESS STORY
AND THEN GREW TO THIS
27/47

28. PICKER ROUTING PROBLEM
THE LOCATIONS THAT MUST BE VISITED
28/47

29. PICKER ROUTING PROBLEM
THE PICK ROUTE
29/47

30. OCAPI ALGORITHM
OPTIMAL CART PICK
• To solve picker routing
problem, we developed the
OCaPi Algorithm
• Calculates the optimal
route to walk
• Also determines optimal
cart handling strategy The Okapi – Our Mascot
30/47

31. ORDER BATCHING PROBLEM
RANDOM SPLIT OF 10 ORDERS À 2 ITEMS INTO TWO PICK ROUTES
31/47

32. ORDER BATCHING PROBLEM
OPTIMAL SPLIT OF 10 ORDERS À 2 ITEMS → 8.3% BOOST
32/47

33. NEURAL NETWORK ESTIMATE OF PICK ROUTE LENGTH
• OCaPi cost landscape
f : (N × R)n
→ R+
is a nice function because it is:
• Lipschitz continuous in the
real-valued arguments
• Piecewise linear in the
real-valued arguments
• Locally sensitive
• Perfect function to model with
Convolutional Neural Networks
with ReLUs:
˜f(x) := (W2(W1x + b1)+ + b2)+
33/47

34. NEURAL NETWORK ESTIMATE OF PICK ROUTE LENGTH
ESTIMATION ACCURACY – ESTIMATED TRAVEL TIME
CALCULATED TRAVEL TIME
34/47

35. OPTIMAL WAREHOUSE STORAGE LOCATION ALLOCATION
80
90
100
110
120
130
140
150
48
52
56
60
64
68
72
76
80
39
42
45
48
51
54
57
60
63
32.5
35.0
37.5
40.0
42.5
45.0
47.5
50.0
35/47

36. 36/47

37. START SMALL AND WIN BIG
• Identify a clear KPI that can be improved with data
• Leverage readily available and easily usable data
• Optimize existing processes, don’t create brand new ones
• E.g. refine recommendations, don’t create a new chatbot
37/47

38. START SMALL AND WIN BIG
• Identify a clear KPI that can be improved with data
• Leverage readily available and easily usable data
• Optimize existing processes, don’t create brand new ones
• E.g. refine recommendations, don’t create a new chatbot
During this process you will:
• Build trust with management and stakeholders
• Make data quality a must have
• Develop a vision for data in the organization
• Gain domain knowledge
• Improve data science skills
37/47

39. 38/47

40. BUILDING TRUST
• Take the time to understand other stakeholders
• Spend time explaining methods and results
• Agree on KPIs and demonstrate improvement
• Start with simpler methods
• Value other’s methods
39/47

41. BUILDING TRUST
• Take the time to understand other stakeholders
• Spend time explaining methods and results
• Agree on KPIs and demonstrate improvement
• Start with simpler methods
• Value other’s methods
Trust is key:
• For obtaining data
• Getting AI into production
• Changing business processes
39/47

42. DATA QUALITY
Joseph Wright the Alchemist
40/47

43. DATA QUALITY
• Data science isn’t alchemy
• Lead can’t be transformed
to gold, really bad data
can’t create value
Joseph Wright the Alchemist
40/47

44. DATA QUALITY
• Data science isn’t alchemy
• Lead can’t be transformed
to gold, really bad data
can’t create value
• But: As data becomes
important for business
processes, data quality will
become a priority too
Joseph Wright the Alchemist
40/47

45. VISION
• In business one would like
to know today what will
make big money
• In science if you already
knew what the
experimental results will
be, there’s no point in
experimenting
• Therefore to extract
business value from with
science experiment first
• And a month long data
science vision session
won’t get you anywhere
41/47

46. DOMAIN KNOWLEDGE
• Data generally has lots of
strange anomalies and
patterns
• Understanding the domain
is key to understanding
these patterns
• Understanding these
patterns is key to creating
value
42/47

47. DATA SKILLS
• We’re always learning
43/47

48. RAMPENSAU
Others will try to take credit for your success. So go ahead, be
a Rampensau and make sure everyone knows who’s
responsible for the success.
44/47

49. 45/47

50. REVIEW
Introduction to Zalando
Data Driven Organizations
AI Projects: A Fraught Landscape
AI Success Story: Warehouse Logistics at Zalando
Start Small to Win Big
46/47

51. Thanks for listening