Danse avec les unicorns : la data science en agile, de l'exploration à l'adoption

Conférence par Wassel Alazhar lors de l'Agile Tour Aix-Marseille 2018.

"La data science sort enfin des grottes des scientifiques. Tout le monde est convaincu que cette discipline changera la donne.
Les outils et les plateformes foisonnent et les startups qui ont fait ce pari sont devenues des unicorns.

Cependant, construire des applications qui apportent une réelle valeur ajoutée reste un vrai challenge. Plusieurs projets prometteurs n'ont pas réussi à dépasser le stade du PoC.

L'année dernière, j'ai accompagné mon client dans son aventure avec une "licorne" californienne qui produit une plateforme de Big data analytics et AI.
Dans cette session, nous allons voir comment faire une data science agile pour l'ajout constant de valeur.

On s'intéressera à :

– l'exploration de la valeur métier
– la collaboration entre différents domaines d'expertise (data science, data integration, app development, UX…)
– la qualité du produit dans un domaine exploratoire

Une agilité à tous les étages !"

Transcript

1. Dances with unicorns Agile datascience from exploration to adoption
2. Un grand merci à nos sponsors et partenaires
3. Oman 11.30 AM
4. Paris 9.45 AM
5. <24h From user’s feedback to production
6. One month later...
7. This project has been abandoned
8. That’s me! Wassel Alazhar Consultant, developer, problem solver @wasselovski https://github.com/jcraftsman
9. Agenda • The full story • What went wrong? • What did we learn? ◦ How to bring value from datascience ◦ Explore and build ◦ Efficient collaboration ◦ Product quality • Why this talk? • Takeaways
10. The full story
11. A global energy leader
12. A global energy leader Produce Deliver SELL
13. A global energy leader Produce Deliver SELL Sensors everywhere! All along the value chain
14. The problem to solve Produce Deliver SELL Sensors everywhere! All along the value chain
15. The problem to solve Two new generation power plants They are exactly the same but... Twin A Twin B
16. The problem to solve Twin B is way more performant (i.e., makes money) Twin A Twin B
17. The solution Datascience can help identifying better operational models for the power plants
18. The solution BIG DATA + DATA SCIENCE = MAGIC
19. The partner
20. The partner A unicorn is a privately held startup company valued at over $1 billion
21. The bill But wait… How much is that? Nevermind. It’s all on me! It’s called innovation. Great!
22. The team Data engineers Data scientists App developers
23. To the Silicon valley Data engineers Data scientists App developers
24. Week after week… Demo after demo
25. It couldn’t be any better
26. SURPRISE Now, it’s all yours! All you have to pay for is the run. Oh! No, thanks. I’m out of it.
27. Deception
28. What went wrong?
29. Building a software!
30. What was the Problem to solve? Do you remember the twin power plants? Twin A Twin B
31. Not what we’ve expected... Problem solved explained quickly No actionable findings
32. Instead we have delivered features! Degradation analysis Anomaly detection The software can detect dust in the steam turbine! PCA???
33. Feature ≠ VALUE
34. What did we learn?
35. Happy ending stories... Predictive maintenance Smart buildings Ice detection Heating and cooling efficiency
36. Business use case discovery Don’t start with a software! Explore Observe Confirm hypothesis or not! Discover
37. Business use case discovery Don’t explore in a dark lab! Get feedback!
38. Business use case discovery A python notebook is not a software! It’s a tool for a study!
39. From study to product delivery Business use case located? Build! VALUE
40. Product delivery Not like this!
41. Wait, what does datascience look like in 2018? How would you write a program for puppy recognition?
42. Wait, what does datascience look like in 2018? You can: • Try to define what a puppy face is • Code all these rules! Or, use Machine learning: • Show a lot of puppy faces examples! You don’t need to tell the algorithm what to do. All you need is to show it a lot of examples!
43. Wait, what does datascience look like in 2018? Take care of your examples (data pipeline) Verify the results (predictions)
44. Putting it all together Discovery: Given a real world pictures sample, would it be possible to recognize a puppy face? The answer is 86% yes, 13% muffins, 1% unknown. Product: Play a dog kibble comercial whenever a puppy picture is displayed!
45. Explore and build Explore: • Gathering data • Cleaning data • Feature engineering • Defining model • Training • Predicting the output => Discover what you are able to do with your data Build: • Data acquisition • Data filtering • Use model configuration • Use model • Training (or use a train set) • Predicting the output => Steadily bring value from your data
46. Explore and build iteratively Explore: • Gathering data • Cleaning data • Feature engineering • Defining model • Training • Predicting the output => Discover what you are able to do with your data Build: • Data acquisition • Data filtering • Use model configuration • Use model • Training (or use a train set) • Predicting the output => Steadily bring value from your data
47. Explore and build iteratively Explore Build Business use case discovery Product delivery
48. Product delivery You’re not done with datascience! They should build together!
49. Building together Code review When? All the time! Who? Everyone! Why? Quality, collective ownership and joy!
50. Building together Pair programming When? All the time! Who? Everyone! Why? Quality, collective ownership and joy!
51. Building together Mob programming When? Whenever you start something new or complex. Who? Everyone! Why? Collective intelligence, collective ownership, quality and joy!
52. Building together TDD Let’s be serious! When? Whenever you change the product’s behaviour. Who? Everyone working on the product! Why? Collective intelligence, collective ownership, quality and joy!
53. Building together TDD Have you ever met a data scientist who write unit tests and refactor? I did! :) It’s hard to imagine doing TDD during an exploratory work though! (i.e., when the target observable behaviour is not yet defined)
54. Product delivery Spikes and user stories
55. Product delivery essentials Don’t lose time repeating boring stuff! Automate! Make data available for everyone! Don’t treat your infra like pets! Destroy and rebuild! Don’t over-engineer though!
56. Product adoption Stay close to the users! Don’t plan too many features! Incorporate feedback!
57. What is agile anyway?
58. Can datascience be agile? It’s still true! Even for: • Big data • AI • Datascience
59. Why this talk?
60. Myths about datascience Well… Things have slightly changed since then… But not that much!
61. Myths about datascience
62. Myths about datascience
63. Unicorns
64. New unicorns - Same old stories You should draw your entire model before you start coding! Open a ticket! You need to hire a machine learning engineer!
65. Takeaways!
66. Takeaways! Make people together! Business value discovery => product delivery Explore and build iteratively Agile is still: • Short feedback • Small increments • Take engineering seriously work learn
67. OCTO © 2018 - Reproduction interdite sans autorisation écrite préalable 67 OCTO Provence recrute ! C’EST AVANT TOUT UN ÉTAT D’ESPRIT START-UP APPUYÉ PAR DES EXPERTISES TECH, AGILE & CHANGE POUR ACCOMPAGNER DIGITALE TRANSFORMATION NOS CLIENTS DANS LEUR Contactez-nous sur PROVENCE@OCTO.COM