SlideShare a Scribd company logo
BUILDING AI-DRIVEN
PRODUCTS
MICHAEL FENG
ABOUT ME
2013: Table extraction from PDF files
ABOUT ME
2014-2015: Document analytics
ABOUT ME
2015-2017: Smart documents
BUILDING AI-DRIVEN PRODUCTS
AGENDA
▸ Someone needs to wear the PM hat
▸ Traditional PM process isn’t a good fit
▸ 4 habits of effective AI product managers
SOMEONE NEEDS TO
WEAR THE PM HAT
SOMEONE NEEDS TO WEAR THE PM HAT
ACADEMIA VS REAL WORLD
Characteristics
▸ Explicit problem
statements
▸ Well-structured
datasets
▸ Defined success
criteria
SOMEONE NEEDS TO WEAR THE PM HAT
ACADEMIA VS REAL WORLD
Goals
▸ Publish research paper
▸ Improve on established
benchmark vs past
methods ImageNet Classification with Deep
Convolutional Neural Networks (2012)
SOMEONE NEEDS TO WEAR THE PM HAT
ACADEMIA VS REAL WORLD
“I’d like you to explore our
server logs and use machine
learning to provide insights
to our customers.”
“That’d be grrreat.”
SOMEONE NEEDS TO WEAR THE PM HAT
AI PRODUCT MANAGERS OWN THE QUESTIONS
‣ What problem are we solving?
‣ Why is that problem valuable?
‣ Do we have relevant data that may solve this problem?
‣ How do we measure success?
‣ Are we building a system that learns?
I DEAL WITH THE GODDAMN CUSTOMERS
SO THAT THE ENGINEERS DON'T HAVE TO.
I HAVE PEOPLE SKILLS!
Product Manager
SOMEONE NEEDS TO WEAR THE PM HAT
YOU’RE WEARING
THE PM HAT
IF THERE’S NO PM
Founder
Data scientist
Manager
TRADITIONAL PM PROCESS
ISN’T A GOOD FIT
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
PROBLEM —> SOLUTION
1. Gather user stories
2. Define requirements
3. Design mockups
4. Perform acceptance
testing
EXAMPLE: NETFLIX FOR DOCUMENTS
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
PROBLEM: USERS DON’T USE THE PRODUCT!
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
SOLUTION: NEW USER TOUR
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
USER STORY
“As a new user, I want to know
what the top 3 features are so that
I can start using them”
MOCKUPS
Did all requirements pass?ACCEPTANCE TESTING
REQUIREMENTS
‣ User see a pop-up with a carousel of top 3 features,
and the background is opaque
‣ User can exit anytime by clicking on background
‣ Works on web, iOS, Android
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
PROBLEM: USERS STILL DON’T USE THE PRODUCT!
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
SOLUTION: IMPROVE RELEVANCE OF
CONTENT RECOMMENDATIONS
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
USER STORY
MOCKUPS
“As a user, I want to see content
shared by people I know”
Add a rule to the algorithm
ACCEPTANCE TESTING
REQUIREMENTS
Show documents shared by people in
same office
Were users are more likely to click
on results shown?
USER STORY #1
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
USER STORY
MOCKUPS
“As a user, I want to see content
related to my work”
Add another rule to the algorithm
ACCEPTANCE TESTING
REQUIREMENTS
Show documents shared by people in
same functional department
Were users are more likely to click
on results shown?
USER STORY #2
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
USER STORY
MOCKUPS
“As a user, I want to see content
about company social events”
Add yet another rule to the algorithm
ACCEPTANCE TESTING
REQUIREMENTS
Show documents shared by people on
the social committee
Were users are more likely to click
on results shown?
USER STORY #3
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
if (A)

return w1 * f1 + w2 * f2 + … + wN * fN

else if (B)…
A BUNCH OF HEURISTICS
TRADITIONAL PM PROCESS ISN’T A GOOD FIT
A set of heuristics isn’t AI because it’s not
a system that learns (i.e improve with
more data)
4 HABITS OF EFFECTIVE
AI PRODUCT MANAGERS
UNDERSTAND
THE MODEL
1
UNDERSTAND THE MODEL
BASIC EXAMPLE: COLLABORATIVE FILTERING
▸ Users “rate”
documents
▸ Sparse ratings matrix
▸ Latent user and
document factors
▸ Predict unknown
ratings
UNDERSTAND THE MODEL
Classes
‣ 2 = Strong interest
‣ 1 = Lukewarm interest
‣ 0 = No interest
Blank = Not rated
Doc 1 Doc 2 Doc 3
User 1 2 1
User 2 1 2
User 3 1 0 ?
RATINGS MATRIX
UNDERSTAND THE MODEL
User vector
1.5
2.0
2.0
-1.0
Document vector
0.5
-1.5
1.0
-2.0
1.75
LATENT FACTORS
Predicted
rating
UNDERSTAND THE MODEL
Doc 1 Doc 2 Doc 3
User 1 2 1
User 2 1 2
User 3 1 0
FITTING THE MODEL
Training
Doc 1 Doc 2 Doc 3
User 1 1.95 0.53 0.91
User 2 1.08 1.35 2.62
User 3 1.33 0.03 1.75
Prediction
Adjust latent factors to minimize
prediction error vs training data
UNDERSTAND THE MODEL
MAKING THE PREDICTION
Show the user documents:
‣ with the highest predicted
ratings
‣ not owned by him/her
GET LABELED
DATA
2
Labeled
Unlabeled data is the new dirt
GET LABELED DATA
Labeled data has explicit labels of
exactly what you’re trying to predict
GET LABELED DATA
Not labeled data
Server logs
Satellite images
Files in S3
LABELED VS UNLABELED DATA
Labeled data
+ which requests came
from malicious IPs
(hack detector)
+ markers of pools
(pool detector)
+ marked confidential
or not
(confidentiality classifier)
GET LABELED DATA
HOW TO GET LABELED DATA
Outsourced labeling
Brute force
HOW LONG DOES IT
TAKE TO GET 10K
SAMPLES?
User feedback
Transfer learning
CONSTRUCT A
FEEDBACK LOOP
3
CONSTRUCT A FEEDBACK LOOP
Display predictions
Collect user
feedback
Add to training set
Re-fit model
CONSTRUCT A FEEDBACK LOOP
For each document shown, assign rating:
‣ 1 if user opens the document once
‣ 2 if the user opens the document again
‣ 0 if the document has been shown 3+
times and not opened
SAME DIMENSIONS AS YOUR TRAINING DATA
CONSTRUCT A FEEDBACK LOOP
Doc 1 Doc 2 Doc 3
User 1 2 1
User 2 1 2
User 3 1 0 1
SAME DIMENSIONS AS YOUR TRAINING DATA
CONSTRUCT A FEEDBACK LOOP
Having the feedback loop on day 1
means you’re not wasting valuable
new labeled data
MEASURE RELATIVE
PERFORMANCE
4
MEASURE RELATIVE PERFORMANCE
THE FIRST VERSION WILL PROBABLY SUCK
MEASURE RELATIVE PERFORMANCE
Relative performance tells you how
much more data you need to get to
“good enough”
MEASURE RELATIVE PERFORMANCE
Accuracy
Amount of
data
“good enough”
How much data
you need to get
to “good enough”
slope
=
relative
perform
ance
PUTTING IT ALL TOGETHER
PUTTING IT ALL TOGETHER
LABELED DATA
MODEL
Predictions
FEEDBACKLOOP
RELATIVE
PERFORMANCE
1. Understand the
model
2. Get labeled data
3. Construct a
feedback loop
4. Measure relative
performance
Thanks!
Michael Feng

mifeng@gmail.com

More Related Content

Similar to Building AI-driven products

Intro to Data Analytics with Oscar's Director of Product
 Intro to Data Analytics with Oscar's Director of Product Intro to Data Analytics with Oscar's Director of Product
Intro to Data Analytics with Oscar's Director of Product
Product School
 
Design + Devops: What We've Learned from Our Developer Friends
Design + Devops: What We've Learned from Our Developer FriendsDesign + Devops: What We've Learned from Our Developer Friends
Design + Devops: What We've Learned from Our Developer Friends
UXPA International
 
Lean User Testing Intro
Lean User Testing IntroLean User Testing Intro
Lean User Testing Intro
Jessica DuVerneay
 
ODSC West 2022 – Kitbashing in ML
ODSC West 2022 – Kitbashing in MLODSC West 2022 – Kitbashing in ML
ODSC West 2022 – Kitbashing in ML
Bryan Bischof
 
DataSpryng Overview
DataSpryng OverviewDataSpryng Overview
DataSpryng Overview
jkvr
 
Jump Start Agile Testing with Acceptance Test Driven Development
Jump Start Agile Testing with Acceptance Test Driven DevelopmentJump Start Agile Testing with Acceptance Test Driven Development
Jump Start Agile Testing with Acceptance Test Driven Development
TechWell
 
Advanced A/B Testing - Jax London 2015
Advanced A/B Testing - Jax London 2015Advanced A/B Testing - Jax London 2015
Advanced A/B Testing - Jax London 2015
Aviran Mordo
 
[UPDATE] Udacity webinar on Recommendation Systems
[UPDATE] Udacity webinar on Recommendation Systems[UPDATE] Udacity webinar on Recommendation Systems
[UPDATE] Udacity webinar on Recommendation Systems
Axel de Romblay
 
Advanced A/B testing - Aviran Mordo
Advanced A/B testing - Aviran MordoAdvanced A/B testing - Aviran Mordo
Advanced A/B testing - Aviran Mordo
JAXLondon_Conference
 
Bridging Current Reality & Future Vision with Reality Maps
Bridging Current Reality & Future Vision with Reality MapsBridging Current Reality & Future Vision with Reality Maps
Bridging Current Reality & Future Vision with Reality Maps
Malini Rao
 
Udacity webinar on Recommendation Systems
Udacity webinar on Recommendation SystemsUdacity webinar on Recommendation Systems
Udacity webinar on Recommendation Systems
Axel de Romblay
 
Using OBIEE and Data Vault to Virtualize Your BI Environment: An Agile Approach
Using OBIEE and Data Vault to Virtualize Your BI Environment: An Agile ApproachUsing OBIEE and Data Vault to Virtualize Your BI Environment: An Agile Approach
Using OBIEE and Data Vault to Virtualize Your BI Environment: An Agile Approach
Kent Graziano
 
Course 8 : How to start your big data project by Eric Rodriguez
Course 8 : How to start your big data project by Eric Rodriguez Course 8 : How to start your big data project by Eric Rodriguez
Course 8 : How to start your big data project by Eric Rodriguez
Betacowork
 
Presentation for ArenaLviv 2017
Presentation for ArenaLviv 2017Presentation for ArenaLviv 2017
Presentation for ArenaLviv 2017
Andriy Dyadyura
 
UserTesting 2016 webinar: Research to inform product design in Agile environm...
UserTesting 2016 webinar: Research to inform product design in Agile environm...UserTesting 2016 webinar: Research to inform product design in Agile environm...
UserTesting 2016 webinar: Research to inform product design in Agile environm...
Steve Fadden
 
Data Science Demystified
Data Science DemystifiedData Science Demystified
Data Science Demystified
Emily Robinson
 
L injection toward effective collaborative filtering using uninteresting items
L injection toward effective collaborative filtering using uninteresting itemsL injection toward effective collaborative filtering using uninteresting items
L injection toward effective collaborative filtering using uninteresting items
Kumar Dlk
 
ONA (organizational network analysis) - enabling individuals to impact their ...
ONA (organizational network analysis) - enabling individuals to impact their ...ONA (organizational network analysis) - enabling individuals to impact their ...
ONA (organizational network analysis) - enabling individuals to impact their ...
Agron Fazliu
 
Linked Open Data Principles, benefits of LOD for sustainable development
Linked Open Data Principles, benefits of LOD for sustainable developmentLinked Open Data Principles, benefits of LOD for sustainable development
Linked Open Data Principles, benefits of LOD for sustainable development
Martin Kaltenböck
 
Agile user story mapping
Agile user story mappingAgile user story mapping
Agile user story mapping
Tam Nguyen, PMP, CSM
 

Similar to Building AI-driven products (20)

Intro to Data Analytics with Oscar's Director of Product
 Intro to Data Analytics with Oscar's Director of Product Intro to Data Analytics with Oscar's Director of Product
Intro to Data Analytics with Oscar's Director of Product
 
Design + Devops: What We've Learned from Our Developer Friends
Design + Devops: What We've Learned from Our Developer FriendsDesign + Devops: What We've Learned from Our Developer Friends
Design + Devops: What We've Learned from Our Developer Friends
 
Lean User Testing Intro
Lean User Testing IntroLean User Testing Intro
Lean User Testing Intro
 
ODSC West 2022 – Kitbashing in ML
ODSC West 2022 – Kitbashing in MLODSC West 2022 – Kitbashing in ML
ODSC West 2022 – Kitbashing in ML
 
DataSpryng Overview
DataSpryng OverviewDataSpryng Overview
DataSpryng Overview
 
Jump Start Agile Testing with Acceptance Test Driven Development
Jump Start Agile Testing with Acceptance Test Driven DevelopmentJump Start Agile Testing with Acceptance Test Driven Development
Jump Start Agile Testing with Acceptance Test Driven Development
 
Advanced A/B Testing - Jax London 2015
Advanced A/B Testing - Jax London 2015Advanced A/B Testing - Jax London 2015
Advanced A/B Testing - Jax London 2015
 
[UPDATE] Udacity webinar on Recommendation Systems
[UPDATE] Udacity webinar on Recommendation Systems[UPDATE] Udacity webinar on Recommendation Systems
[UPDATE] Udacity webinar on Recommendation Systems
 
Advanced A/B testing - Aviran Mordo
Advanced A/B testing - Aviran MordoAdvanced A/B testing - Aviran Mordo
Advanced A/B testing - Aviran Mordo
 
Bridging Current Reality & Future Vision with Reality Maps
Bridging Current Reality & Future Vision with Reality MapsBridging Current Reality & Future Vision with Reality Maps
Bridging Current Reality & Future Vision with Reality Maps
 
Udacity webinar on Recommendation Systems
Udacity webinar on Recommendation SystemsUdacity webinar on Recommendation Systems
Udacity webinar on Recommendation Systems
 
Using OBIEE and Data Vault to Virtualize Your BI Environment: An Agile Approach
Using OBIEE and Data Vault to Virtualize Your BI Environment: An Agile ApproachUsing OBIEE and Data Vault to Virtualize Your BI Environment: An Agile Approach
Using OBIEE and Data Vault to Virtualize Your BI Environment: An Agile Approach
 
Course 8 : How to start your big data project by Eric Rodriguez
Course 8 : How to start your big data project by Eric Rodriguez Course 8 : How to start your big data project by Eric Rodriguez
Course 8 : How to start your big data project by Eric Rodriguez
 
Presentation for ArenaLviv 2017
Presentation for ArenaLviv 2017Presentation for ArenaLviv 2017
Presentation for ArenaLviv 2017
 
UserTesting 2016 webinar: Research to inform product design in Agile environm...
UserTesting 2016 webinar: Research to inform product design in Agile environm...UserTesting 2016 webinar: Research to inform product design in Agile environm...
UserTesting 2016 webinar: Research to inform product design in Agile environm...
 
Data Science Demystified
Data Science DemystifiedData Science Demystified
Data Science Demystified
 
L injection toward effective collaborative filtering using uninteresting items
L injection toward effective collaborative filtering using uninteresting itemsL injection toward effective collaborative filtering using uninteresting items
L injection toward effective collaborative filtering using uninteresting items
 
ONA (organizational network analysis) - enabling individuals to impact their ...
ONA (organizational network analysis) - enabling individuals to impact their ...ONA (organizational network analysis) - enabling individuals to impact their ...
ONA (organizational network analysis) - enabling individuals to impact their ...
 
Linked Open Data Principles, benefits of LOD for sustainable development
Linked Open Data Principles, benefits of LOD for sustainable developmentLinked Open Data Principles, benefits of LOD for sustainable development
Linked Open Data Principles, benefits of LOD for sustainable development
 
Agile user story mapping
Agile user story mappingAgile user story mapping
Agile user story mapping
 

Recently uploaded

Full-RAG: A modern architecture for hyper-personalization
Full-RAG: A modern architecture for hyper-personalizationFull-RAG: A modern architecture for hyper-personalization
Full-RAG: A modern architecture for hyper-personalization
Zilliz
 
How to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For FlutterHow to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For Flutter
Daiki Mogmet Ito
 
GenAI Pilot Implementation in the organizations
GenAI Pilot Implementation in the organizationsGenAI Pilot Implementation in the organizations
GenAI Pilot Implementation in the organizations
kumardaparthi1024
 
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdf
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdfAI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdf
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdf
Techgropse Pvt.Ltd.
 
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdf
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdfUnlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdf
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdf
Malak Abu Hammad
 
AI 101: An Introduction to the Basics and Impact of Artificial Intelligence
AI 101: An Introduction to the Basics and Impact of Artificial IntelligenceAI 101: An Introduction to the Basics and Impact of Artificial Intelligence
AI 101: An Introduction to the Basics and Impact of Artificial Intelligence
IndexBug
 
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development ProvidersYour One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
akankshawande
 
Mariano G Tinti - Decoding SpaceX
Mariano G Tinti - Decoding SpaceXMariano G Tinti - Decoding SpaceX
Mariano G Tinti - Decoding SpaceX
Mariano Tinti
 
“I’m still / I’m still / Chaining from the Block”
“I’m still / I’m still / Chaining from the Block”“I’m still / I’m still / Chaining from the Block”
“I’m still / I’m still / Chaining from the Block”
Claudio Di Ciccio
 
20240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 202420240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 2024
Matthew Sinclair
 
Choosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptxChoosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptx
Brandon Minnick, MBA
 
Things to Consider When Choosing a Website Developer for your Website | FODUU
Things to Consider When Choosing a Website Developer for your Website | FODUUThings to Consider When Choosing a Website Developer for your Website | FODUU
Things to Consider When Choosing a Website Developer for your Website | FODUU
FODUU
 
Building Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and MilvusBuilding Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and Milvus
Zilliz
 
Best 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERPBest 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERP
Pixlogix Infotech
 
UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6
DianaGray10
 
Infrastructure Challenges in Scaling RAG with Custom AI models
Infrastructure Challenges in Scaling RAG with Custom AI modelsInfrastructure Challenges in Scaling RAG with Custom AI models
Infrastructure Challenges in Scaling RAG with Custom AI models
Zilliz
 
Taking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdfTaking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdf
ssuserfac0301
 
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
Edge AI and Vision Alliance
 
Columbus Data & Analytics Wednesdays - June 2024
Columbus Data & Analytics Wednesdays - June 2024Columbus Data & Analytics Wednesdays - June 2024
Columbus Data & Analytics Wednesdays - June 2024
Jason Packer
 
Programming Foundation Models with DSPy - Meetup Slides
Programming Foundation Models with DSPy - Meetup SlidesProgramming Foundation Models with DSPy - Meetup Slides
Programming Foundation Models with DSPy - Meetup Slides
Zilliz
 

Recently uploaded (20)

Full-RAG: A modern architecture for hyper-personalization
Full-RAG: A modern architecture for hyper-personalizationFull-RAG: A modern architecture for hyper-personalization
Full-RAG: A modern architecture for hyper-personalization
 
How to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For FlutterHow to use Firebase Data Connect For Flutter
How to use Firebase Data Connect For Flutter
 
GenAI Pilot Implementation in the organizations
GenAI Pilot Implementation in the organizationsGenAI Pilot Implementation in the organizations
GenAI Pilot Implementation in the organizations
 
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdf
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdfAI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdf
AI-Powered Food Delivery Transforming App Development in Saudi Arabia.pdf
 
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdf
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdfUnlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdf
Unlock the Future of Search with MongoDB Atlas_ Vector Search Unleashed.pdf
 
AI 101: An Introduction to the Basics and Impact of Artificial Intelligence
AI 101: An Introduction to the Basics and Impact of Artificial IntelligenceAI 101: An Introduction to the Basics and Impact of Artificial Intelligence
AI 101: An Introduction to the Basics and Impact of Artificial Intelligence
 
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development ProvidersYour One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
 
Mariano G Tinti - Decoding SpaceX
Mariano G Tinti - Decoding SpaceXMariano G Tinti - Decoding SpaceX
Mariano G Tinti - Decoding SpaceX
 
“I’m still / I’m still / Chaining from the Block”
“I’m still / I’m still / Chaining from the Block”“I’m still / I’m still / Chaining from the Block”
“I’m still / I’m still / Chaining from the Block”
 
20240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 202420240605 QFM017 Machine Intelligence Reading List May 2024
20240605 QFM017 Machine Intelligence Reading List May 2024
 
Choosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptxChoosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptx
 
Things to Consider When Choosing a Website Developer for your Website | FODUU
Things to Consider When Choosing a Website Developer for your Website | FODUUThings to Consider When Choosing a Website Developer for your Website | FODUU
Things to Consider When Choosing a Website Developer for your Website | FODUU
 
Building Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and MilvusBuilding Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and Milvus
 
Best 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERPBest 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERP
 
UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6UiPath Test Automation using UiPath Test Suite series, part 6
UiPath Test Automation using UiPath Test Suite series, part 6
 
Infrastructure Challenges in Scaling RAG with Custom AI models
Infrastructure Challenges in Scaling RAG with Custom AI modelsInfrastructure Challenges in Scaling RAG with Custom AI models
Infrastructure Challenges in Scaling RAG with Custom AI models
 
Taking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdfTaking AI to the Next Level in Manufacturing.pdf
Taking AI to the Next Level in Manufacturing.pdf
 
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
“Building and Scaling AI Applications with the Nx AI Manager,” a Presentation...
 
Columbus Data & Analytics Wednesdays - June 2024
Columbus Data & Analytics Wednesdays - June 2024Columbus Data & Analytics Wednesdays - June 2024
Columbus Data & Analytics Wednesdays - June 2024
 
Programming Foundation Models with DSPy - Meetup Slides
Programming Foundation Models with DSPy - Meetup SlidesProgramming Foundation Models with DSPy - Meetup Slides
Programming Foundation Models with DSPy - Meetup Slides
 

Building AI-driven products

  • 2. ABOUT ME 2013: Table extraction from PDF files
  • 5. BUILDING AI-DRIVEN PRODUCTS AGENDA ▸ Someone needs to wear the PM hat ▸ Traditional PM process isn’t a good fit ▸ 4 habits of effective AI product managers
  • 7. SOMEONE NEEDS TO WEAR THE PM HAT ACADEMIA VS REAL WORLD Characteristics ▸ Explicit problem statements ▸ Well-structured datasets ▸ Defined success criteria
  • 8. SOMEONE NEEDS TO WEAR THE PM HAT ACADEMIA VS REAL WORLD Goals ▸ Publish research paper ▸ Improve on established benchmark vs past methods ImageNet Classification with Deep Convolutional Neural Networks (2012)
  • 9. SOMEONE NEEDS TO WEAR THE PM HAT ACADEMIA VS REAL WORLD “I’d like you to explore our server logs and use machine learning to provide insights to our customers.” “That’d be grrreat.”
  • 10. SOMEONE NEEDS TO WEAR THE PM HAT AI PRODUCT MANAGERS OWN THE QUESTIONS ‣ What problem are we solving? ‣ Why is that problem valuable? ‣ Do we have relevant data that may solve this problem? ‣ How do we measure success? ‣ Are we building a system that learns?
  • 11. I DEAL WITH THE GODDAMN CUSTOMERS SO THAT THE ENGINEERS DON'T HAVE TO. I HAVE PEOPLE SKILLS! Product Manager SOMEONE NEEDS TO WEAR THE PM HAT
  • 12. YOU’RE WEARING THE PM HAT IF THERE’S NO PM Founder Data scientist Manager
  • 14. TRADITIONAL PM PROCESS ISN’T A GOOD FIT PROBLEM —> SOLUTION 1. Gather user stories 2. Define requirements 3. Design mockups 4. Perform acceptance testing
  • 15. EXAMPLE: NETFLIX FOR DOCUMENTS
  • 16. TRADITIONAL PM PROCESS ISN’T A GOOD FIT PROBLEM: USERS DON’T USE THE PRODUCT!
  • 17. TRADITIONAL PM PROCESS ISN’T A GOOD FIT SOLUTION: NEW USER TOUR
  • 18. TRADITIONAL PM PROCESS ISN’T A GOOD FIT USER STORY “As a new user, I want to know what the top 3 features are so that I can start using them” MOCKUPS Did all requirements pass?ACCEPTANCE TESTING REQUIREMENTS ‣ User see a pop-up with a carousel of top 3 features, and the background is opaque ‣ User can exit anytime by clicking on background ‣ Works on web, iOS, Android
  • 19. TRADITIONAL PM PROCESS ISN’T A GOOD FIT PROBLEM: USERS STILL DON’T USE THE PRODUCT!
  • 20. TRADITIONAL PM PROCESS ISN’T A GOOD FIT SOLUTION: IMPROVE RELEVANCE OF CONTENT RECOMMENDATIONS
  • 21. TRADITIONAL PM PROCESS ISN’T A GOOD FIT USER STORY MOCKUPS “As a user, I want to see content shared by people I know” Add a rule to the algorithm ACCEPTANCE TESTING REQUIREMENTS Show documents shared by people in same office Were users are more likely to click on results shown? USER STORY #1
  • 22. TRADITIONAL PM PROCESS ISN’T A GOOD FIT USER STORY MOCKUPS “As a user, I want to see content related to my work” Add another rule to the algorithm ACCEPTANCE TESTING REQUIREMENTS Show documents shared by people in same functional department Were users are more likely to click on results shown? USER STORY #2
  • 23. TRADITIONAL PM PROCESS ISN’T A GOOD FIT USER STORY MOCKUPS “As a user, I want to see content about company social events” Add yet another rule to the algorithm ACCEPTANCE TESTING REQUIREMENTS Show documents shared by people on the social committee Were users are more likely to click on results shown? USER STORY #3
  • 24. TRADITIONAL PM PROCESS ISN’T A GOOD FIT if (A)
 return w1 * f1 + w2 * f2 + … + wN * fN
 else if (B)… A BUNCH OF HEURISTICS
  • 25. TRADITIONAL PM PROCESS ISN’T A GOOD FIT A set of heuristics isn’t AI because it’s not a system that learns (i.e improve with more data)
  • 26. 4 HABITS OF EFFECTIVE AI PRODUCT MANAGERS
  • 28. UNDERSTAND THE MODEL BASIC EXAMPLE: COLLABORATIVE FILTERING ▸ Users “rate” documents ▸ Sparse ratings matrix ▸ Latent user and document factors ▸ Predict unknown ratings
  • 29. UNDERSTAND THE MODEL Classes ‣ 2 = Strong interest ‣ 1 = Lukewarm interest ‣ 0 = No interest Blank = Not rated Doc 1 Doc 2 Doc 3 User 1 2 1 User 2 1 2 User 3 1 0 ? RATINGS MATRIX
  • 30. UNDERSTAND THE MODEL User vector 1.5 2.0 2.0 -1.0 Document vector 0.5 -1.5 1.0 -2.0 1.75 LATENT FACTORS Predicted rating
  • 31. UNDERSTAND THE MODEL Doc 1 Doc 2 Doc 3 User 1 2 1 User 2 1 2 User 3 1 0 FITTING THE MODEL Training Doc 1 Doc 2 Doc 3 User 1 1.95 0.53 0.91 User 2 1.08 1.35 2.62 User 3 1.33 0.03 1.75 Prediction Adjust latent factors to minimize prediction error vs training data
  • 32. UNDERSTAND THE MODEL MAKING THE PREDICTION Show the user documents: ‣ with the highest predicted ratings ‣ not owned by him/her
  • 35. GET LABELED DATA Labeled data has explicit labels of exactly what you’re trying to predict
  • 36. GET LABELED DATA Not labeled data Server logs Satellite images Files in S3 LABELED VS UNLABELED DATA Labeled data + which requests came from malicious IPs (hack detector) + markers of pools (pool detector) + marked confidential or not (confidentiality classifier)
  • 37. GET LABELED DATA HOW TO GET LABELED DATA Outsourced labeling Brute force HOW LONG DOES IT TAKE TO GET 10K SAMPLES? User feedback Transfer learning
  • 39. CONSTRUCT A FEEDBACK LOOP Display predictions Collect user feedback Add to training set Re-fit model
  • 40. CONSTRUCT A FEEDBACK LOOP For each document shown, assign rating: ‣ 1 if user opens the document once ‣ 2 if the user opens the document again ‣ 0 if the document has been shown 3+ times and not opened SAME DIMENSIONS AS YOUR TRAINING DATA
  • 41. CONSTRUCT A FEEDBACK LOOP Doc 1 Doc 2 Doc 3 User 1 2 1 User 2 1 2 User 3 1 0 1 SAME DIMENSIONS AS YOUR TRAINING DATA
  • 42. CONSTRUCT A FEEDBACK LOOP Having the feedback loop on day 1 means you’re not wasting valuable new labeled data
  • 44. MEASURE RELATIVE PERFORMANCE THE FIRST VERSION WILL PROBABLY SUCK
  • 45. MEASURE RELATIVE PERFORMANCE Relative performance tells you how much more data you need to get to “good enough”
  • 46. MEASURE RELATIVE PERFORMANCE Accuracy Amount of data “good enough” How much data you need to get to “good enough” slope = relative perform ance
  • 47. PUTTING IT ALL TOGETHER
  • 48. PUTTING IT ALL TOGETHER LABELED DATA MODEL Predictions FEEDBACKLOOP RELATIVE PERFORMANCE 1. Understand the model 2. Get labeled data 3. Construct a feedback loop 4. Measure relative performance