Presentation used by Preyas Joshi during the Webinar ML Models for Product Managers. The audience for this presentation includes PMs, CEOs, and VPs of Product in any tech or non-tech organization.
1. ML Models in Product Management
Preyas Joshi
16 years in Product Management
Amazon, Ex-CTO of Fundation, entrepreneur, fintech
Led multi-disciplinary teams – PM, S/W Engineering, Applied Sci
P&L for multi-billion dollar businesses
Launched several multi-million dollar businesses
Product, Tech Advisor for several startups
Patent holder in secure QR code tech
Father and husband. Love to read, hike and stay fit.
2. ML Models in Product Management
Disclaimer
All opinions I state today, either in this video or when we
chat after this session, are my own and not the views of my
current or previous employer.
Examples, projects or user stories do NOT depict, nor
speculate about projects or experiments with my current or
former employers.
3. ML in Product Management
Preyas Joshi
What to expect
TYPES OF
MODELS
USER STORIES ORGANIZATIONAL
STRUCTURE
WHY ML?
WHY NOW?
HOW DO YOU
IDENTIFY AN
ML IDEA?
Supervised
Unsupervised
Reinforced
CRM, Lending,
Home Services,
Metaverse, Retail
Organizational
structures for ML
teams
Is ML right for
your
organization?
When ML fails
Some ideas
and
models
4. ML in Product Management
Preyas Joshi
Types of Models
Regression
Binary
Supervised Unsupervised
Association
Recommender
Clustering
Detect anomaly
Use your own
intelligence
Train the model
Run the model
• Analyze data
• Find similarities across
X vectors
• Group data into
clusters
• Associate known
attributes /
behavior / events
to link each other
Predict a value
Is this Yes/No,
X/Y?
Multi class
Is this of type
X, Y or Z?
Extract text Text Analysis
Incrementally
rewarded
intelligence
Reinforced
Model based Model free
Markov (MDP)
Q-Learning, SARSA (State-Action-
Reward-State-Action
5. ML in Product Management
Preyas Joshi
5
Types of Problems
The global machine learning market is predicted to grow 14X from $8.43 billion (2019) to $119.8B (2027).
$119.8B
Lending
Problem: As a lender, I want to predict the default rate of a business, based on a large
number of variables (industry, tenure, owner’s credit score, zip code, revenue, and
thousands of other variables, , so I can lend them money at the right rate, while
protecting investors.
Problem: As a student loan lender, I want to predict the likelihood of the loan being paid
based on the student’s college, profession, grades, GPA, and zip code, so I can support
more students while protecting investors.
Solution: Phase 1 – Train a model to predict the outcome of
a loan based on existing results. 100X Iterate and refine
model.
Regression, binary and multiclass classification, reinforced learning
Solution: Phase 1x - A back-tested self-learning
credit decisioning model based on prior data
and outcomes.
Industry
User Story / Epic
Possible solutions
ML model
6. ML in Product Management
Preyas Joshi
6
Types of Problems
The global machine learning market is predicted to grow 14X from $8.43 billion (2019) to $119.8B (2027).
$119.8B
Retail
Problem: Launching a new product in the market is risky. How can a
manufacturer produce inventory just in time, based on real-time demand?
Solution: Build an A/B test driven eCommerce tech integrated with manufacturing
that predicts demand based on order velocity.
Supervised: Binary regression, multiclass, clustering, reinforced learning
Industry
User Story / Epic
Possible solutions
ML model
7. ML in Product Management
Preyas Joshi
7
Types of Problems
The global machine learning market is predicted to grow 14X from $8.43 billion (2019) to $119.8B (2027).
$119.8B
Infrastructure
Problem: Can I proactively recommend maintenance for a bridge or machine
before the problem occurs?
Solution: Train a model to assess pictures of existing infrastructure that’s in disrepair to
predict the maintenance cycle based on current indicators.
Supervised: Binary regression, multiclass. Unsupervised: Association
Industry
User Story / Epic
Possible solutions
ML model
8. ML in Product Management
Preyas Joshi
8
Types of Problems
The global machine learning market is predicted to grow 14X from $8.43 billion (2019) to $119.8B (2027).
$119.8B
Infrastructure
Problem: As a Support Engineer, I want to reduce the time I spend searching
for troubleshooting tools to solve a technical problem, so I can resolve issues
quickly.
Solution: A tool recommendation model based on previously used tools to solve
similar cases
Supervised: multiclass. Unsupervised: Clustering + Association
Industry
User Story / Epic
Possible solutions
ML model
9. ML in Product Management
Preyas Joshi
9
Why ML?
48 percent of CIOs plan do deploy ML driven tech in the next 12 months. Look around. If you’re not
currently using ML in your industry, your competitor is.
48%
Should I invest in ML? Why now?
• Depends on your product lifecycle, competitive space, and opportunities
• Data health, readiness, access, data pipelines and infrastructure
• Where should you allocate your ML investments?
• How do you learn from failure? Culture of open-ended, rapid experimentation.
• Do you learn from success - Ethical dilemmas and removal of bias
• Organizational attitude to experimentation and high failure tolerance. Think like a VC!
10. ML in Product Management
Preyas Joshi
10
When can
ML projects fail?
Root causes
• Organization overinvested in ML
• Low engagement with business and competitive environment
• Low fidelity innovation
• Risk aversion
• Low experimentation
• Works sometimes. But carefully review the underlying use case.
Recommendations
• Review by internal board of ML advisors
• External audits
• Incubator mechanism
“To a man with a hammer, everything looks like a nail” – Mark Twain
11. ML in Product Management
Preyas Joshi
Organizational
Structure
ML
TEAM
PEOPLE, FINANCE
PRODUCT, DESIGN
TECH, IT
BIZ DEV, CUSTOMER
MARKETING, SALES
SUPPLY CHAIN
12. ML in Product Management
Preyas Joshi
12
Organizational
Structure
48 percent of CIOs plan do deploy ML driven tech in the next 12 months. Look around. If you’re not
currently using ML in your industry, your competitor is.
48%
Which team does an ML org belong to?
• Several options here.
• ML team and the hierarchy does not need to have a fixed structure / location
• Evolve as the business evolves
• Centralize vs break-up into parts
• Team / skill specialization
• Business lifecycle
• Consult (specific project based advise)
13. ML in Product Management
Preyas Joshi
Organizational
Structures
ML
TEAM
FINANCE
DESIGN
TECHNOLOGY
BUSINESS
DEVELOPMENT
MARKETING & SALES
SUPPLY CHAIN
Service org
M
L
FINANCE
DESIGN
TECHNOLOGY
BUSINESS
DEVELOPMENT
MARKETING
& SALES
SUPPLY
CHAIN
ML
M
L
Respective ML pods
FINANCE
DESIGN
TECHNOLOGY
BUSINESS
DEVELOPMENT
MARKETING
& SALES
SUPPLY
CHAIN
ML
External consultant
14. ML in Product Management
Preyas Joshi
Organizational
Structure
ML
TEAM
FINANCE
DESIGN
TECHNOLOGY
BUSINESS
DEVELOPMENT
MARKETING & SALES
SUPPLY CHAIN
Service org
• Needs multiple SMEs within the ML org
• High program/project management effort
• Teams could get underserved because
they didn’t identify their ML use case
• Deeper ML expertise within ML team
helps career progression and deeper learning
• Risk of generic solutions due to repetitive reuse of the same solution
• Teams need to understand when ML can be used for their use case
15. ML in Product Management
Preyas Joshi
Organizational
Structure
• Each team identifies their respective use case for ML
• Each ML team can develop specialized solutions for their business
• Could lead to silos
• Could lead to higher attrition if team doesn’t have
mentors and a growth path
• Allows for the ML team to internally verticalize along Functional,
Technical, and Product lines
• Each ML team needs to deeply understand business
M
L
FINANCE
DESIGN
TECHNOLOGY
BUSINESS
DEVELOPMENT
MARKETING
& SALES
SUPPLY
CHAIN
ML
M
L
Respective ML pods
16. ML in Product Management
Preyas Joshi
Organizational
Structure
• Temporary engagement with an existing ML PM leader
• Low overhead, low risk. Test and Learn. Invest incrementally.
• Hire temporary ML staffing. Double down when the
right problems are identified
• Each business group proposes / identifies a problem
they want to solve
FINANCE
DESIGN
TECHNOLOGY
BUSINESS
DEVELOPMENT
MARKETING
& SALES
SUPPLY
CHAIN
ML
External consultant
17. ML in Product Management
Preyas Joshi
17
How do you identify
an ML Idea?
• Teams present their use case to
the ML team
• ML team assesses validity of the
problem and whether (clean,
structured) data exists to test
and build outcomes
• Risk of identifying several small
problems, while big ideas
could get lost
• Need a mechanism to bubble
up big ideas
Office Hours Subject Matter Experts External Advisor
• ML PM in ML team work with
SMEs in various orgs
• ML PM assesses validity of
problems, and documents
detailed user stories and use
cases
• Presented to the ML experts on the
team. Proposals are promoted /
rejected based on validity
• Develops ML strength within the
PM team
• Consult an external ML Product
Advisor
• Review whether your problems,
data, infrastructure, and
problem space is conducive for
ML driven problem solving
18. ML in Product Management
Preyas Joshi
Poll
What Next?
EVALUATE ML
RELEVANCY
FOR A
PROBLEM
DEEP DIVE
INTO EACH ML
MODEL
WRITING
REQUIREMENTS
FOR ML
ENGINEERING
TEAMS
EXAMPLES OF
STARTUPS WHERE
PMS ARE USING
ML
OTHER