Introduction to ArtificiaI Intelligence in Higher Education
Omnyscope e245 march 2014 final
1. omnyscope
Lessons Learned
Interviews: 107 / Hypotheses: 66
(Interviews this week: 10)
Initial idea:
Sentiment Centered News Aggregation
We provide an easy interface to consume news
articles with diverse opinions
Final Video: http://www.youtube.com/watch?v=9srMlWcgzKA
2. Arijit Banerjee
Pararth Shah
Shantanu Joshi
Sujeet Gholap
Background
MS CS
MS CS
MS CS
MS CS
Systems
UI/UX
Hustler
Hacker, Designer
Expertise
Algos
AI/Machine Learning
Role on team
Hacker, Designer
Hustler,
Product Picker
-- Four engineers with zero business experience --
3. Our Journey
Ton of
learnings!
Product/Market Fit!
Optimized cooking:
Engineer’s fantasy!
Got into
the
class!
Resegment!
Reco-as-a-service
MVP
Not a viable
business
Wireframe
Recipe
Recommendations
First
presentation
Restart
Data access
issues
Pivot
No revenue stream!
No market
“This is either the
dumbest or the
smartest idea I
have seen lately”
4. Week 1: Sentiment based news aggregation
Key Partners
Key Activities
Content
Generators:
Forming new
partnerships to
obtain data
News Sites
Editorials
Personal blog
sites
Building web and
mobile MVP
Key
Resources
IT infrastructure
Intellectual
Property
Cost
Structure
Value
Propositions
360 degree view
on a topic of your
choice
Bursting the filter
bubble
Provides an easy
medium for
collecting a
diverse set of
opinions
MVP:
Specialised
product for one
customer
segment
Customer
Relationships
Customer
Segments
Feedback on
content quality
Professional
users:
Recommendation
s
Investors
Business
analysts
Channels
PR firms
Individuals:
Content
aggregator - web
and mobile app
Browser plugin
Revenue
Streams
Server and bandwidth
Advertisements
Paid APIs for data access
Freemium model
Researchers
Young adults
Travel planners
5. Here’s what we did
•
Talked to:
o Day traders
o Investors
o Brokers
o Research analysts
o Journalists
•
Built a wireframe:
6. Here’s what we found
•
•
Summary of key hypothesis: Burst the filter bubble and provide a holistic
view of a topic of interest to an investor
Customer feedback/insights:
o Not useful for professional investors → Bloomberg terminal + direct
access to analyst reports
o Individual investors lack the time to learn about a stock in detail and
either invest in well known stocks or mutual funds
o Rely on understanding general trends rather than specifics
o Google and Yahoo finance + other crowdsourced websites are more
than satisfactory for personal investors
o Domain experts refer to research articles
o "... too much technical analysis was of little use and I would rather look
at the total environment using Google searches"
7. RESTART: Here’s why
•
•
•
•
After two and a half weeks, we did not find a definite product-market fit
with either investors or journalists
All of us were excited about the original idea (bursting the filter bubble) but
it had no market
None of us were excited about the “investor” market segment
We had not anticipated that news content would be so closely guarded
by organizations like Bloomberg, Reuters
10. Results from talking to customers
Summary of key hypothesis: Optimize recipes, provide step by step
navigation, adapt to multiple cooks
What we did: Interviewed students from different backgrounds, majors
What we learned:
•
•
•
•
•
Most people who cook frequently consider cooking as a recreational/
therapeutic activity, and were not looking to “optimize”
Most recipes already have step by step navigation, however, they are
missing visual and aromatic cues
Most grad students almost never have all the ingredients for a recipe, so
they use recipes as a reference and improvise
Recipe is unnecessary to most after cooking it twice or thrice
Students only cook the stuff that they know well in order to save time
13. Results from talking to customers
•
•
Summary of key hypothesis: Solve the “manual entry problem” and
provide personalized recommendations
Customer feedback/insights:
o Hard to keep track of everything, end up throwing stuff bought in bulk
o Don’t use inventory apps - too much manual entry
•
Our solution:
o Leverage the fact that people buy the same kind of ingredients
o Have them enter the first couple of times and we learn from that
•
Teaching team: Build an MVP and get tangible feedback
14. There’s a better way
●
●
●
●
●
Provide personalized recommendations from the start
Approximate inventory based on behavior on the website
Solve cold start problem with a very short survey ( Thanks mor.sl and Netflix)
No manual entry
The more you use it the better it becomes
16. Results from talking to customers
Summary of key hypothesis: Focus more on personalized
recommendations rather than inventory management
What we did:
•
•
Built an MVP and showed it to customers for feedback
Ran Google AdWords campaign to measure signup rate
What we learned:
•
•
•
Recipe recommendations based on preferences and available inventory
was a “must-have” or “nice-to-have” for most customers we interviewed
CAC from AdWords is very high - must rely on virality for demand creation
MVP Feedback
o Some users prefer videos over text (provide both)
o Recommendations should not provide a barrage of recipes with same
20. Results from talking to customers
Potential Revenue Stream: Targeted ads and premium version.
What we did:
•
•
Interviewed grad students to gauge willingness to pay
Interviewed recipe website owners to understand their revenue model
What we learned:
•
Large fraction of students not willing to pay, while some others find
our solution appealing enough to pay
o Range from $1 one time fee to $20 annual subscription
o Will pay for content not recommendation.
•
Grad students is a bad market segmentation for targeted ads
o More attractive segmentations: vegetarian, vegan, gluten-free,
diabetic
23. Results from talking to customers
Potential Revenue Stream: Recommendation as a service
What we did: Interviewed recipe websites and online foodstuff delivery
oriented startups
What we learned:
•
High traffic recipe websites are not keen on having a third-party
recommendation service on their site
o Allrecipes.com, Yummly have working recommendation engine
developed in house
o SimplyRecipes.com does not want to modify a working model
•
Food-related services which sell directly to consumer have a greater
incentive to increase user engagement and retention
o But they are not willing to share data with 3rd party services
27. Here’s what we did
•
•
Summary of key hypothesis: Resegment the market to target frequent
cooks with personalized recommendations and novice cooks with amateur
cooking instruction videos
We talked to frequent cooks to gauge their willingness to pay for
personalized recipe recommendations
● We recorded cooking
instruction videos and asked
novice cooks to compare
them with videos by
professional chefs
https://www.youtube.com/watch?v=Jgf-a6Tmmb8
28. Here’s what we found
•
•
•
Frequent cooks willing to pay not more than a one-time charge of $3-$5 for
the convenience of getting relevant recipe recommendations
Novice cooks are interested in tips/hacks related to common cooking
methods for amateur cooks, but did not see any huge value addition in the
amateur videos as compared to professional videos
We posted our cooking video on the Stanford India Association FB group:
No. of members on Facebook Group
474
No. of video views
116
CTR
No. of subscribers
Signup rate
CAC
24.5%
4
3.45%
N/A
29. Revenue Flows (3)
$3-$5 one
time
payment
Expert
Premium Version
Curated
Training
Videos
omnyscope
Free Version
Novice
32. Not a viable business!
•
•
•
High customer acquisition costs
o AdWords are very expensive and signup rate is low
o Not sure whether service will be viral enough to drive demand creation
Shaky revenue model
o Some customers have shown willingness to pay, but majority do not
perceive it as a major value addition over the free services available in
this space
o Advertisements will not generate significant revenue from the target
customer segment
We are not going to pursue this idea further
33. Lessons Learned
•
Customer discovery
o Not just “get out of the building”, also necessary to “get out of the
mindset” by interviewing people who think differently from you
o Keep all boxes of the canvas under consideration during every
interview in every week, don’t just focus on the topic of the week
•
•
Competition
o In an overcrowded space, if you find an unserved customer need, ask
why nobody else has filled the need yet. There might be learnings
about impracticality of solution or non-existence of revenue streams
Teamwork
o Customer discovery is painful, it’s important for everyone to be
motivated
o Choose a product which excites everyone on your team
34. Domain Knowledge Acquired
In addition to lessons learned, we acquired tangible domain knowledge
about many areas, as a by-product of our customer interviews and
numerous pivots:
•
•
•
•
•
•
•
How day traders, personal investors and brokers make their decisions
How financial information flows between research analysts and key
decision makers, via organizations like Reuters, Bloomberg, etc
How journalists manage their research, interviews and writing process
How grad students manage their cooking activities
How high-traffic recipe websites increase engagement, and what are the
current challenges they face (eg. monetizing mobile traffic)
How online delivery services manage logistics issues and partnerships
How startups decide on building technology in-house or paying a 3rd
party service
35. What we’d have done differently
•
•
•
Come prepared
o Doing initial customer discovery before the start of the class would
have saved us from the painful first two weeks
“Blue Ocean” is better than “Red Ocean”
o We went from a product with no market to a product with an
overcrowded market - we wish we’d have chosen a relatively
underserved market
Prioritize search for revenue model early on
o We spent significant time finding a product/market fit with grad
students, only to later realize that they are a difficult customer
segment to monetize
36. Investment Readiness Level: 4
9. Validate metrics that matter
8. Validate left side of canvas
7. High-fidelity MVP
6. Validate right side of canvas
5. Validate product/market fit
4. Low-fidelity MVP
3. Problem/solution validation
2. Mkt size/competitive
analysis
1. Complete first pass canvas
4
38. Week 1: Sentiment based news aggregation
Key Partners
Key Activities
Content
Generators:
Forming new
partnerships to
obtain data
News Sites
Editorials
Personal blog
sites
Building web and
mobile MVP
Key
Resources
IT infrastructure
Intellectual
Property
Cost
Structure
Value
Propositions
360 degree view
on a topic of your
choice
Bursting the filter
bubble
Provides an easy
medium for
collecting a
diverse set of
opinions
MVP:
Specialised
product for one
customer
segment
Customer
Relationships
Customer
Segments
Feedback on
content quality
Professional
users:
Recommendation
s
Investors
Business
analysts
Channels
PR firms
Individuals:
Content
aggregator - web
and mobile app
Browser plugin
Revenue
Streams
Server and bandwidth
Advertisements
Paid APIs for data access
Freemium model
Researchers
Young adults
Travel planners