Data Science Talk
AMOL SAHASRABUDHE
ANKIT JAIN
WHY THIS PRESENTATION
DEMOCRATIZE
BRAINSTORMING
2
• Why and What of Data Science
• Data Science @Roadrunnr
• Multi Drop Logic
• Demand Prediction
• Future Work
Agenda
3
What is Data Science ?
4
Data is Everywhere
5
HOUSE OF CARDS
MAKING SENSE OF DATA
6
HOUSE OF CARDS
WHO USES DATA SCIENCE
7
HOUSE OF CARDS
WHO IS A DATA SCIENTIST
8
HOUSE OF CARDS
WHAT DO THEY DO
9
HOUSE OF CARDS
WHAT DO THEY DO
IDEA EXPERIMENT VALIDATE SPEC DEPLOY
MODELING
SKILLS : Math, Statistics, Programming and Domain Knowledge
10
HOUSE OF CARDS
DATA SCIENCE@ROADRUNNR
11
Use data as a raw input to develop products to:
• Minimize Estimated Time of Arrival (ETA)
• Improve Reliability
at minimal Costs
VALUES
12
Multi Drop Clubbing Logic
• Improve the multi drop grouping logic to:
• Reduce number of touchpoints/order (5-15%)
• Reduce distance travelled per touch point
• Increase driver satisfaction
13
Existing Multi Drop Logic
HUB
Red Cluster can be done away with in this scenario
14
Ideal Clubbing
HUB
DB should be able to cover points on his way
15
How to Achieve Ideal Clubbing?
HUB
D1 < D2
Join two points which have minimum distance between them
• Joining two points
16
Joining Two Groups
• Join on the basis of shortest distances and not center distances to
achieve ideal clubbing
HUB
• S1, S2 shortest distances
• C1, C2 center distances
17
Delhi West Hub Orders (Existing Clubbing)
18
Delhi West Hub Orders (New Clubbing)
19
Algorithm Limitations
• Sub optimal solution due to accuracy vs complexity trade off
• Performance limited by accuracy of Lat,Long of drop points
• Not optimized for size and weight of shipments
• Routing not included as a part of this version (V2)
• Considers only bike as a carrier (V2)
20
Demand Prediction
21
Demand Prediction (Impact)
30% stock-outs during peak hours
70% demand during peak hours
Supply Planning
Driver Placement
Surge Pricing
22
Demand Prediction(Actual Demand)
Koramangala
High Variance in hourly demand
23
Demand Prediction(Method)
24
Demand Prediction(Actual and Predicted)
Koramangala
25
Demand Prediction(drill down)
Koramangala 8PM
Predictability in number of orders in this hour
26
Demand Prediction(Model Failure)
Electronic City
Demand is too erratic to be captured by modeling
Bad predictions
27
Demand Prediction(Future Work)
Category specific demand
Prediction for smaller clusters rather than localities
Prediction over smaller intervals
Stock-out incorporation
28
Data Science Future Projects
Surge pricing
Wait time prediction for food orders (deployment pending)
Carrier Selection (Extension of multi drop logic)
Driver Fraud detection
Product size estimation using image
29
THANK YOU
30

Data Science Projects @ Runnr