21. ML
Such description
So much text
Accept
Reject
Moderation queue
Automatic
moderation system
Duplicate
detection
Forbidden
items
Other ML
models
22. ML
Such description
So much text
MP
Automatic
moderation system
Moderation panel
Accept
Reject
Moderators
s3
ES
Duplicate
detection
system
Hashes
Accept
Reject
Moderation queue
23. ML
Such description
So much text
MP
Automatic
moderation system
Moderation panel
Accept
Reject
Moderators
s3
ES
Duplicate
detection
system
Hashes
Accept
Reject
Moderation queue
Index listings & images
24. ML
Such description
So much text
MP
Automatic
moderation system
Moderation panel
Accept
Reject
Moderators
s3
ES
Duplicate
detection
system
Hashes
Accept
Reject
Moderation queue
Detect duplicates
25. ML
Such description
So much text
MP
Automatic
moderation system
Moderation panel
Accept
Reject
Moderators
s3
ES
Duplicate
detection
system
Hashes
Accept
Reject
Moderation queue
Moderate duplicates
26. ML
Such description
So much text
MP
Automatic
moderation system
Moderation panel
Accept
Reject
Moderators
s3
ES
Duplicate
detection
system
Hashes
Accept
Reject
Moderation queue
Collect feedback
29. Plan
● Data Science at OLX
● Way of Working
● MLOps maturity model
● Improvisation
30. A project like this is very complex
We need a team (or multiple teams) to make it work: it’s a joined effort of many
people working together
31. Roles in teams
● Product Manager (PM)
● Engineering Manager (EM)
● Software Engineers
○ Backend Engineers (BE)
○ Data Engineers (DE)
○ ML Engineer (MLE)
○ Site Reliability Engineers (SRE)
○ Frontend Engineers (FE)
○ Mobile Engineers
● Product Analysts (PA)
● Data Scientists (DS)
32. Team A
Team B
Team C
Product
PM
PM
PM
Head of
Product
PA
PA
Head of
Analytics
DS
DS
DS
Manager
Data Tech
EM
EM
EM
Head of
Engineering
BE
DE
BE
FE
BE SRE
FE SRE
FE
Matrix structure
33. Feature teams
● A cross-functional team with experts in different areas
● All work together on one feature/product
● All have the same goal!
● Anyone can work on anything, as long as it helps achieve the goal
PA DS DE BE SRE
EM
PM
34. Goal setting
● OKRs, set quarterly
● Great alignment tool: other teams know what you’re doing
● Whatever team is doing, should be in line with their OKRs
Example:
● O
○ Catch more fraudsters
● KRs
○ Precision of model A improves from 30% to 60% while staying at the same recall level
○ Model B is tested in 5 key markets
35. Plan
● Data Science at OLX
● Way of Working
● MLOps maturity model
● Improvisation
36. MLOps Maturity Levels
● Level 0: No MLOps
● Level 1: DevOps but no MLOps
● Level 2: Automated training
● Level 3: Automated model deployment
● Level 4: Full MLOps automation
https://docs.microsoft.com/en-us/azure/architecture/example-scenario/mlops/mlops-maturity-model
38. Model creation
● Training
○ Laptop
○ AWS Batch
○ AWS Sagemaker
● Experiment tracking
○ Central MLFlow server
● Version controlling
○ Code — always
○ Data — rarely
○ Models — rarely
39. Model release
● Manual release — for PoCs and less mature teams
● Automatic release via CI/CD (gitlab) — for the rest
● Metric-based automated retraining/release — rarely
● No handover to SWE — DS/team own the full cycle
40. Application integration
● Unit and integration tests — always
● Rely on software engineers to integrate to OLX backend
● A/B tests — often
41. Plan
● Data Science at OLX
● Way of Working
● MLOps maturity model
● Improvisation
42. I’m happy to talk more about
● Processes
● Our data platform
● Experimentation
● Model deployment
● And other things!