3. Logan Kopas
Data Team Lead
A little about me
I started out working as a full stack developer at NC
Smart Call. I went back to school to do my masters
in bioinformatics and deep learning. During my
masters I began working at 7shifts and I’m loving it.
I get to combine everything I’ve learned from
undergrad, working at NC Smart Call, and my
masters, as well as learn new skills and face new
challenges. Additionally, I started a data science
consulting company with Conor Lazarou to help
other companies get started with data and
analytics.
9. Est. July 2019
Data Science Data Analysis
Data Scientist
ML expert
Data Warehouse
Developer
Data storage and
distribution
Data Analyst
Most data science-y analyst
Data Analyst
Tableau expert
Data Team Lead
No special skills
Conor Tolo Will Kiara
Data
Developer
Data Pipelines
Jessie
12. Technology
- Apache beam (python)
- Postgres
- Ansible
- That’s it.
Building a data warehouse
Hosting
- Compute Engine VM
- Persistent Disk Storage
14. Building a data warehouse
Pipelines
Storing data is only partially
useful.
- We need to get the data to our 3rd party tools
so that our teammates can use it
15. Building a data warehouse
Pipelines
Problem
- Intercom wants info on “users” and “companies”
- Salesforce wants info on “leads”,
“opportunities”, “accounts”, and “contacts”
16. Building a data warehouse
Pipelines
Solution
- Push all events into Roughneck
- Perform nightly update through Roughneck as
well
Event A
● User ID
● Company ID
● Event name
Enrich from DB
● Lead ID
● Intercom ID
● Account Owner
● Etc.
●
Salesforce
Intercom
Convert
● Proper fields and
formatting for each
destination
17. Events are a large part of our
incoming data.
- Currently we dump it all into a data swamp
- We also pipe the clean stuff into Amplitude
- We have plans for an event warehouse
Building a data warehouse
Events
18. Event Warehouse Hopes and
Dreams
- Graph database
- Great for questions like
- How many steps do users take from starting
to build a schedule to publishing it?
- What are the most common paths that arrive
on the billing page?
Building a data warehouse
Events
https://snowplowanalytics.com/blog/2018/03/26/building-a-model-for-atomic-event-data-as-a-graph/
session_start
view_home_p
age
publish_
schedule
User X
User Y
User A
for
next next
notify
notify
20. PoP Model - Probability of
Paying
- Tracks company stats as they go through trial
- Score from 0-1
- Allows sales to focus on customers that need an
extra nudge
- Indicator of Marketing lead quality
AI for internal use
PoP Model
21. PoP Model - v1
- Random Forest
- Based on different events and properties
- Integrated POS system
- Input employees birthdays
- Published schedule
- Employee logins
AI for internal use
PoP Model
22. PoP Model - v2
- XGBoost
- Incorporates events at different times in the trial
- Number of employee logins in week 1
- Days to first published schedule
AI for internal use
PoP Model
23. Section heading
Page heading
Other internal AI tools
● Plan suggestion model*
● Upsell suggestion model*
● Customer churn prediction model*
*All under development
Photo by NeONBRAND on Unsplash
27. Employee Engagement Score
● Tracks employee engagement
● Can be an early indicator of churn
or other issues
AI as a product
Intro
Autoscheduler
● Generates a schedule
automatically
○ Based on predicted sales,
optimal labour, previous
employee schedules,
events, etc.
28. AI as a product
Autoscheduler
Autoscheduler recipe
● Dash of Sales Forecasting
● 1 part LabourNet
● 1 part ShiftMaker
● 1 part ShiftFiller
● Rinse glass with dash of sales forecasting
● Blend LabourNet and ShiftMaker together,
strain into glass
● Sprinkle ShiftFiller on top
● Garnish and serve
Photo by Mgg Vitchakorn on Unsplash
34. Autoscheduler
ShiftMaker
Shift
● Creates shifts to meet the
labour demand curve
● Adheres to restaurant rules
such as: min/max shift length,
start/end time, etc.
● Also splits, combines, expands,
trims, and adjusts shifts to
meet requirements
36. Autoscheduler
ShiftFiller
ShiftFiller
● Traditional AI that assigns employees to shifts
○ Heuristically solves the “nurse scheduling problem”
● Adheres to hard rules
○ Only 1 employee per shift, only roles the employee is assigned to,
weekly OT thresholds
● Maximizes score based on soft rules
○ Common working times/days, target hours
● Stochastic hill climbing