Tech-Forward - Achieving Business Readiness For Copilot in Microsoft 365
AI-Powered Facial Animation and Product Recommendations
1. Artificial Intelligence
Paving The Way to The Exponential Age
Muhammad Ghifary, Ph.D
Head of Engineering @ GovTech Edu
Adjunct Lecturer @ STEI ITB
2. Overview
Modern Intro to AI
Use Case #1: Performance-Driven
Facial Animation (Movie VFX/CGI)
Use Case #2: Product
Recommendation (E-Commerce)
Use Case #3: BRILINK Agent
Acquisition (Banking)
Recent Trend: Generative AI
4. Artificial Intelligence (AI)
The theory and development of computer systems able to perform tasks that normally
require human intelligence
Insight
Discover patterns
or make
predictions
Automate Internal
processes
Cognitive
Engagement
Product
Process
02
03
01
Source: Deloitte
Interaction
Communicate with humans through digital or analogue
mediums
Decision-making
Generate rules from general data and apply specific
profiles against those rules
Customization
Generate rules from specific profiles and apply general
data to optimize outcomes
Foresight
Determine the probability of future events
Pattern detection
Recognize (ir)regularities in data
Source: World Economic Forum
5. Machine Learning
From Deductive to Empirical Science
Problem
Definition Dataset Construction Data Transform
Model Training &
Evaluation Model Serving
Set the research goal
01
Make a hypothesis
02
Collect the data
03
Build a model and test your
hypothesis
04
Analyze your results
05
Reach a conclusion
06
Refine the hypothesis and
repeat
07
x y
f
known known
unknown
Input (X) Output (Y) Application
email spam? spam filtering
customer info loan approved? credit scoring
face picture is it you? face recognition
product, user info click? feedback? recommender system
6. AI Today
Algorithm(esp. Machine Learning) + Data + Computing Power and Storage
https://www.mckinsey.com/capabilities/quantumblack/our-insights/an-executives-guide-to-ai
7. Superhuman Ability in Game
AlphaGo
Source: https://www.bbc.com/news/technology-35785875
AlphaStar
Source: https://deepmind.com/blog/article/alphastar-mastering-real-time-
strategy-game-starcraft-ii
11. AI Organization and Roles
Data
Engineering
Modeling
Business
Analysis
Deployment
AI
Infrastructure
Data
Scientist
Machine
learning
Engineer
Data
Analyst
Software
Engineer-ML
Machine
learning
Researcher
Software
Engineer
Data Engineering Modeling Deployment Business Analysis AI Infrastructure
Source: Workera Report 2020 - AI Career Pathways: Put Yourself on the Right Track
13. Performance-
based Facial
Animation
• Creating an animation of a realistic and expressive human face is one of the
greatest challenges in CGI
• The human face is an extremely complex biomechanical system that is very
difficult to model
• While some animators may be able to produce realistic facial animation, the
consistent production of large amounts of flawless animation is not practical
• Simply mimicking the desired expression is far faster, easier, and more natural
than adjusting dozens of sliders
14. Facial Deep Learning Solver (FDLS)
A machine learning
approach to solving face
animation weights
Framing the animation
solving as a multi-
dimensional regression
problem
17. Avatar: The Way of Water
[SIGGRAPH Asia 2022] Animatomy: an animator-centric, anatomically inspired system for 3D facial modeling,
animation and transfer
20. Product
Recommendation
A computerized system that suggests
goods and services by predicting user’s
preference and ratings.
Recommendation may be combined with
personalization.
>Rp200B per month additional income
for sellers
>45% of traffics to product detail end up
clicking the recommendation
21. E-commerce
Data
(User Feedback)
User click indicates some
level of interest. Clicks are
abundant. But noisy!
Okay, I’m interested in
buying this stuff!
I definitely like it. I
have paid for the
product!
I love this item!
22. What can we
learn from the
recommendation
carousel regarding
user interactions?
23. Item Quality
Features
1. Product Rating
2. Seller
Feedbacks
3. Revenue
4. etc.
1. Title Similarity
2. Price Ratio
3. Category
4. etc.
Comparison
Features
[L. Evalina et al. ICACSIS 2019] Towards Improving Similar Item
Recommendation for a C2C Marketplace
Enhance Ranking Through
Learning-to-Rank (LTR)
27. BRILink Agent Recommendation
Only 2% agents reach
“Juragan” after being
acquired
Definition & Daily Activity
PAB BRILink Agent
Existing Business Process
Branchless banking
BRILink Agent
Acquire new potential agent
Account Officer
Serve banking transactions & PPOB
Utilize EDC & BRILink Mobile Apps
Manage agent to reach BEP, Jawara, Juragan
Educate EDC & BRILink mobile features
1
AO finds potential existing
customers to be acquired as
BRILink Agents
Current Problem
2
Agents that accept the offer
from AO will be equipped with a
starter kit (EDC, BRILink Mobile
Apps, etc)
3
AO acting as a coach / facilitator
for some agents in day-to-day
business
4
Monitor the productivity and
track the progress towards the
financial target
Directionless
marketing activity
~50% inactive agents
(0 transaction)
Lack of knowledge of
potential agents
Hard to
reach target
No visibility of
marketing activities
50%
28. AI Roles within BRILink Agent Lifecycle
Location scoring as an auxiliary information for the recommendation considering
geospatial data
~29K agents to be upgraded
for the next 6 month
~229K new agents
can potentially be
acquired, with 50%
of them are
expected to achieve
BEP
Acquisition Activation
Upgrade
Location Scoring
~208K can be activated and
potentially reach BEP
Machine
Learning
Rule-Based
Rule-Based
as a features
29. BRILink Agent Acquisition with AI
Identify Potential Agent from Various Features
Number of BRI
Branch & ATM nearby
Point of Interest
(traditional market, harbor, etc.)
Density
(divided into urban, suburban, and rural
areas)
Business Sector
(grocery store, counter, etc.)
Customer Age
Saving Credit
Amount
Saving Debit
Amount
Lending
Outstanding
Lending Payment
Amount
Credit Frequency
Lending Location
Saving Demography
30. Modeling with Light Gradient Boosting Machine (LGMB)
Combining XGBoost with GOSS and EFB
Why LightGBM ?
Gradient Boosting Machine with same
accuracy as conventional GBDT
Highly Time Efficient
LightGBM reduce data size using GOSS and
reduce features size using EFB
Reduce Data Size
GOSS EFB
Sparse Feature
Bundling
eg. result feature of One
Hot Encoding
As features
PEK_WIRASWASTA,
PEK_PNS, etc. are
combined into 1 again
during calculation
Tree Based
Using “tree”
representing the base
model / function
Ensemble
Combination of multiple “weak”
ML models that can generate
“strong” prediction
Gradient Boosting
Minimize errors / residuals of
ensemble models through
gradient-based training
Histogram Data
Determine split with histogram-
based solution to reduce
computing cost without losing too
much accuracy
LightGBM Diagram Process
lighten the
burden of
calculating
GBDT so that
the time is
shorter
Gradient-based One-side
Sampling (GOSS)
Reduce data size by sampling based on
pseudo-residuals (select top instance
then performs random sampling on
the instances with small gradients)
Exclusive Feature Bundling
(EFB)
Bundle sparse features togethers to
reduce number of features
Training Sample
Selection
eg. Eliminate samples
with very small
transactions and
deposits
LGBM will help process 10-20x faster because it
reduces the amount of data and bundle sparse features
additional algorithm
Source: Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., … Liu, T.-Y. (2017). Lightgbm: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30, 3146–3154.
31. Embedding Explainable AI into User Experience
Model
SHAP Explainer
Accurate user recommendation scores are produced by a complex model built on the top of complex
features. Users (PAB) need to be provided by simple and valid explanations to make the predicted scores
more actionable.
Customer data
Success
Probability
Score
SHAPley value
for each feature
Explainable
Feature
Score
Explainability
Feature Description
is mob banking Cust’s m-banking status
Cust’s sector of business
Loc Total Agent Num of agent around Cust
Example of explainable features
Feature Description
std amt debit l3m variance of debit amount
median amt credit l6m median of credit amount
std loc total agen l6m
variance of agent’s number
around cust
Example of complex features
business sector
32. BRILink Agent Acquisition Recommendation
Presenting the shortest route to
visit the top N recommended
agents
Explainable AI is utilized to help
the follow-up process of the
officers
Human-in-the-loop feedback for
better AI model in the next
iteration
1 2 3
Bandung
Denpasar
33. Measuring Impact through A/B Testing
Productivity of Recommended Agents
24%
new agent reach BEP
without bribrain
52%
new agent reach BEP
with bribrain
in Regional Office,
Acquisition Growth
-18%
Acquisition Growth
+6%
#Users of BRILink Agent Recommendation
Period: Jan – April 2022
34. Impact of BRILink AI Aqcuisition
Overall Summary
Periode Data: Januari - April 2022
Acquisition
User access to BRIBrain Activation BEP Upgrade
Num of agents
2,292
Frequency
176,369
Sales Volume
Rp142,755,363,343
Fee based
Rp223,647,045
Num of agents
28,177
Frequency
5,623,227
Sales Volume
Rp5,137,048,241,164
Fee based
Rp7,134,284,132
Num of agents
18,049
Frequency
20,659,632
Sales Volume
Rp34,119,118,377,991
Fee based
Rp29,046,642,147
Sales Volume Generated
~Rp39,4 Trillion
Total Fee Based Generated
~Rp36,4 Billion
37. DALL-E 2: Arts made much easier
“An astronaut playing basketball with cats in space
as a child’s illustration”
“An astronaut riding a horse in a photorealistic style”
https://openai.com/dall-e-2/
43. Basic Principle of Prompting
Principle 1: Write clear and specific instruction
• Use delimiters: ```, “””, ---, <>, XML tags
• Ask for structured output (e.g., HTML, JSON)
• Check whether conditions are satisfied, check assumptions required to do the task
Principle 2: Give the model time to think
• Specify the steps required to complete the task
• Instruct the model to work out its own solution before rushing to a conclusion
Source: https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/
44. 3 Things in AI Thinking
Know how AI works
1
Know the difference
between AI and
human
2
Know how to work
with AI
3
45. Source: How AI can save our humanity (Kai Fu Lee, https://www.youtube.com/watch?v=ajGgd9Ld-Wc )
46. Source: How AI can save our humanity (Kai Fu Lee, https://www.youtube.com/watch?v=ajGgd9Ld-Wc )
47. OECD AI Principles
“Ultimately, AIs will dematerialize, demonetize and democratize all of
these services, dramatically improving, the quality of life for 8 billion
people, pushing us closer towards a world of abundance.”
--
Peter H. Diamandis
Chairman of X Prize Foundation