AI

A
vs
B

AI

எப்ப� இ�க்க �றீர்கள்? How are you?

Apa Khabar? 你好吗?

A Resource from the AI

Strategy for the Government
 TABLE OF CONTENTS

Chapter 1 – Introduction 3

Chapter 2 – Understanding AI 5

How AI Works 5

How AI Systems are Developed and Deployed 9

Chapter 3 – Applications of AI in the Public Sector 11

9 Common AI Applications in the Public Sector 12

Chapter 4 – How do I Start an AI Project? 19

Defining the Problem 19

Collecting and Assessing Data 25

Choosing a Solution Provider 28

a. Tap on Government Resources 28

b. Procure from the Industry 34

c. Research Collaborations 36

Deployment, Operation and Maintenance of the Model 38

Chapter 5 – Developing AI Capabilities 40

Annex: A Compendium of Use Cases 44

2
Public Sector AI Playbook
 CHAPTER 1

Introduction

The Smart Nation Digital Government Group (SNDGG)

has developed an Artificial Intelligence (AI) Strategy for the
Government that seeks to drive the adoption of AI in the
Government’s core business areas to achieve significant benefits.

While many agencies in the Government have started

to scope AI projects, significant value has yet to be
delivered from the use of AI. Two key gaps identified
What are
in the strategy were (i) the lack of awareness of major
ops-tech
AI use cases and (ii) the lack of ops-tech capabilities to
capabilities?
initiate AI projects by public officers.
Ops-Tech is the way
ops and tech teams
This playbook provides public officers, especially non- work together from
technical officers, a guide on how AI can be adopted early stages of planning
and conceptualisation,
in their area of work and shares a range of AI projects
to co-develop and co-
implemented throughout the public service. This deliver capabilities.
playbook will enable you to:

Understand the basic concepts of AI

Identify opportunities to adopt AI in your agency

Learn how to start an AI project

Leverage central support for AI adoption

3
Public Sector AI Playbook
CHAPTER 1 : I NTRODUCTION

Why AI?
AI represents the next frontier As a general-purpose technology, there are significant
of our Digital Government opportunities to apply AI in the public sector.
ambitions. Advances in
machine learning have now Policy-Making & Planning: AI can be used to support
decision-making through more accurate predictions
brought practical applications
and ground sensing
of AI within reach. Machine
learning enables AI systems Service Delivery: AI can enable more personalised,
to improve their ability to responsive and anticipatory services
perform a particular task by
Internal Operations: AI can optimise processes to drive
learning from data over time.
manpower, cost and time savings

AI Strategy for the Government

The vision is for Government The AI Strategy for the Government has five recommendations:
agencies to use AI to deliver
high impact outcomes in their 1 Identify common AI applications that can be proliferated
core business. To achieve this
vision, we will need to define 2 Productionise AI tools by building centrally
and prioritise high impact AI
use cases.
3 Identify signature AI use cases to demonstrate impact

4 Proliferate AI capabilities by broadening and

deepening technical capabilities
This playbook focuses on
recommendations #1, #3 and
#4. The full AI Strategy for the 5 Strengthening trust in the Government’s use of
Government can be found here. AI by establishing governance

4
Public Sector AI Playbook
 CHAPTER 2

Understanding AI

How AI works

A What is AI
AI is a field referring to the collective
advancements around building
systems that mimic what humans
can do – see, speak, hear, move and
most importantly, learn. This means
that AI systems are able to adapt their
approach to solving a problem based
on new data without the need for
manual reprogramming of the system.

B What AI can and cannot do

Science fiction has created an impression intelligence” and can only perform
that an AI system should understand or a specific set of tasks. For one, AI
learn any intellectual task that a human algorithms cannot distinguish between
being can, and even attain self-awareness cause and effect or draw inferences
and consciousness. This form of AI is outside of the data on which it is trained.
called Artificial General Intelligence (AGI), Indeed, AI algorithms remain a “curve-
which most experts would consider fitting exercise”, drawing best fit lines to
decades or even hundreds of years away successfully classify different data points
from realisation. The truth is, as of now, or to predict the next data point.1
AI algorithms remain far from “true

1
Professor Judea Pearl, winner of the Turing Award (the “Nobel Prize of computing”) has argued that “All the impressive achievements of deep learning amount to just curve fitting.”

5
Public Sector AI Playbook
CHAPTER 2 : U N DERSTAN DI NG AI

Even with these current limits, there are many things AI can do and
do better or faster than a human could.

Here are some things that AI systems CAN do very well:

1 Analyse
Derive new insights from data

A Classification B Regression C Clustering

Identify which category an Estimate the Find patterns/

observation belongs to relationship groupings in the data
between variables

2 Read
Natural Language Processing
Understand how humans communicate using language

3 See
Computer Vision
Interpret and understand digital images and videos

4 Listen
Speech Recognition
Process speech into readable text

6
Public Sector AI Playbook
CHAPTER 2 : U N DERSTAN DI NG AI

C Defining terminology
Let’s try to clarify some terms:

ML in AI ML in Data Science

ML in AI projects is often ML in data science projects

productionised, resulting in uses ML algorithms to perform
AI systems producing data modelling to extract
output B, given A. knowledge and insights.

Machine
Learning Data
AI (ML) Science
Uses algorithms that are not
A set of tools The study of
explicitly programmed but
that mimic extracting
learnt and adapted from
human knowledge
processing data.
intelligence. and insights
from data.
Deep Learning
Uses artificial
neural networks
inspired by the
human brain.

Non-ML AI Non-ML Data Science

Besides ML, AI Besides ML, studying

systems also use the data involves
methods likes other processes and
rule-based systems fields, such as data
and symbolic management,
systems to mimic descriptive analytics
human intelligence. and data visualisation.

7
Public Sector AI Playbook
CHAPTER 2 : U N DERSTAN DI NG AI

Additionally, there are a few different types of Machine Learning methods:

What is it When it’s used How it works Use Cases

The algorithm You know how 1. Label the training data and • Predict the call volume
uses labelled to classify the define the output variable in call centres
data to data and want
2. The algorithm is trained • Understand the factors
learn the to define the
on the data to learn the for driving the sales of
Supervised relationship output of the
connection between input a product
Learning between AI system.
and output
inputs and
outputs. 3. Trained model is applied to
new data

The algorithm You do not 1. Unlabelled data is given • Recommender systems

organises know how to to the algorithm
• Segment customers for
the data by classify the
2. The algorithm infers a targeted marketing
searching data or would
Unsupervised structure to the data
for common like to uncover • Identify topics
Learning 3. The Algorithm identifies from a text
characteristics. patterns.
groups of data exhibiting
similar behaviour

The algorithm To learn a series 1. The algorithm takes an • Optimise the driving
learns to of actions that will action in the environment behaviour of
perform a lead to the best self-driving cars
2. It receives a reward
task and outcomes.
or punishment • Balance the load of
Reinforcement maximise
electricity grids in
Learning rewards by 3. The algorithm optimises
varying demand cycles
trial and error. the best series of actions
for maximising total
rewards available

Deep Learning

What is it: How it works:

Deep Learning is a common machine learning technique The first layer of “neurons” receives and
that uses the three methods above but is often supervised processes the data. Subsequent layers further
learning. Inspired by the human brain, it can be described as process and interpret. The network is trained
interconnected layers of “neurons” that learn progressively by weighting the inputs. For example, If
complex data features at each layer. It is used dominantly the network classifies an image correctly,
for unstructured data (e.g. images), sequences (e.g. audio weights contributing to the correct answer are
recordings or text), or time-series data. increased, while other weights are decreased.

8
Public Sector AI Playbook
CHAPTER 2 : U N DERSTAN DI NG AI

How AI systems are Developed and Deployed

STEP STEP STEP

1 2 3
Data Collection
Model Training Deployment
and Preparation

Algorithm New Data

It’s a tomato!

Output

Raw Data Training Data Trained Model Deployed Model Prediction

TRAINING PHASE INFERENCE PHASE

Training Phase STEP 1 Data Collection and Preparation

The training data is Training Validation Evaluation

collected and undergoes Dataset Dataset Dataset
any necessary processing
The set used to train the model The set used for The set that is only used
(e.g. removing frequent evaluation when the model
of the model — is completely trained
duplicates, formatting). used for finetuning — aiming to simulate the
of the model state of data the model
In addition, for many will face in the real world
projects, the training
data needs to be labelled,
requiring significant
effort. After processing
the data, it is split into
three different sets for
use in an AI project.

Whole Dataset
9
Public Sector AI Playbook
CHAPTER 2 : U N DERSTAN DI NG AI

Training Phase STEP 2 Model Training

ta
n Da
atio Ev
a
d

lu
li
Ru

Va

ate
Run

n
l Final

Ev
de C ua
Mo Model

la
ti o
nD
a Validate ata
Algorithm Run Training D at the Model Evaluate

dj r

s
A Choose B Train
us
t Hy ete D Evaluate
p erp ara m
the Algorithm the Model the Model

Once the data is ready, the A Choosing the Algorithm

next step is then to build and
Engineers have to choose the algorithm best suited for the task.
train the AI model, which
follows a process of:
B Training the Model
The algorithm runs the training data to perform a given task
(e.g. making a prediction), resulting in a trained model.

C Validating the Model

Next, the model needs to be validated against the
Definitions: Validation dataset, which must be different from the
Training dataset. With each run, the hyperparameters of
Algorithm
the AI model are adjusted. After repeating the process
A series of instructions
executed automatically multiple times, the performance of the model improves.
by a computer

Model
D Evaluating the Model
The algorithm which has Finally, once the model has also performed well at the
been trained on data for a validation process, it is tested against the Test dataset.
specific problem
Depending on the result, the model either has to go back
to training with more datasets or is ready for use!

Inference Phase STEP 3 Deployment

When the model has been trained, it is ready to be deployed

and predict new input data based on what it has “learned”. The
model is deployed for use in an agency’s work processes, and its
performance is continually monitored. When performance falls,
for example, due to changes in incoming data distribution, the
model will need to be retrained.
10
Public Sector AI Playbook
 CHAPTER 3

Applications of AI
in the Public Sector

How to identify opportunities to

adopt AI in your agency?
Broadly, there are two ways AI can support you with at work:

Automate repetitive tasks Augment human capabilities

Are there repetitive tasks at work Do you wish you had more insights from data
that you wish could be automated? to back up your decisions? AI can support by
Tasks that AI can automate tend to be providing predictions and detecting patterns
consistent and repetitive. There are not perceivable by humans. There are three
two archetypes of AI applications that archetypes of AI applications that augment
automate repetitive tasks: human capabilities:

• Recognition • Personalisation
• Conversation Systems • Pattern and Anomaly Detection
• Forecasting and Decision Support

11
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

We have identified nine common AI applications that

fall under five archetypes of AI applications:

Common AI Application Use Case

Recognition

Video analytics for Kerbside

Video analytics Loading Bay Pilot (URA)
Automate Repetitive Tasks

Information extraction Extraction of unstructured

and summarisation financial data (DOS)

Conversation systems

Chatbot Municipal Services Chatbot (MSO)

Speech and Transcription of parliamentary

language tools sitting (Parliament)

Personalisation

Customisation of job
Personalised services recommendations (WSG)

Pattern and Anomaly Detection

Augment Human Capabilities

Understanding
?

Sentiment analytics SME issues (ESG)

Enabling holistic
Feedback analytics urban planning (URA)

Detection of Fraudulent
Fraud analytics Claims (SSG)

Forecasting and Decision Support

Identification of riskier
?

Predictive modelling
?
A
B
workplaces (MOM)

12
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

Recognition
Structured vs
In your role, do you regularly look through Unstructured Data
text, audio or video to recognise key Structured data is stored and represented in
information to be extracted or an object of a manner that can be easily interpreted. For
example, tabular data that stores names, height
interest to be identified or counted? and blood-type. Unstructured data is stored
in a format that cannot be directly interpreted.
Such tasks are extremely tedious and prone to For example, an image of a cat or a MP3 file.
human error. Recognition AI models identify and These data have semantic meaning that require
interpretation beyond the raw data stored digitally.
determine objects within text, audio, video, image
or other primarily unstructured data.

Information
Video Extraction and
Analytics Summarisation

What is this? What is this?

An AI application that can identify An AI application that shortens long
objects and monitor their behavior pieces into a coherent summary with
within video footages. main points in a document or extracts
essential details from text documents.

Use case example: Use case example:

Analysing Usage of Kerbside Loading Extracting Unstructured Data for
Bays using Video Footage Business Statistics

Department of Statistics (DOS)

To tap on the vast amount of information
Urban Redevelopment Authority (URA)
that can be derived from unstructured
To efficiently monitor utilisation of
data (e.g. detailing of fixed assets under
the pilot Kerbside Loading Bays,
notes), DOS is developing an AI solution
URA deployed a video analytics
to automatically identify, extract and
model to measure vehicle dwell
validate the required information from
times from video footages.
financial statements.

Find out more here Find out more here

13
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

Conversation Systems
In your role, do you engage It takes an intensive amount of manpower to
citizens and businesses or engage every stakeholder individually and to
transcribe meetings. AI conversation systems
record conversations?
enable machines to comprehend and mimic
how humans interact across mediums such as
text and speech.

Speech and
Chatbot Language
Tools
What is this? What is this?
An AI application that can simulate An AI application that converts
a conversation through messaging audio on different formats to text
applications, websites, mobile apps or or translates text of one language
over the phone. to another.

Use case example:

Streamlining the feedback process for
Use case example:
municipal issues
Improving efficiency and
Municipal Services Office (MSO) accuracy in the transcription of
To bring the feedback reporting process to parliamentary sittings
residents, guide them along the reporting
process and route the case to the correct
agency, MSO tapped on the Virtual Parliament
Intelligent Chat Assistant (VICA) platform Parliament deployed an AI product,
to develop an AI-powered chatbot Transcribe, to reduce manpower and time
available on widely used messaging required as well as increase the accuracy
platforms like WhatsApp and Telegram. of parliamentary sitting transcriptions.

Find out more here Find out more here

Whatsapp Telegram

Give it a
try here

14
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

Personalised
Personalisation Services
In your role, do you provide What is this?
digital services to citizens An AI application that can curate
or businesses? recommendations and customise a
user’s experience on the platform
It is not possible to manually according to their individual needs
customise services for every and preferences.
individual. Personalisation
engines can develop profiles of Use case example:
each individual that can adapt
Customising job recommendations
over time to the individual’s
for individual jobseekers
unique preferences.

Workforce Singapore
To reduce time and effort spent
on finding suitable jobs, WSG
developed an AI model which made
personalised job recommendations
based on a jobseeker’s profile (e.g.
skills), preferences and past activity
on MyCareersFuture.

Find out more here

15
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

Sentiment
Pattern and Analytics
Anomaly Detection
What is this?
In your role, do you need to An AI application that can detect
regularly scan documents to the sentiments and understand
find trends, highlight common the issues of concern reflected in
topics or spot atypical claims? text documents.

On top of being laborious, there is a Use case example:

limit to a human’s ability to detect and Understanding issues facing SMEs
identify patterns and anomalies when
the data is vast. AI algorithms are able Enterprise Singapore (ESG)
to learn patterns in the data and find
To gain deeper insights on
higher-order connections between data
the types of issues faced by
points to see if it fits existing patterns
SMEs when COVID-19 hit in
or if it is an outlier.
2020, ESG developed a model
for detecting emerging issues
based on the engagement
reports of ESG’s SME centres.

Find out more here

16
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

Feedback Fraud
Analytics: Analytics
What is this? What is this?
AI applications that can analyse An AI application that can
feedback and identify trends identify transactions or claims
for planners and policymakers. which might potentially be
fraudulent for further investigations.
Use case example:
Translating varied public feedback
Use case example:
into insights for planning Optimising the detection of
fraudulent grant claims

Urban Redevelopment Authority (URA)

SkillsFuture Singapore (SSG)
To gain a holistic understanding of
public sentiment on planning and To efficiently filter for fraudulent
infrastructure issues, URA is leveraging grant claims, SSG built two anomaly
on AI which can geo-tag email feedback detection AI models which output a
and in the future, create a feedback dashboard to show training providers
dashboard to analyse feedback trends which have displayed high-risk
across time and space, allowing for claims, providing a starting point for
quicker intervention. investigations.

Find out more here Find out more here

17
Public Sector AI Playbook
CHAPTER 3 : APPLICATIONS OF AI I N TH E PU BLIC SECTOR

Predictive A
?

Forecasting and
A
Modelling B
Decision Support
What is this?
In your role, do you need to An AI application which predicts
the probability of an outcome
identify probabilities or predict
using a set of variables.
outcomes based on a given set
of conditions? Use case example:
Increasing inspections to riskier
Whether it is because of a lack of workplaces across sectors
time or analytics support, decision-
making at times might not be backed Ministry of Manpower (MOM)
up with insights learned from data. MOM developed a predictive
AI enables predictive analytics to AI model that identifies riskier
help officers make decisions about workplaces by providing a risk
future outcomes based on existing or ranking of companies based on their
past behaviour patterns. likelihood of safety contraventions,
helping them target and plan
enforcement better.

Find out more here

18
Public Sector AI Playbook
 CHAPTER 4

How do I Start an AI Project?

This chapter provides a guide to help you

in starting an AI project when:

1 Defining the Problem

2 Collecting and Assessing Data

3 Choosing a Solution Provider

4 Deployment, Operations and Maintenance of the Model

STEP

1 2 3 4

Defining the Problem

The first step to any AI project is the problem definition.
AI projects without a proper problem definition process
might end up wasting critical time, resources and effort.

19
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Defining the Problem:

Identify opportunities for AI

There are two ways to find opportunities to adopt AI:

1 2

Identify the pain points in your Examine if existing processes can

work processes which AI can solve be improved using the Common
AI Applications (Chapter 3)

Articulate the Pose a question

Problem Statement Having thought through these three
Rather than a starting point of “My boss aspects, try sharpening the problem
asked me to use AI to improve our work”, statement into a question.
articulate the real pain point in your work
processes. The issue at hand usually E.g.: “Monitoring the video feeds is
burdening for the team”
belongs to one of or a combination of
these three categories: “How can we reduce the number
of hours and manpower needed to
monitor the video feeds?”
Cutting down time and effort
Saving cost
Enhancing performance

20
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Defining the Problem:

Identify how AI can solve the problem

Early in the process, ensure that AI is the right

approach to the problem. Below are factors to assess
if AI is the right tool to solve your problem statement:

1
Data based
There is data that could represent
your decision-making process.

Repetitive
Decision-making is repeated
on a large scale.

Non-deterministic
It is difficult to articulate logical rules on
the decision-making process.

21
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Defining the Problem:

Define success for the project

Identify metrics
Articulate the measure by which the AI solution
is going to improve your work process.

E.g., Hours spent, manpower, risk

identification accuracy

Measure baseline
?
A current benchmark helps the problem to be clearly
defined, and you know what exactly you are trying to
solve or improve. This also helps evaluate the expected
and eventual outcome/impact of the solution later.

E.g., 50 hours/month being spent performing a task

Set a target
Set and quantify the expected outcomes and GOALS
improvements from the project. This is to:

• Set the right expectations

• Measure success and outcomes
after implementation

E.g., Cut hours spent each month by 30%

TIP
Try to think of outcomes and improvement using the frame of

• Cutting down time and effort • Enhancing performance • Saving cost

22
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

To help you better understand this process, let’s look at how ESG did it!

ESG, in partnership with five Trade Associations and Chambers, has a network of SME centres
that provide SMEs with basic business advisory and capability upgrading workshops.
The SME centres also facilitate projects to meet the needs of SMEs.

Articulate the Problem Statement

Identify the issue at hand Problem Statement:
A better grasp of SMEs’ concerns allows ESG to collaborate Manual categorisation of engagement reports
with their partners to better re-orientate the services at the is time consuming and labour intensive. How
SME Centres to meet their needs. Such concerns are typically can ESG increase this process’s efficiency to
reflected in the engagement reports that the SME Centres’ gain a better grasp of SMEs’ concerns and
Business Advisors log when SMEs come forward to seek meet their needs?
business advice. However, sieving and categorising the 3,000
monthly reports is laborious and time consuming.

Identify how AI can solve the problem

Data based Non-deterministic Repetitive

There is data that could represent It is difficult to articulate Decision-making is repeated
your decision-making process. logical rules on the decision- on a large scale.
making process.
Yes, the task is Yes, there are 2,000–3,000
Yes, different engagement
entirely based on the engagement reports to be
reports might use different
engagement reports classified each month
vocabulary to describe the
logged by SME Centres
same problem

Define success for the project

Identify metrics Identify how things would improve after the solution
The time and manpower needed to review is implemented
and classify the engagement reports.
Cutting down time and effort
Measure baseline Rather than the existing process of manually tagging each report
Previously, to understand sentiments, into key areas of concern for business sensing each month, an
the team had to review 2,000–3,000 automated classifier could perform the classification. Officers
engagement reports manually every month. need to only review 5% of the reports every other month and
view a dashboard to review trends and emerging issues.

23
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Support

Getting more support for the first steps of your AI project

DSAID Design Thinking Workshop

The Data Science and Artificial Intelligence DSAID conducts workshops to help agencies
Division (DSAID) was established as a identify AI implementation opportunities and
capability centre to enable GovTech and the upskill organisational understanding of AI. With
whole-of-government (WOG) to formulate customised and proven frameworks, these workshops
effective policies and deliver citizen-centric are essential as a first step for agencies. Users
services through data-driven insights and are brought through the journey of challenging
decision-making. This will allow evidence-based assumptions, redefining problems and collectively
policy-making, enhanced productivity in public discovering solutions that fit their problem
service provision and better-targeted services to statements. The systematic approach guides users to
meet citizens’ needs. their end goal, starting small from proof-of-concepts
to scaling up for impact by leveraging private or
Click here to contact DSAID regarding the
public sector solutions and platforms.
Design Thinking Workshop

AISG AI Discovery
AISG is a national AI programme launched by The AI Discovery programme is based on AISG’s
the National Research Foundation. It brings experience in providing advisory services to
together Singapore-based research institutions businesses to identify, scope and develop AI
to grow local AI research capabilities and solutions. It consists of two 2.5-hour phases.
develop new AI tools and solutions to meet
national and industry needs. AISG offers two
programmes targeted at organisations that are
Ideation Phase:
Helps an agency identify areas of opportunity
taking the first steps in adopting AI solutions.
for their organisations to explore

AI Clinic
The AI Clinic is a theme-based 2.5-hour Prioritisation Phase:
workshop that aims to help participants Helps an agency prioritise use cases based
understand practical AI use cases on feasibility and potential business value
specific to your agency’s sector and
explore potential AI solutions and
While the AI Clinic is a workshop for participants
possible ways to implement them.
from different organisations within one sector, the AI
Discovery workshop is a one-to-one engagement with
AISG for close consultation on AI use case ideation
and prioritisation.

Click here to contact AISG regarding the

two programmes

24
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

STEP

1 2 3 4

Collecting and Assessing Data

Poor training data results in poor AI performance. An AI application is
dependent on the training data and the quality of labels used to develop
the model. As such, the output can only be as good as the input – or as
the phrase goes, “Garbage in, garbage out!”.

If the data is not ready for analysis, is insufficient, or non-existent, the problem is not
ready to be tackled. Therefore, it is important to first know the data needs of the project.

Data Collection Data Labelling

Consider what datasets are Annotate data (e.g. images, text files, videos etc.) with labels for
needed for this project and training the algorithm (for supervised learning). This is a critical
how to obtain these datasets. step, but also often highly manual and time consuming.

TIP

How much data is needed to train a decent model? It depends.

1. Using pre-trained models. 2. Transfer learning. 3. Training from scratch.

Many use cases can utilise Many of use cases can be You may need to collect
pre-trained models, such trained from pre-trained between thousands and tens
as detecting people in model by utilising transfer of thousands of examples if
images, and translating learning. For example, if you you are training your AI model
text. In this scenario, want to train a model to detect from scratch. That said, recent
you will not need to unicorns in images, you can advances in machine learning
collect data for training, utilise model weights learnt have achieved significant
but collect some data from large-scale general object breakthroughs in allowing
to estimate the model detection datasets, which image and language problems
performance. will require less data. In this to achieve reasonable
scenario, you can expect to performance with few than 50
collect thousands of examples examples.
to train the model.

25
Public Sector AI Playbook
 STEP

1 2 3 4

Assessing the state of data

Assess the quality and quantity of the data.

TIP
Refer to the Data
Strategy Playbook

A section of the Data Strategy

Playbook is dedicated to identifying
data needs for use cases. It provides
systematic guidance for sourcing and 3 Principles of Data Quality:
using data for data projects.
1
You can also use this framework for Completeness
further discussion with the Chief Data
Do you have all the required data, both in terms
Officer/IT team to ensure the data
of the variables and the quantity of data?
needs of your project are addressed.

2
Accuracy
How well do the values in the dataset match the
TIP true characteristics of the entities described by
the dataset?
This article on DSAID’s Medium
blog focuses on the steps for data
3
collection and preparation in building
a computer vision model. While your
Veracity
project might not be a computer How credible is the data? Did the data originate
vision model, this could help you get from a reliable source?
a better sense of how to go about data
collection and assessment.

26
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

To better understand where to start, let’s briefly introduce the Government Data Architecture.

Government Data Architecture

The Government Data Architecture (GDA) facilitates efficient and
secure data sharing of clean, authoritative datasets across public
agencies. It does so by designating and building:

Single Sources of Truth (SSOTs) that acquire, clean and maintain high quality core data
Trusted Centres (TCs) that fuse and distribute core datasets

There are 4 Trusted Centres for different types of data:

Individual Business
Department of Statistics Department of Statistics

Geospatial Sensor
Singapore Land Authority Smart National Digital Government Group

Central platforms for data users to request, download and analyse datasets

Vault GeoSpace
For Individual and Business datasets, Vault is the one-stop platform in Geospatial data can be accessed
the GDA which enables officers to search, request and obtain datasets. from https://geospace.gov.sg
To see what datasets are available for the project, you are free to browse
and preview datasets available on Vault without a request. Sensor
Sensor data can be accessed from
Access Vault at https://vault.gov.sg/. https://tcs.gov.sg

Central data
discovery and Data Access & Data Data
Data Acquisition Data Fusion
distribution Distribution Exploitation Destruction
platforms

Data Users

Organisational Trusted Centres (TCs)

SSOT
Structures

Sectoral Data Hubs (SDHs)

Read more about
the GDA here

27
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

STEP

1 2 3 4

Choosing a Solution Provider

There are three options to develop an AI solution for your
agency depending on your project requirements:

A. Government B. Industry
- Data Science and AI Division (DSAID) Support
- Central AI Products C. Research

A
Government Project Consultancy
DSAID Support DSAID also provides technical expertise and solution
architecting services to agencies. By providing technical
DSAID works with government agencies consultancy on software and hardware systems design
using data science and artificial and deployment, DSAID supports agencies in enhancing
intelligence to improve policy outcomes, their systems’ design, capabilities and performance.
service delivery and operational
efficiency. DSAID supports partner
agencies in building in-house expertise, Forward Deployed Teams (FDTs)
formulating data and AI strategies and FDTs are teams of data scientists and engineers
setting up the necessary infrastructure. from GovTech. They are embedded to agencies on
The following paragraphs detail the a permanent basis as their core data science team,
support offered by DSAID to agencies. enabling success by cultivating relationships with agency
IT teams and business owners. Their roles include:
Project Partnerships
DSAID provides support for developing AI • Executing High-Impact Use Cases
solutions through project partnerships with • Developing Workforce Capabilities
agencies. With a focus on shared ownership • Improving Accessibility and Quality of
of success, post-project handover and Data and Tools in the Agency
sustainability, project partnerships involve
Read more about FDTs in this article on
agency staff throughout the project to allow DSAID’s Medium blog.
for stronger buy-in, more effective agile Click here to contact DSAID about a project
delivery and capability development. consultancy or FDTs.

28
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Central AI Products
Some problem statements might be able to be solved using existing
Central AI Products. There are currently eight Central AI Products that
support the development and deployment of AI in the government.
These are products that agencies should consider before other options.

Service DeliveryChat Assistant (VICA)

Virtual Intelligent Percy
Virtual Intelligent Chat Assistant (VICA) Percy
Virtual Intelligent Chat Assistant (VICA) Percy
???
An AI-powered personalisation
Provides a human-like
??? software platform that can
conversational experience. An
Provides a human-like alsoAI-powered
maintain thepersonalisation
data security
??? software platform that can
conversational experience. of
Ancitizens.
AI-powered personalisation
Provides a human-like also maintain the data security
software platform that can
conversational experience. of citizens.
also maintain the data security
of citizens.

Internal Operations
Video Analytics System Transcribe
Video Analytics System Transcribe
Video Analytics System Transcribe
A scalable and modular An AI platform converting
system which empowers speech in audio or video
A scalabletoand
agencies usemodular
Video An AI platform converting
into text.
system which
Analytics as a empowers
service. speech in audio or video
A scalable and modular An AI platform converting
agencies to use Video into text.
system which empowers speech in audio or video
Analytics as a service.
agencies to use Video into text.
Analytics as a service.

SG Translate Fraud Detection

SG Translate Fraud Detection
SG Translate 你好 Fraud Detection
A customised Government 你好 A platform that provides
translation engine using AI to 你好 insights for evaluating and
A platform that provides
A customised
produce Government
translations. investigating grant fraud.
translation engine using AI to insights for evaluating and
A customised Government Hello A platform that provides
investigating grant fraud.
produce translations. insights for evaluating and
translation engine using AI to
produce translations. Hello investigating grant fraud.
Hello

Policymaking and Planning

GovText Analytics.gov
GovText Analytics.gov
GovText Analytics.gov
A platform that provides
A text analytics platform that
centralised access to
streamlines multiple text A platform used
that provides
A text analytics platform that commonly analytics
analytics processes. centralised access to
streamlines multiple text products
A platformforthat
data exploitation.
provides
A text analytics platform that commonly used analytics
analytics processes. centralised access to
streamlines multiple text products for data exploitation.
commonly used analytics
analytics processes.
products for data exploitation.

If you wish to find out more about any of these Central AI Products, contact the Product Manager here

29
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Virtual Intelligent Chat

Percy
Assistant (VICA)
Percy is an AI-powered personalisation
software platform that allows agencies
to easily configure personalised
experiences on their email and web
channels, improving the interaction
between the Government and citizens.
VICA provides a human-like
conversational AI experience for Percy can:
government interaction, leveraging • Enable personalisation of citizen
modern Natural Language Processing engagement on email and web
Engines with higher accuracy and more without the need for technical
know-how
advanced conversational capabilities than
the previous AskJamie chatbots. • Maintain data security of citizens
as compared to many other
VICA can: commercial tools
• Provide answers to basic queries
HDB’s MyNiceHome taps on Percy to
• Perform agency transactions
recommend relevant HDB web articles
for the user
to users based on their browsing history.
• Direct users to the correct agency It further tracks the performance of the
chatbot to attend to them recommendations using click-through rate.
The MyNiceHome team achieved web
• Guide citizens in the process of
personalisation simply by integrating with
giving their input
Whole-of-Government Application Analytics
• Escalate conversations to human (WOGAA) and annotating additional
support agents within the same information on their website. Percy then
chat session automatically built a custom personalisation
model to serve relevant content.
• Support chat interactions on
multiple channels including web,
Whatsapp, and Telegram

This will enable end-users to get

their questions and requests handled
quickly and easily.

MSO’s Chatbot taps on VICA as a

Central AI product. Find out more
about this use case in the Annex.

30
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

Video Analytics System Transcribe

The Video Analytics System (VAS) is

a scalable and modular system that
empowers agencies to use Video
Analytics as a Service, train video
analytics models and exchange
videos and data generated from video
analytics across whole-of-government.

AI models using the VAS can:

• Identify and count objects of
interest from government-deployed
video cameras

• Access on-demand videos that belong

to your agency or other agencies for
operation or research work Transcribe is an AI platform that
enables auto transcription technologies,
URA used the VAS to analyse usage converting speech either in audio or video
of Kerbside Loading Bays via video into text. The platform is accessible from
footages. Refer to the Annex to read both the internet and intranet at https://
more on this use case. www.transcribe.gov.sg via Google Chrome
or Microsoft Edge.

With Transcribe, users can:

• Upload an audio or video file

for transcription

• Speak into the Transcribe web

application for instantaneous
voice transcription

Parliament used Transcribe to

automate the transcription of
parliamentary sittings. Read more
about this use case in the Annex.

31
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

SG Translate Fraud Detection

SG Translate is a customised Fraud Detection is a platform that

Government translation engine provides insights for evaluating and
that uses AI to produce decent investigating grant fraud across WOG.
first-cut translations with localised This will reduce the time needed by
terms in all four official languages. public officers to consolidate data and
insights for investigations and audits
SG Translate can: and provide a common framework for
describing an entity’s nature of breach,
• Serve as a translation resource
which can be used for grant application
for public officers that produces
consistent output, especially processing.
for the translation of distinctly
Singapore content Fraud Detection can:

Over a short span of two years, • Automate checks against

public officers have used SG government data sources such
as CPF and MOM and flag
Translate to generate about 200,000
discrepancies in employment size
first-cut translations. During the
national COVID-19 response, SG • Alert if the applicant/vendor or their
Translate facilitated Government directors/shareholders are blacklisted
communications with the public, or have incarceration records
reducing the time taken to translate
• Draw data from agencies to create a
each Gov.sg WhatsApp message by 25- network of people and companies and
50% as translation requests soared. the relationships between them. It
will flag any conflicts of interest in the
applications for grants requiring that
எப்ப� Apa Khabar?
applicants and vendors not be related
இ�க்க �றீர்கள்?

你好吗?
How
are you?

32
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

GovText
GovText is a text analytics platform From a collection of
that streamlines multiple text documents (e.g. feedback,
analytics processes such as topic survey, discussion notes),
modelling, enabling public officers GovText can be used to:
to analyse and sieve out insights
from texts. When there is a need • Discover key topics in a
to sieve through a high volume of collection of documents
text data, many hours are required.
• Extract the insights
This product potentially reduces
the time and manpower needed to • Reveal any unknown trends
extract information from texts. from the text data

Analytics.gov
Analytics.gov is a platform that On Analytics.gov, officers can:
provides centralised access
to commonly used analytics • Create data science notebooks
products for data exploitation. • Use data visualisation
This will reduce ground collaboration platforms
challenges public officers
• Access curated Open Source
face in the procurement, package repositories
deployment and maintenance (Python and R)
of data analytics tools, such as • Leverage on integration
an AI model. with Vault

33
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

B Procure from the Industry

If DSAID support and the Central AI products are

not suitable, your agency can procure an AI solution
commercially via the “Green Lane Approach”:

Green Lane Approach

The Green Lane is for government agencies
to procure technology products and services
from firms that have qualified under the
various programmes endorsed by GovTech,
such as the Accreditation@IMDA.

This approach aims to:

• Improve process efficiency during the tender $$$
and quotation evaluation

• Support growth and scaling of Singapore-

based tech product companies and start-ups

Any procurement of technology products and

services, if not covered under an international
agreement, will need to follow the Green Lane
Procurement approach. The flowchart on the
next page guides you through the process.

34
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

AC Panel
Agencies are obliged to consider the Panel of Find the list of AI solution providers
Accredited Companies (AC Panel) first before in the AC Panel here.
any other frameworks or companies. AC Panel
consists of Singapore-based frontier tech
product companies and promising start-ups
that have been evaluated under the IMDA SG:D
Accreditation and SG:D Spark programmes.

To expedite the procurement decision

cycle, agencies can discuss specific project
requirements with any AC Panel supplier to
assess whether the AI solution meets your needs.

DSAI Bulk Tender

GovTech established GVT(T)20005 Data Science • Cat A: Software and Professional/
& Artificial Intelligence (DSAI) software and Support Services
services block and bulk tender that will reduce • Cat B System Integration and
Maintenance Services
lead time and effort required from Agencies
• Cat C: Project and
when procuring DSAI products/services.
Consultancy Services

This bulk tender is valid for a period of three For more information, click here.
years, from 3 May 2021 to 2 May 2024, for the
supply of three categories:

Other Sourcing Methods

If neither the AC Panel nor the Bulk
Tender offering are suitable, then
agencies are no longer obliged under the
Green Lane Procurement approach and
can consider other sourcing methods.

35
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

C Research Collaborations

If commercially available solutions are unable or unfeasible to solve your

problem statement (especially if it is more complex and difficult), the third
option for developing an AI solution for your agency is via collaboration
with Singapore’s ecosystem of researchers. There are a few possibilities:

AI Singapore (AISG)
Your agency can collaborate with AISG on IRAS’ AI-enabled Evaluation of
your AI project via its flagship programme, Live Chat Service Quality
the 100 Experiments (100E).
The Inland Revenue Authority of Singapore
(IRAS) handles an increasing volume of
In this programme, AISG supports projects
live chats by taxpayers every year – about
where a minimum viable model can be
235k live chat requests were received in
developed and deployed within 9 to 18
FY2020. Frontline supervisors monitor the
months. AISG will provide matching funds
quality of service rendered by inspecting
of up to $250,000 per 100E project for
live chats regularly through sampling. The
researchers from Singapore’s autonomous service quality monitoring process needs
universities, A*STAR research institutes, other to ensure consistency and quality in the
Singapore-based publicly-funded research delivery of services and can be tedious and
institutions or the AISG engineering team to resource-intensive.
develop an AI minimum viable model to work
for your agency’s problem statement. Your Through the 100E collaboration, AI models
agency will be required to match this funding were deployed to automate the scoring of
in kind and cash. chat dialogues based on specific service
quality metrics. This enabled IRAS to
The fine-tuned AI minimum viable model will efficiently and effectively evaluate the
be handed to your agency to proceed with performance of each agent by aggregating
the integration, deployment and maintenance the predicted scores. IRAS has also
benefited from the collaboration with
of the end solution undertaken by GovTech
AISG by hiring an AI apprentice (who had
or in partnership with a commercial solution
developed IRAS’ AI models) to continue
provider. Your agency can also consider
their efforts in improving, maintaining and
recruiting the AI apprentices from AISG who
scaling up the AI models.
have been working on the project to build up
internal capability in integrating, deploying
and maintaining the fine-tuned model.
Contact AISG here for more information on the 100E

36
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

TRANS Grant Intelligent Case Retrieval

System for Singapore Judiciary

The Judiciary needed an intelligent case

retrieval system for supporting effective
retrieval of key legal documents in
The Public Sector Translational R&D decided cases. Commercially available
Grant (TRANS Grant) supports research tools only enabled searches
translation of R&D into applications and based on the hard text of published
taps on the local research community judgements without any intelligent
to solve public sector challenges, capabilities. Legal commercial AI vendors
demonstrate the feasibility of new ideas were also not willing to explore R&D in
and encourage agencies to experiment AI solutions that fit the requirements of
the Judiciary without a hefty price tag.
and deploy innovative solutions using
digital technology.
Therefore, the Supreme Court and State
Courts worked with NTU SPIRIT Lab to
The desired outcome is for the public
develop the Intelligent Case Retrieval
sector to deploy the solutions upon
System (ICRS), an AI solution that can
successfully completing the proof-of- understand case details, identify points
concept and good results demonstrated of law, relevant precedents and possible
during testing. Your agency may outcomes, to guide the judge. From a
approach the following TRANS Labs to Proof-of-Concept, the ICRS has been
discuss possible collaborations: deployed as a full-fledged AI decision
and research support system internally
for judges and judicial officers across the
• Smart Nation Applied R&D Lab Judiciary to replace the hardcopy of the
(SNAL) from A*STAR legal bluebook.

• Smart Platform Infrastructure

Research on Integrative
Technology (SPIRIT) from NTU Contact GovTech here for more information on the
TRANS Grant

37
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

STEP

1 2 3 4

Deployment, Operation and

Maintenance of the Model
Deployment and Operation
Successfully implementing an AI model to solve your problem
statement goes beyond developing an AI model. Crucially, there are
important considerations, which can be categorized in the classic
framework of People, Process and Technology.

Technology
Depending on the needs of your project, there are some important technical
considerations which you will have to discuss with the engineers.

• Data Pipelines • Deployment environment

While the AI model could be trained by batch Should you deploy your AI model
data that has been collected, data pipelines are on-premise or on the Government
needed to input data into the AI system. It is Commercial Cloud? This would depend
also important to specify the destination of the on the control needed of the data, and the
AI model’s output. Will the output be directly compute resources your model requires.
used by a human to make decisions? Or will the
output be used as input for another application?

Latency Requirement
• Regularity of output
Latency requirement describes the speed
Is the model required to output upon at which an application or user requires
request, or only at specified timings? How an AI model’s output.
quickly does your agency require the model
predictions? This will affect the model’s
latency requirement.

38
Public Sector AI Playbook
CHAPTER 4 : HOW DO I START AN AI PROJ ECT?

People Maintenance
An AI model’s performance might
deteriorate over time as data evolve
and models built become outdated.
It is important to track a model’s
performance and maintain it. To
ensure a systemised maintenance,
you should think about:

Tracking performance
How will you track the model’s performance?
If intended users do not know or Use historical data to backtest, examine data
want to use the AI solution, the distributions and test assumptions.
benefits reaped will be limited no
matter the quality of the technology.
Retraining the model
Some ways to encourage adoption How often should you retrain the model? If
of the model amongst users in the you receive new data periodically, retraining the
agency include:
model can be scheduled accordingly. Will you
retrain the entire model on new data? This might
• A clear communication plan on the
capabilities and limitations of the not be necessary, depending on the model.
AI system
• Addressing user resistance Infrastructure for retraining
and demands
Does the project have the infrastructure for
• Training users users to benefit from retraining? This could be in “containers”
the AI system
– isolated applications offline, or online,
allowing the model to learn continuously.
Process
It is important that the AI solution
is integrated into the existing work
TIP
processes. The right processes
should be in place to support the
If you need more help on deploying, operating
adoption of the new AI solution. and maintaining the model, you can reach
out to DSAID for their Project Consultancy
services mentioned earlier in this Chapter.

39
Public Sector AI Playbook
 CHAPTER 5

Developing
AI Capabilities

With AI increasingly used across the

Government and its functioning markedly
different from technology most are familiar
with, there is a need to proliferate AI
capabilities across an agency’s workforce
and further develop ops-tech capabilities to
harness AI effectively.

Crucially, most officers need to be literate with

AI technology while more need to be able to
identify opportunities where AI can be employed
and even customise AI products to suit the
needs of their agency.

The next page provides an overview of the

different archetypes of officers with regard to
AI competency. Linked to each archetype is a
directory of relevant courses which officers can
attend to achieve various competencies. Kindly
refer to the course websites for the latest details.

40
Public Sector AI Playbook
CHAPTER 5 : DEVELOPI NG AI CAPABI LITI ES

Literacy Business User Power User

Archetype
Objective

Knowledge and Ops-tech capabilities Capabilities to

understanding of to identify AI use cases customise AI systems
AI fundamentals and products for use
and applications

• Demonstrate • Relate some use • Perform basic data

intuition (not cases where modelling for a set
technical machine learning of data to identify
comprehension) on (e.g. clustering, and recognise
the workings of some regression, patterns and predict
common machine classification) has future trends
learning algorithms. been used to solve using techniques
Competencies

specific problems. in classification,

regression, clustering,
forecasting.

• Utilise proven
off-the-shelf products,
auto-ML tools,
open-source tools and
libraries for analysis.

• Develop appropriate
evaluation criteria
to determine the
performance of
the models and
recommend further
improvements.

41
Public Sector AI Playbook
CHAPTER 5 : DEVELOPI NG AI CAPABI LITI ES

Type of Courses Competencies

Data Science and AI Literacy ePrimer (LEARN App)

AI for Everyone (Coursera/Andrew Ng)
AI for Everyone Workshop (AISG)
Artificial Intelligence for Everyone - A Practical Experience (Republic Polytechnic)

Literacy

Introduction to AI
An Introduction to Code-Free Machine Learning (Republic Polytechnic)
Introduction to AI and Machine Learning (Singapore Polytechnic)
Getting Started with Artificial Intelligence (IBM)
Introduction to Machine Learning Techniques (code-free) (CSC)

Business Users Specific Applications

Predictive Analytics
Predictive Analytics for Business Users (Temasek Polytechnic)

Text Analytics
Text Mining – Making Sense of Unstructured Data (Republic Polytechnic)
Text Analytics for Business Users (Temasek Polytechnic)

Conversation Systems
Developing Your First Chatbot (Republic Polytechnic)
Creating Your Digital Voice Assistant (Republic Polytechnic)

Introduction to AI
Artificial Intelligence for Techies – A Hands-On Approach (Republic Polytechnic)
Data Science and Machine Learning Bootcamp with R (udemy on CSC LEARN)
AI for Industry (AISG)
Certificate in Artificial Intelligence (SMU)

Power Users Specific Applications

Predictive Analytics
Predictive Analytics with Programming (Temasek Polytechnic)

Text Analytics
A Primer in Text Analytics – Unstructured Text and Sentiment Analysis (CSC)
Text Analytics with Programming (Temasek Polytechnic)

Video Analytics
Face Recognition in Action (Nanyang Polytechnic)

For an updated and comprehensive list of courses you can take,

check out the Data Science and AI Training Road Map

42
Public Sector AI Playbook
CHAPTER 5 : DEVELOPI NG AI CAPABI LITI ES

Featured Course:

An Introduction to Code-Free
Machine Learning (Republic Polytechnic)
The National AI Office and DSAID have worked with Republic
Polytechnic to adapt this course for public sector officers, with
hands-on exercises on the Government’s Central AI Products

Overview of Machine Learning

Machine Learning workflow

Five Hands-on exercises using GovTech and private sector platforms

• Solve a regression problem on a cloud provider’s platform

• Train an object detector on GovTech’s Video Analytics System

• Perform topic modelling on GovTech’s GovText platform

• Perform sentiment analysis on a cloud provider’s platform

• Use speech-to-text technologies using GovTech’s Transcribe platform

Instil awareness of major AI applications

from use cases in the Government

Develop capability for identifying opportunities to

apply AI at work and for scoping out an AI project

43
Public Sector AI Playbook
 ANNEX

Video Analytics
Analysing Usage of Kerbside Loading Bays (KLBs)
Using Video Footage (URA)

Problem Statement: What the team did:

Despite a high volume of deliveries, Through collaboration with GovTech, a Video
delivery drivers faced challenges Analytics model named “Platesnap” was
in commercial areas that lacked developed to extract vehicle and dwell time
designated loading/unloading facilities. data from video footages captured from KLBs.
They were parking illegally, resulting Using Platesnap, which is currently available
in traffic disamenities for other road on GovTech’s Video Analytics System (VAS),
users and risking summons. The pilot the team efficiently monitored and analysed
involved the conversion of selected dwell times within the KLBs. For example, by
kerbside parking lots into paid Kerbside analysing the footage, it could be shown that
Loading Bays (KLBs). There was a need median dwell time was less than 8 minutes,
to monitor the usage of these bays to and 84% of goods vehicles dwelled less than
evaluate the pilot for its effectiveness 20 minutes. Additionally, analysis also found
and collate learning points for scaling that four of the five KLBs had utilisation rates
up and future deployment. that exceeded 50% of the utilisation rates of
regular parking lots along the same street,
between the hours 9am–530pm, indicating
that the KLBs were well-utilised by delivery
drivers.
Deliveries Only

Outcome/Impact:
Platesnap extracted meaningful information
from video data automatically and quickly,
allowing analysis to be done efficiently without
manually looking through hours of video
footage. This is cost-efficient significantly
reducing the man-hours, therefore allowing
for a long-term, large-scale evaluation of KLBs.

44
Public Sector AI Playbook
AN N EX

Information extraction
and summarisation
Extracting Unstructured Data for
Business Statistics (DOS)

Problem Statement: What the team did:

While DOS has been relying on The team worked with a commercial
structured data from financial AI solution provider to develop an AI
accounts, there is also a rich source of solution, using advanced semantic and
information and new insights that can reasoning algorithms to automatically
be derived from the unstructured data identify, extract, cleanse and validate
(e.g. details of fixed assets under the the required information from financial
notes). However, considerable manual statements. As a Proof-of-Concept
effort is currently required to extract (PoC), the team identified about 30 data
and process relevant information from items such as details of fixed assets,
the unstructured part of the financial overseas subsidiaries and ultimate
statements. This limits the number investors from the financial statements
of financial statements that can be for the model to extract. The solution
processed for data compilation. was assessed to successfully extract the
required data from the unstructured
information in the financial statements
with reasonable accuracy. The experience
and knowledge gained in the PoC helped
the team to plan and scale up the actual
implementation of the AI solution.

Outcome/Impact:
The AI solution enables DOS to
improve operational processes
and data quality and facilitate
the development of more timely
business indicators.

45
Public Sector AI Playbook
AN N EX

Chatbot
Streamlining the feedback process
for municipal issues (MSO)

Problem Statement: What the team did:

Residents who encountered municipal MSO has developed an AI-
issues have been consistently deterred from powered chatbot hosted on the
reporting them because they do not know VICA platform, which simplifies
which government agency to approach or do the feedback provision process
not have the time. This results in municipal for residents. It is available on
issues not being surfaced in time for proper widely-used messaging platforms
resolution. Consequently, this impacts the (i.e. WhatsApp and Telegram),
residents’ standard of living and erodes their and utilises AI predictive
confidence in the Government. capabilities to guide users along
the feedback reporting process,
For those who do report them, crucial prompt residents for mandatory
information is often left out, resulting in case details in real-time, and
additional time spent on agencies seeking predict and route the case to the
clarification before they can adequately act correct agency.
on the case.

Outcome/Impact:
With the AI Chatbot, residents can report
municipal issues on existing messaging
platforms they are familiar with, not
worry about which agency to contact and
whether or not all required information has
been provided. Not only will this improve
residents’ feedback reporting experience,
but it also speed up resolution.

Whatsapp Telegram

46
Public Sector AI Playbook
AN N EX

Speech and
Language Tools
Improving efficiency and accuracy in transcription
of parliamentary sittings (Parliament)

Problem Statement: What the team did:

Transcribing hours of Parliamentary Parliament worked with GovTech to
sittings to compile a full verbatim customise the transcription Central AI
transcript quickly took hours of Product, Transcribe, to integrate the
work and engagement of additional solution into the workflow process. The
manpower on sitting day to help type customised solution functioned within
out raw transcripts. Parliament wanted the Government intranet environment to
to explore how technology or AI could ensure compliance with cybersecurity IMs,
be adopted to reduce reliance on contextualised to Parliament’s lingo, and
human effort and improve turnaround included an optional spell check feature to
time for the raw transcripts. increase the spelling accuracy in the raw
transcripts. In April 2021, the customised
Outcome/Impact: solution was further enhanced with a live
Parliament was able to reduce manpower transcription feature that could machine-
every sitting day and turn around more transcribe spoken text instantaneously,
accurate, 15-minute-long transcript within following feedback from users.
1.5 hrs, down from 2.5 hrs previously.

Transcript

47
Public Sector AI Playbook
AN N EX

Personalisation Services
Customising job recommendations
for individual jobseekers (WSG)

Problem Statement: What the team did:

While people no longer have to flip JumpStart is a data science platform
through the newspapers to find job that focuses on providing AI services
ads, today’s online job portals still have for labour market use cases, such as job
their drawbacks. Jobseekers are often or course recommendations. GovTech
overwhelmed with the number of options developed a platform, JumpStart, that
out there, and it takes significant time and tapped on various data sources like skills,
effort to sort through the listings to find CVs, user clickstreams, applications and
relevant jobs. Not only might jobseekers job descriptions to build models that
accidentally miss out on potentially made personalised job recommendations
suitable jobs, this time and effort might based on a jobseeker’s skills and
also be better used to help prepare preferences on MCF.
their applications. With over 50,000
job listings, MyCareersFuture (MCF)
needed a recommendation system to Outcome/Impact:
help jobseekers find suitable job matches.
After implementing these new models
Being able to deliver personalised and
on MCF, job applications on the platform
timely job recommendations would
increased by 9.1% and successful job
help to improve both the process and
placements using MCF increased by 21.5%.
outcomes of job matching.

48
Public Sector AI Playbook
AN N EX

Sentiment Analytics
Understanding issues facing SMEs (ESG)

Problem Statement:

ESG has a network of SME centres that provides SMEs with

basic business advisory, capability upgrading workshops and
facilitation of projects to meet the needs of SMEs. When
COVID-19 hit in 2020, there was a need for deeper insights on
the types of issues faced by SMEs and to better understand if
new issues were also surfacing for micro and small enterprises.

What the team did: Outcome/

Impact:
The SME Centres’ Business Advisors typically log engagement
reports with SMEs who come forward to seek business advice. Previously, to understand
The team overseeing the SME Centres collaborated with sentiments, the team
DSAID’s forward-deployed team in ESG and developed reviewed around 3,000
a deep learning classifier to mine this data and produce engagement reports and
regular insights. categorised the issues
manually. The ESG
Furthermore, the team developed a model for detecting division now only reviews
emerging SME issues. When tested against historical data, the under 5% of the data
model proved it could quickly pick up some COVID-19 related every other month with
concerns from SMEs, such as concerns over employees having the model, saving the ESG
to transition from on-site to working from home. The model division about 30 hours
also showed that overseas expansion was still a relevant issue each month.
among SMEs who consulted SME centres, even during the
Circuit Breaker period.

49
Public Sector AI Playbook
AN N EX

Feedback Analytics
Translating varied public feedback
into insights for planning (URA)

Problem Statement: What the team did:

Citizen feedback on planning and infrastructure URA had previously collaborated with
issues are sent to varied agencies – e.g. LTA for Institute for Infocomm Research,
road congestion, NParks for the lack of parks within A*STAR to develop an AI model which
an area, or PUB for ponding and drainage issues. geo-tags email feedback to various
However, feedback from various sources need to be locations. URA next plans to use
looked at holistically to gain a complete understanding machine learning and visual analytics
of public sentiment, especially for cross-cutting issues. to translate the varied types of public
This enables better public engagement preparation, feedback across agencies in the sector
and a coherent set of measures for each agency’s into insights of topics of concern, across
respective policies and planning. different scales, from islandwide down
to a neighbourhood level.

Outcome/Impact:
Where feedback was previously manually geo-
tagged and thus at times inaccurate or absent, the
model now automatically geotags, improving the
work process. Separately, once developed, planners,
architects and policymakers in various agencies
will have a feedback dashboard that can analyse
trends across time and space. This will enable
them to identify emerging topics, enabling them to
intervene quickly and be less reactive. Furthermore,
this tightens the feedback loop between ground
sentiment and policy/plan formulation.

50
Public Sector AI Playbook
AN N EX

Fraud Analytics
Optimizing the detection of
fraudulent grant claims (SSG)

Problem Statement:
SkillsFuture Singapore receives more than
600,000 grant claims per year. The agency
needs an efficient way to filter for fraudulent
grant claims, such as claims for training that
did not occur.

What the team did:

SSG worked with GovTech to build two
anomaly detection unsupervised machine
learning algorithm to better detect novel
types of fraudulent behaviour. Business
users proposed and validated the features
which the algorithm was trained to
recognise. The product was an AI model
that outputs a dashboard showing the
training providers with the highest anomaly
scores. The key features that contributed
to the score are highlighted, providing a
starting point for investigations.

Outcome/Impact:
Backtesting the model on historical
data, the algorithm would have enabled
early detection of a known case five
months earlier, when less than 1% of
the claims have been disbursed.

51
Public Sector AI Playbook
AN N EX

Predictive Modelling
Increasing inspections to riskier
workplaces across sectors (MOM)

Problem Statement:
The Occupational Safety and Health
Division in MOM carries out enforcements
and inspections to ensure that workplaces
maintain acceptable safety and health
standards. Previously, these inspections
were planned primarily on the general risks
of each sector. For instance, a majority of
construction worksites are inspected at
least once a year. However, more frequent
inspections should be planned for riskier
workplaces regardless of sector to achieve
accident prevention outcomes and optimise
resources. There was a need to identify
riskier workplaces to target and plan
enforcement better.

What the team did: Outcome/Impact:

The team first hypothesised potential With the same inspection resources,
company or workplace parameters that the team was better able to focus on
correlated to the infringement or accidents workplaces with severe contraventions
and analysed their relevancy. Thereafter, the and allowed for differentiated
relevant data was used to develop predictive interventions based on the risk score of
AI models for various sectors, providing a the companies, such as more targeted
risk ranking of companies based on their inspections on companies with higher
likelihood of safety contraventions. risks. In particular, for the manufacturing
sector, the model significantly improved
the hit rate of the workplaces inspected,
thus targeting where it matters and
saving resources.

52
Public Sector AI Playbook
Brought to you by: