Session 1
đThis first session will cover an introduction to Generative AI & harnessing the power of large language models. The following topics will be discussed:
Introduction to Generative AI & harnessing the power of large language models.
Whatâs generative AI & whatâs LLM.
How are we using it in our document understanding & communication mining models?
How to develop a trustworthy and unbiased AI model using LLM & GenAI.
Personal Intelligent Assistant
Speakers:
đGeorge Roth - AI Evangelist at UiPath
đSharon Palawandram - Senior Machine Learning Consultant @ Ashling Partners & UiPath MVP
đRussel Alfeche - Technology Leader RPA @qBotica & UiPath MVP
Developer Data Modeling Mistakes: From Postgres to NoSQL
Â
AI and ML Series - Introduction to Generative AI and LLMs - Session 1
1. Executive Roundtable
Mumbai | Delhi | Bangalore | Jakarta | Kuala Lumpur | Seoul | Tokyo | Brisbane | Singapore | Melbourne | Sydney | Manila | Hong Kong | Taipei | Bangkok
George Roth
AI Evangelist
Introduction to Generative AI &
harnessing the power of large
language models
2. 2
Agenda
What is generative AI and what are the LLMs ?
01
02
03
04
How to develop a trustworthy and unbiased AI model using LLM
& GenAI ?
How are we using it in DU and Communications Mining ?
Personal Intelligent Assistants
4. 4
LLM
⢠Large Language
Models are a type of
AI system that works
with language.
⢠In the simplest of terms,
LLMs are next-word
prediction engines.
⢠Examples:
OpenAIâs GPT-4
Googleâs PaLM
Metaâs LLaMA
Hugging Face -
Bloom
Foundational Models
âLLMsâ specifically refers
to language-focused
systems, while
âfoundation modelâ is
attempting to stake out a
broader function-based
concept, which could
stretch to accommodate
new types of systems in
the future. (Stanford
University)
AI Driven Chat Bots
⢠UX for LLMs
⢠Chat GPT stands for
chatbot generative pre-
trained transformer
⢠They have LLMs behind
them
⢠Use prompts for
conversation
⢠Examples:
- Open AI Chat GPT
- Google BARD (multi
modal)
Fine Tunning
⢠To use LLMs you need
to fine tuning and
distillation
⢠Fine Tuning
Examples:
- Reinforcement
Learning with Human
Feedback (Open AI)
- Active Learning
(UiPath)
Generative AI â system that generates content
5. 5
David Barber
Director at UCL Center for AI;
Distinguished Software Engineer â UiPath
âOne of Arthur C. Clarkeâs famous
âthree lawsâ is: âAny sufficiently
advanced technology is
indistinguishable from magicââŚ.
To many people, generative AI systems
like ChatGPT and GPT-4 can seem
like magic...â
7. 7
1. First technology in the human history that can make decisions and is creative ( Yuval
Harari)
2. Nobody really understands how makes the decisions â is âopaqueâ like the brain (David
Barber â UCL and UiPath)
3. It âhallucinatesâ â makes up stuff if doesnât jknow the answer to a question
4. It âthinksâ differently that the human does â uses numbers not words
5. It can store âtons of info - almost all available info
6. Cannot make the distinction between âgood and badâ (Daniel Dines - UiPath)
7. It fundamentally changes the âwhite collar workâ
8. Allows âmass customizationâ â(Daniel Susskind â Oxford University)
Why Generative AI is so disruptive ?
8. 8
1. Chat GPT cost was 100 K per day in December 2022 â 3 Million per month
2. A fine tuning of a model and a query can be very expensive â it can take days using very
expensive resources. Example: When a user submits a prompt to GPT-3, it must access all
175 billion of its parameters to deliver an answer.
3. Aside of cost, there is an âenvironmentâ: cost, inferences use a lot of computer power that
pollute the environment
4. Solution â use âsmaller modelsâ , fine tunned for specific needs
5. Use hybrid solutions that use the Generative LLMs only when needed, combine these with
traditional searches
The cost of using Generative AI can be high !
9. 9
Generative AI: Consumer vs. Enterprise
Consumer
Generative AI
Enterprise
Generative AI
Primary Use Cases Fun content, games, creative writing Contact center, doc processing, etc.
Accuracy Needs Low (inaccuracies/hallucinations are OK) Very High
Training Data General internet data
Trained and adapted with domain and
company specific
Security and Privacy Shared deployment or end point is OK
Must be deployed within enterprise
environment
Open Standards Not relevant or necessary
Flexibility to choose best-in-class and
avoid vendor lock-in is required
10. 10
Trustworthy & Responsible AI
Trusted
Generative AI
Veracity and
Validation
Content
Moderation
Prompt
Engineering
Intellectual
Property
Human
Oversight
Source: Deloitte AI Institute and MIT, 2020
12. 12
The UiPath Business Automation Platform
PROCESS MINING TASK MINING COMMUNICATIONS MINING IDEA CAPTURE & MANAGEMENT
LOW-CODE DEVELOPMENTâ UI & API AUTOMATIONâ PROCESS ORCHESTRATIONâ
INTELLIGENT
DOCUMENT PROCESSINGâ
INTEGRATED NLP & AI/ML
Discover
Continuously uncover opportunities for process and task
improvements âhelping you identify the highest ROI areas
Automate
Get more done with a digital workforce that seamlessly collaborates with your
people âand automates work via UI and API, powered with native integrated AIâ
Your Applications Your
People
Your
Processes Systems of record â ERP, CRM, HCM | Communications and Collaboration | Personal Productivityâ
ANALYTICS CONTINUOUS TESTING
UNIFIED MANAGEMENT
& GOVERNANCEâ
FLEXIBLE DEPLOYMENT
Operate
An enterprise-grade foundation to run and optimize
a âmission critical automation program at high scale
13. 13
1 2 3
Understand Act
Receive
End-to-end intelligent document
processing (IDP) solution
Extracts relevant data from the
documents
Requests or
communications with
attached documents:
⢠Multiple languages
⢠Various formats
⢠Handwriting
⢠Signatures
⢠Skewed & low-quality scans
⢠Checkboxes
⢠Tables
Extracts key intent, sentiment and
context data from messages
Human in the loop
Asking employees to validate the
results if required or in case of
inaccuracies and exceptions.
UiPath Automation
Route the extracted actions and data
to downstream systems for further
processing.
14. 15
Demo
⢠GPT Based labelling
⢠GPT Based Extraction
⢠GPT Based Q&A
⢠DU and Comms Mining Demo
15. 16
Demo 3:Clipboard AI â Financial Systems Demo
Address Change
Address Change
Financial Clipboard AI Demo
16. How to develop a
trustworthy and
unbiased AI model using
LLM & GenAI ?
17. 18
Stuart Russell
Professor of Computer Science The University of California, Berkeley
âThe most robust and general solutionâone that does not
require cumbersome and potentially restrictive licensing
authoritiesâis for the software object to come with its own proof
of safety that the hardware can check efficiently. In essence,
this means switching from (A) machines that run anything
unless itâs known to be malicious to (B) machines that run
nothing unless itâs known to be safe. Obviously, making this
switch is a huge lift for governments, industry, and users, but it
can be accelerated if software vendors release new versions of
their products that will run only on type-B machinesâ
18. 19
Bias
⢠Real and potential harms
to protected categories
of individuals Harms
arise from several
causes, including data
sets polluted by historical
biases in society, data
sets that fail to represent
protected categories
adequately, and a
misunderstanding of the
sociotechnical context in
which a machine
learning system will be
applied.
Manipulation
⢠Social media
recommender systems
determine what billions
of people read and
watch every day. They
have more power over
human cognitive intake
than any dictator in
history. Yet they remain
largely unregulated.
Disinformation and
deep fakes
⢠LLMs can create
individualized
disinformation on a huge
scale to disrupt societies
and pervert democratic
processes. There are
already more than 300
fully automated ânewsâ
websites consisting of
AI-generated and largely
fake or content-free
news articles.
Impact on employment
⢠While classical economics
discounts the possibility of
long-term technological
unemployment, more
recent research
acknowledges its
inevitability as AI systems
begin to outperform large
sections of the population
in a broad range of tasks.
Generative AI Risks â Stuart Russell
19. 20
⢠A new regulatory agency for AI
⢠the EU Parliament has recently inserted clauses requiring the creation of an EU-wide AI Office
⢠International Coordination of AI
⢠An international coordinating body seems essential; proponents differ as to whether it should be modeled on the IAEA,
ICAO, IMO, etc.
⢠The OECD has formed an Expert Group on AI Futures . Details here and here.
⢠The World Economic Forum has formed a Global Council on the Future of AI,its focus is on the regulation of generative
AI.
⢠UNESCO, after developing and unanimously passing its Recommendation on the Ethics of AI, formed a High-Level
Expert Group on Implementation. Its mission is to help member states turn principles into laws.
⢠GPAI (the Global Partnership on AI) has a Working Group on Responsible AI.
⢠The European Union has drafted an AI Act covering many of the issues related to AI; Details here and here.
Regulating AI (Stuart Russell)
20. 21
AI without the automation is like the brain without the body!
Graham Sheldon â UiPath CPO
AI-powered automation
Open | Flexible | Responsible
Supported by UiPath Built with UiPath or BYO
Docs Screens Tasks Processes
Solutions
Infrastructure
Integration Service â Validation Station â Active learning â Fine tuning â Guardrails â Auditing
Generative AI Specialized AI
Context
HITL
UI
API
Action
People
Comms
Docs
Data
Processes
AI-powered automation
Delivering enterprise automation with Specialized AI
21. 22
Use any Generative AI model in your automations
Amazon
Google
OpenAI
Falcon-40B â AI21 Labs â Stability AI
Public Preview (via SageMaker)
PaLM 2
Public Preview (via Vertex AI â July)
GPT-4
General Availability
Azure OpenAI GPT-4
General Availability
Generative AI Connectors
Others Connector Builder / BYO
Microsoft
24. 25
â˘Chatbots have been mainstream in the eCommerce sector since their inception.
â˘Voice assistants use automatic speech recognition and Natural Language Processing to give
vocal responses to queries. Examples: Siri and Google Assistant products.
â˘AI avatars are 3D models designed to look like humans, that are used for entertainment
applications, or to give a human touch to virtual customer support interactions. Example NVIDIA
Omniverse Avatar Platform.
â˘Domain-specific virtual assistants are highly specialized implementations of AI virtual
assistants designed for very specific industries, and are optimized for high performance in
travel, finance, engineering, cybersecurity, healthcare and other demanding sectors.
Technologies: Speech to text, Computer Vision, NLP, Deep Learning, Augmented Reality,
Specialized AI etc.
A virtual assistant, also called an AI assistant or digital assistant, is an application program that
understands natural language voice commands and completes tasks for the user.
What is a Virtual Assistant
25. 26
Automating white collar tasks with current technologies
Task Task Task Task Task
Job
Task Task Task Task Task
Job Job
Task Task Task Task
Totally Automated Jobs
Automated Manual
Tasks
Automated Manual
5%
65%
From the work of Professor Daniel Susskind â Oxford University
26. Two examples from Hyper Hack 2023:
Team Name: Doctor SimplifAI
Team Name: MedicAI
27. 28
Revant Singh
Director @ t
revant.singh@simplifynext.com
Vincent Parnabass
Architect @ SimplifyN t
vincent.parnabass@simplifynext.com
Ikshit Dhawan
Consultant @ t
ikshit.dhawan@simplifynext.com
The Team
Rong Tao Zheng
Intern @ SimplifyNext
rongtao.zheng@simplifynext.com
28. 29
Intelligent Assitant for Doctors : Patientâs Journey and
Pain Points during a Doctor Visit
Register and wait for
consultation
Meet the doctor for
consultation
Doctor spends more attention
typing what youâre saying
than focusing on you / your
mannerisms
Doctors give advice on top off
their head and may miss
important points, leading to
mistakes
Personal experience:
Chicken pox misdiagnosed
as viral hives; 2nd Doctor
almost forgot the most
critical point in recovery
monitoring (symptom that
could mean brain swelling /
haemorrhaging)
Wait for prescription
at the same time
Doctor takes time to do all
post-consultation
documentation
15 mins to 1 hour
5 â 20 mins
Doctorâs pain points:
⢠Burnout due to the amount of time
spent on documentation (4 in 10
medical practices had a physician
retired early or resigned in the past
year due to burnout*)
⢠Slow EHR and EMR** systems
diminish care quality
MeanwhileâŚ
*August 2022 Poll of Physicians by STAT and the Medical Group
Management Association
**EHR: Electronic Health Record | EMR: Electronic Medical Record
29. 30
Proposed Solution
Send the differentiated text to
LLMs have it summarise the
content for us.
Voice differentiation to identify patient
and doctor before sending to Large
Language Model.
In addition to what the patient tells
us, we will also leverage on other
data sources (lab reports, outputs
of medical imaging).
Creates checklist for the Doctor to
ensure nothing is forgotten,
medications to take, all emailed to
the patient instantly.
Get summary of what is going on
from the model based on all inputs
(conversation, lab reports, diagnostic
imaging), editable for the doctor to
update.
1. Voice-to-text conversion 2. Summarise conversation
Using Pattern 1: Reader/Writer
3. Augment with other sources
(Multimodal AI)
4. Summary recommendation 5. Patient checklist and
medication emailed on the spot
Using Pattern 2: Analyst
Demo
31. 32
Problem Statement / Globally
1 (FDA Food and Drug Administration, 2019)
WHO World Health Organization
2 Claesson CB, Burman K, Nilsson JLG, Vinge E. Prescription errors detected by Swedish pharmacists. Int J Pharm Pract. 1995;3:151-6.
3 Khoja T, Neyaz Y, Qureshi NA, Magzoub MA, Haycox A, Walley T. Medication errors in primary care in Riyadh City, Saudi Arabia. East Mediterr Health J. 2011;17:156-9.
4 12 Zavaleta-Bustos, Miriam, Lucila Isabel Castro-Pastrana, Ivette Reyes-HernĂĄndez, Maria Argelia LĂłpez-Luna, and Isis Beatriz BermĂşdez-Camps. Prescription Errors in a Primary Care University Unit: Urgency
of Pharmaceutical Care in Mexico. Revista Brasileira De CiĂŞncias FarmacĂŞuticas Rev. Bras. Cienc. Farm 2008;44:115-25.
5 Chan M, Nicklason F, Vial JH. Adverse drug events as a cause of hospital admission in the elderly. Intern Med J. 2001;31:199-205.
6 Patel KJ, Kedia MS, Bajpai D, Mehta SS, Kshirsagar NA, Gogtay NJ. Evaluation of the prevalence and economic burden of adverse drug reactions presenting to the medical emergency department of a tertiary
referral centre: a prospective study. BMC Clin Pharmacol. 2007;7:8.
7 Pirmohamed M, James S, Meakin S, Green C, Scott AK, WalleyTJ, et al. Adverse drug reactions as cause of admission to hospital: prospective analysis of 18820 patients. BMJ. 2004;329:15-19.
The problem is likely more pronounced in
the elderly, because of multiple risk factors,
one of which is polypharmacy (5)
It has been estimated that in some
countries approximately 6-7% of hospital
admissions appear to be medication
related, with over two-thirds of these
considered avoidable and thus, potentially
due to errors (6-8)
The FDA receives more than 100,000 reports every year
that are associated with medication errors. (1)
A Swedish study found a medication error rate
of 42%. (2)
A study from Saudi Arabia reported that just under 20% of
primary care prescriptions contained errors (3)
Another study in Mexico observed that 58% of
prescriptions contained errors (âŚ)(4)
32. 33
Problem Statement / Individually
Medicine boxes without the braille alphabet.
Patient Information Leaflets text is in very small font. Itâs very difficult to read
for elderly and visually impaired people.
Over-the-counter medicines could cause crucial complications for
pregnant, breastfeeding women and children.
The risk for unsupervised athletes to take medicines which
can cause doping effects.
33. 34
Proposed Solution
Public UiPath App that enables individuals to
â access the key information from patient
information leaflets of medicines,
â access the relevant information about their
conditions from the official medical database,
â ask their personal questions related to
medicines or conditions.
Attended Robot which
â monitors the official medical system that
doctors use,
â gathers the patient details and selected
medicine information,
â analyses the medicine-patient suitability,
â finally returns the recommendations to the
doctorâs screen with a callout in real-time.
1 2
34. 35
Details & Technologies Used
End-User
Individuals
⢠Elderly People
⢠People with Visual Impairments
⢠Pregnant, Breastfeeding Women
⢠Children
⢠Athletes
⢠Tourists
⢠Doctors
User
Department
Any
Industries Health Care / Any
UiPath Products Used
⢠Studio & Studio Web,
⢠Orchestrator,
⢠Integration Service,
⢠Cloud Robots,
⢠Apps
Other â Integrations / APIs
/ Technologies Used
⢠OpenAI,
⢠Fixie Agent API â Conversational AI apps
using LLM
⢠Google Speech Solutions
35. 36
Benefits / MedicAI App
Ensuring users quickly and effortlessly find the most
relevant and accurate information with ease within
a vast data collection
Inclusive accessibility for all users by voice-enabled QA robot
The advantage of asking personalized questions in a
natural and conversational interaction
Enhanced functionality on mobile â Medicine
identification from a photo capture
Search for information in any country in any language with
seamless multi-lingual support
36. 37
Benefits / Doctorâs Assistant
Quickly and effortlessly get the patient-medicine
suitability recommendation
Never miss any side effect that can affect not only the
patientâs health but also their lifestyles
38. 39
George Roth
AI Evangelist
Presenter
george.roth@uipath.com
LinkedIn
Biography
George Roth is an entrepreneur originally from Cluj, Romania, now
residing in Los Gatos, CA, a town in Silicon Valley.
He joined UiPath in May 2018 following the acquisition of his
companyâs Intelligent Document Processing platform that was the
foundation for the Document Understanding platform at UiPath. He
was the CEO of Recognos Inc. for 18 years and was a pioneer in
using NLP and AI for document processing.
George is a big fan of Foundational Models and Generative AI and
of promoting the AI based automation platform of UiPath.
Prior to his work with Recognos and UiPath, George co-founded
multiple companies and worked as a system architect in the San
Francisco Bay Area. He holds a Master's Degree in Mathematics
and Informatics from the University Babes Bolyai in Cluj, Romania.
George is actively involved in organizations like Alianta.org, and
Romanian United Fund, and he serves as the Honorary Consul of
Romania in the San Francisco Bay Area.
In his free time, he enjoys sports, music, diplomatic activities, and
supporting Romanian startups.
Thank you