AC Atlassian Coimbatore Session Slides( 22/06/2024)

This is the combined Sessions of ACE Atlassian Coimbatore event happened on 22nd June 2024
The session order is as follows:
1.AI and future of help desk by Rajesh Shanmugam
2. Harnessing the power of GenAI for your business by Siddharth
3. Fallacies of GenAI by Raju Kandaswamy

1. Actionable Strategy for implementation of AI Help Desk
Gen AI and Future of Help Desks

2. Agenda
1.Top Workplace Applications
2.Common Help Desks
3.AI Help Desk
4.Traditional vs AI Help Desk
5.Benefits of AI Help Desk
6.Implementation Strategy
7.Getting Started

3. Top Workplace Applications
Initiative ROI Considerations
Content Creation High
Reduced content creation costs, Increased content output, Improved
content personalization
Data Analysis & Insights High
Improved decision-making, Identification of new opportunities, Reduced
time spent on manual data analysis
Help Desk Automation High
Reduced support operations costs, Improved customer satisfaction,
Frees up human agents for complex tasks
Enterprise Search Moderate
Improved information retrieval, Increased knowledge sharing, Improved
employee satisfaction
Code generation Moderate
Increased developer productivity, Reduced development time and costs,
Potentially fewer coding errors

4. Common Help Desks
IT Help Desk
Provide technical support and resolve technology-related issues
Common queries:
-I've forgotten my password. How do I reset?
-I need access to Jira
-VPN isn't working for me?

5. Common Help Desks
HR Help Desk
Address employee-related inquiries and support HR processes
Common queries:
-How do I request time off?
-How do I change my 401(k) contribution?
-What's our work-from-home policy?

6. Common Help Desks
Customer Service Help Desk
Provide support and assistance to external customers
Common queries:
-How do I change our account administrator?
-How do I cancel my subscription?
-We're experiencing [specific error]. How can we resolve this?

7. AI Help Desk

8. AI Help Desk

9. Traditional vs AI Help Desk
Complex Form Filling
Interface
Natural Chat
Interface
vs

10. Traditional vs AI Help Desk
9-5 Availability 24*7 Availability
vs

11. Traditional vs AI Help Desk
Slow Manual
Resolutions
Unlimited Instant
Resolutions
vs

12. Benefits
Easily and cost-effectively scale your
support operations with Gen AI powered
copilots
95%
90%
60%
Guaranteed end-user
satisfaction
High
acceleration
rate
High automatic
resolution rate

13. Implementation Strategy
1.AI Assessment
2.AI Knowledge Engineering
3.Chatbot Prototype
4.Continuous Adapative Learning
5.Advanced AI Automations
6.Human Agent Augmentation
7.AI Insights
8.AI Security and Governance

14. Implementation Strategy
1. AI Assessment
Asses your existing help desk processes
and identify areas for automation

15. Implementation Strategy
2. AI Knowledge Engineering
Organize existing company documentation,
policies, and product information

16. Implementation Strategy
3. AI Chatbot Prototype
Build a simple FAQ knowledge base

17. Implementation Strategy
4. Continuous Adaptive Learning
●Set up Continuous learning by connecting to
knowledge sources like Confluence, SharePoint
●Train the chatbot on historical help desk tickets
and past conversations in Slack/Teams

18. Implementation Strategy
5. Advanced AI Automations
●Understand complex questions
●Mimic human actions in business apps
●Provide personalized responses
●Understand images and videos

19. Implementation Strategy
6. AI Agent Augmentation
●Rephrase answers for tone adjustment
●Summarize customer conversations
●Identify situations that need human handover
●Understand emotional tones and trigger escalations
●Turn conversations into knowledge assets

20. Implementation Strategy
7. AI Insights
Monitor and track automation rates,
identify gaps and opportunities

21. Implementation Strategy
8. AI Security and Governance
-Anonymize training data, user queries
-Ensure company data is not used train LLMs
-Ensure ISO, SOC2 and GDPR compliance

22. Getting Started
Build a custom Gen AI chatbot using
-Data layer for RAG (e.g., LlamaIndex, LangChain)
-Foundation models (e.g., Open AI, Claude, Gemini)
-Vector databases (e.g., Pinecone)
Leverage purpose-built vendor products like Enjo AI

23. Get Started with Enjo AI
1.Personalized Enjo demo
2.Help desk automation potential assessment
3.14 days no obligation free trial
4.3 months guided Pilot program

25. Siddarth Kengadaran

26. theproductguy.xyz
Who am I?
➔Product Consultant | Strategy and Design
➔Information Technology and Psychology
➔Convenor - The Product Space
➔Organizer - Google Developer Groups and Friends of Figma, Coimbatore

27. How Generative AI works?
Table of contents
The Rise of Generative AI
What is Generative AI
capable of?
Assessing Your Business
Needs
Future Trends and
Opportunities
Conclusion
01
02
03
04
05
06

28. Artificial
Intelligence (AI)
Artificial Intelligence (AI) refers to the simulation of
human intelligence in machines that are
programmed to mimic human actions and cognitive
processes.
The Rise of Generative AI

29. Logical Reasoning &
Problem-Solving
Abstract Thinking
Learning & Adaptation
Memory
Language &
Communication
Perception &
Sensory Processing
Emotional
Intelligence

30. Social Intelligence
Creativity &
Imagination
Decision-Making
Metacognition
Spatial Reasoning
Numerical &
Quantitative Skills
Practical Intelligence
Moral & Ethical
Reasoning

31. Expert systems, rule-based systems, automated reasoning,
theorem proving, constraint satisfaction algorithms.
Deep learning, neural networks, generative models (e.g.,
GANs, VAEs), reinforcement learning.
Natural language processing (NLP), natural
language understanding (NLU), natural
language generation (NLG), machine
translation, chatbots, language models (e.g.,
GPT-4).
Machine learning (supervised, unsupervised,
semi-supervised, and reinforcement learning), adaptive
systems, transfer learning, lifelong learning systems.
Knowledge graphs, semantic networks, databases,
memory-augmented neural networks, long short-term
memory (LSTM) networks.
Computer vision, speech recognition, audio
processing, sensor fusion, image and video
recognition systems.

32. Affective computing, sentiment analysis, emotion
recognition systems, empathy bots.
Social robots, conversational agents, virtual assistants,
social network analysis.
Meta-learning, self-improving AI, automated
machine learning (AutoML), reflective agents.
Generative adversarial networks (GANs), creative AI, music
composition AI, art generation AI, creative writing AI.
Decision support systems, recommendation engines,
optimization algorithms, predictive analytics.
Robotic perception, pathfinding algorithms,
spatial analytics, autonomous navigation
systems, 3D modeling.

33. Data analytics, statistical analysis software, financial
modeling AI, algorithmic trading systems.
Robotics, autonomous systems, smart appliances,
context-aware computing.
Generative adversarial networks (GANs), creative AI, music
composition AI, art generation AI, creative writing AI.
AI ethics frameworks, fairness-aware AI, explainable AI
(XAI), bias detection and mitigation tools.

34. Artificial
Intelligence[AI]
Machine
Learning [ML]
Natural Language
Processing [NLP]
Deep Learning
Vision Speech
Robotics
Planning
Expert
Systems
Neural Networks
Generative AI

35. The Rise of Generative AI
Machine Learning (ML)
Machine Learning (ML) is a subset of AI that
enables systems to automatically learn and
improve from experience without being
explicitly programmed.
Deep Learning
Deep Learning is a subset of machine
learning that uses neural networks with
multiple layers to learn hierarchical
representations of data.

36. Generative AI
Generative AI falls under the umbrella of Machine
Learning, particularly within the realm of deep
learning. It's a specialized type of model that
leverages neural networks (often very large and
complex ones) to generate new data that resembles
the data it was trained on.
The Rise of Generative AI

37. ✦Abstract Thinking
✦Language & Communication
✦Creativity & Imagination

39. 1966
2017
2023
OpenAl GPT-3
May: OpenAl releases GPT-3, the largest language model to date with 175 billion parameters.
Microsoft Introduces GPT-4
March: Microsoft debut OpenAl's GPT-4 likely a multimodal trillion parameter version of GPT-3
Introduction of Transformer Models
Transformer Models are introduced through papers like Google's Transformer: A Novel
Neural Network O Architecture for Language Understanding and Attention Is All You Need,
Vaswani et al., 2017.
2020
2024
Meta introduces LLaMA 3
June: AI model that surpasses previous versions in terms of versatility and language generation,
with better contextual understanding and reduced biases.
Statistical Language Model (N-gram model)
2022

40. Statistical Language Model (N-gram model)
An n-gram model breaks text down into chunks of n consecutive words (or
"grams") to predict the next word in a sequence. Let's use a 3-gram (trigram)
model for simplicity.
Our model has been trained on a large corpus of text, and it has learned that
after the sequence "The cat is on the", the most probable next words are
"roof", "floor", "bed", or "mat", let's say.
It knows nothing more than the statistical probability of each of these words
appearing after the input sequence based on its training data.
So, if "roof" appeared most frequently in its training data after the phrase
"The cat is on the", it would predict "roof" as the next word.

41. Neural Network Language Model (like GPT-4)
These models take a more sophisticated approach. They don't just look at
the immediate previous words, but they understand the entire context of the
input and have a notion of word meaning derived from their training data.
Now, if we had a more nuanced sentence like:
"The cat spotted a mouse. Quietly, it started to climb. The cat is on the..."
Despite the commonality of phrases like "the cat is on the floor/bed/mat", a
neural network model like GPT-4 might predict "chase" or "prowl", as it
can understand from the earlier part of the sentence that the cat is likely
pursuing the mouse, and "climb" implies an upward movement, possibly
indicating something like a table or a counter.

42. Large
Vision-Language
Models

43. Model
The result of the machine's learning process. The model holds the patterns
and insights the computer discovered from the training data, allowing it to
make predictions or take informed actions on new information.
Foundation
Model
Adapted Models
Domain-Specific
Models
Task-Specific
Models
Hybrid Models
Multimodal
Models
Explainable &
Interpretable Models
Personalized
Models

44. Foundation Model
BERT, GPT-n,
DALL-E,..
Adapted Models
BioGPT
Domain-Specific Models
BloombergGPT
Task-Specific Models
Whisper
Hybrid Models
Multimodal Models
Gemini
Explainable & Interpretable Models
Personalized Models
Apple Intelligence

45. Data
Text
Images
Audio
Structured
Data
3D Signals
Video
Foundation
Model
Tasks
Question &
Answering
Summarization
Generation
Extraction
Paraphrase
Search
Classification
Analysis
Captioning
Recognition
Translation
Rephrase
ReasoningPrediction
EnhancementSegmentation
Deciding &
Planning
Conversion

47. Generative pre-training
Fine-tuning
Retrieval-augmented
generation (RAG)
Prompt engineering
Complexity
Accuracy
Cost
Time to Implement
Domain Specificity
Flexibility

48. Prompt engineering
Complexity
AccuracyCost
Time to
Implement
Domain
Specificity
Flexibility

49. Retrieval-augmented
generation (RAG)
Complexity
AccuracyCost
Time to
Implement
Domain
Specificity
Flexibility

50. Fine-tuning
Complexity
AccuracyCost
Time to
Implement
Domain
Specificity
Flexibility

51. Generative pre-training
Complexity
AccuracyCost
Time to
Implement
Domain
Specificity
Flexibility

53. LLM OS
Agents
RAG
Chat Bot
Question & Answers
Levels of LLM Apps
Predicts answers based on patterns learned
from a vast corpus of text.
Engages in interactive dialogues by
generating contextually relevant responses.
Retrieves and incorporates information
from external knowledge sources to
enhance responses.
Executes actions in external systems based
on user requests and retrieved information.
Orchestrates multiple agents and processes,
managing complex tasks and workflows
through a unified interface.

54. ✦MAKER
Train and build custom models
✦SHAPER
Tune foundational Industry Models
✦TAKER
Use pre-trained ML API models and point to
your apps

55. Thank you!
theproductguy.xyz

56. Thefallacies ofGenerative AI
Opportunities& Challenges inproductionizing

57. Introduction
WhoamI?
Overview ofSession Goals
•Fallacies ofGenerative AI
•Opportunities & Challenges
•Strategiesforproduct
•Impactsand Future ofAI
Agenda

58. TheAIPyramid
Don'tcareaboutAI
UseAI at work
IntegrateAIin
Enterprisesolutions
BuildAI
fromscratch
Big techlandscape
Opportunity
Opportunity

59. Opportunity Radar

60. Hype Cycle
Trends that would reach productivity < 2 years
•Retrieval Augmented Generation (RAG)
•GenAI Enabled applications
•GenAI Workload accelerators
•GenAI enabled virtual assistants

61. Common Fallacies
šFallacy 1: Generative AI is a Magic Bullet
šMyth: One magicpillfor all our problems
šReality: Needs specific training and fine-tuning, richness of AI comes from your data
šFallacy 2: Generative AI Understands Context
šMyth: Perfect comprehension of context and nuance
šReality: Limitations in understanding complex, nuanced queries
šFallacy 3: Generative AI is Completely Autonomous
šMyth: No human intervention needed
šReality: Requires human oversight and validation

62. Data Powers AI - Challenges
The data needed to power AI are scattered
Data are not NLP friendly
Solution: Transform your data platform as single source of truth with NLP friendly JSON schemas
Data platforms are
not built for AI
AI may not be able to chew our complex production schemas
Solution: Keep the data for AI mostly in denormalized form like a datawarehouse
Complex production
data
Duplication of data results in Bias
Solution: Qualitative assessment of dataData duplication

63. Challenges in Productionizing
Bias and Hallucination
Explanation of bias in AI models
Examples of hallucination and its
impact
Inference and Infrastructure
Technical challenges in deploying
Generative AI
Infrastructure requirements and
associated costs
Cost Considerations
Budgeting for AI development and
maintenance
Cost vs. benefit analysis

64. ProductStrategy
šNot allproducts andsolutionneedGenerativeAI
šAsuccessful Generative AIbased productexhibits
oGenerative AIis not aforce fit into theproduct
oProduct deliversvalueeven withoutGenerativeAI
oSolves areal userpainpoint or a problem
oUseof generative AI amplifies thevaluedelivered
oAutomates a complexset of operations within the
product

65. IdeatoProduct
PoC/MVPwithcommercialAPIs(ChatGPT, Gemini)
Richbusinessdomaindataset
Self hostedLLMmodels
A/B&RCwithfeedback
Fine tuning
01
02
03
04

66. Model
Landscape

67. Team
Composition
AI architect
Data Scientist
Data Engineer
Developer
Business Analyst
/ Scrum MasterQuality Analyst
Product
Manager
NLP Engineer
(Optional)
MLOps
Engineer
(Optional)

68. OtherConsiderations
šAIassistedsoftwaredelivery
šAIassistedcontentcreation
šAGI?
šAI andworkforce
šSuccess Stories
šAdobe – Generative Fill
šMicrosoft 360 – GenAI Designer

69. Recap
šOpportunities andchallenges
šYour AI Strategy

70. Questions?

71. Thank you!
Raju Kandaswamy
https://www.linkedin.com/in/rajukandasamy/
https://medium.com/@raju.kandasamy