DnT Infotech

LARGE
LANGUAGE
MODEL
Present by Atharva Virkar
FLOW
1 Introduction 6 What’s the Hype

2 Traditional Vs Transformers 7 Workflow

3 LLMs 8 Concept Breakdown

4 Gen AI 9 Real life Adaptation and

Example
5 Introduction to Agentic AI
INTRODUCTION
In recent years, we’ve entered a new era of artificial intelligence, one thats
driven by Large Language Models (LLMs) and Generative AI (GenAI). These
technologies are reshaping how we interact with machines, automate tasks,
and generate content across industries.

It all started with a breakthrough research paper published by

Google Brain in 2017 titled “Attention is All You Need.”

This paper introduced the Transformer architecture, the

building block of all modern LLMs and GenAI systems.
TRADITIONAL METHODS
VS
TRANSFORMERS
 Before Transformers, it was RNNs & LSTMs Era.

RNN (Recurrent Neural Networks) : A neural network designed to handle

sequential data by having connections that form directed cycles. It remembers
previous inputs using a hidden state.

It struggles with remembering long-term memory .

Used for: Time series prediction, Speech recognition, Language modeling

LSTM (Long Short-Term Memory) : A type of RNN with a more complex

TRADITIONAL architecture using gates (input, forget, output) to better remember and

METHODS manage long-term and short-term dependencies.

Used for : Text generation, Machine translation, Sentiment analysis

Major setbacks of RNN’s and LSTM’s are :

Sequence handling
Memory limitations
Bottleneck with content vector
Training challenges
Limited scalability
 For example we are building an email autocompletion assistant, which needs to
predict the next word "I wanted to follow up on the meeting we had last..."
Words such as week, Friday, or month.

RNNs take on it will be :

RNN reads the sentence word by word and tries to remember what came
before.
It might forget the early context ("I wanted to follow up") and focus only on
EXAMPLE OF recent words like “we had last”.
Because it has a limited memory, by the time we reach "last...", the memory
WORKING of “meeting” is faint or lost. So, it might predict something vague like "time".
RNN fails when the important clue was far back in the sentence.
RNN & LSTM LSTMs take on it will be :

LSTM also reads word by word, but it has gates that help it remember that
the sentence started with “follow up on the meeting”.
It has 3 gates : Forget gate, Input gate, & Output gate. Which decides what
to ignore, what to remember and what to show to the next step.
So even if “meeting” was several words ago, it remembers that it’s important.
It is more likely to predict "week" or "Friday" because it knows we’re
referring to a meeting.
 RNN Structure: Think of it as a Relay Race
RNN processes the sentence one word at a time.
Each word passes its meaning to the next like a baton in a relay race.
There's one shared memory (hidden state) passed along the chain.
Early information (like "meeting") has to travel through every step to reach the end.

The message gets weaker with each step — like a story passed from person to person. So by the time we
reach "last...", the memory of “meeting” is faint or lost.

To summarise it, we can say RNN is a one linear path, memory passed forward step-by-step, prone to
forgetting earlier parts.
STRUCTURE OF RNN

LSTM Structure: Like a Smart Assistant with a Notepad.

Still processes word-by-word, but with internal gates:

Forget gate: Decides what to ignore
Input gate: Decides what to remember
Output gate: Decides what to show to the next step

Maintains a separate memory cell, like a notepad, so it can hold onto key facts.

It remembers important words (like “meeting”) for a long time without fading. This helps it
make better decisions even many steps later.

To summarise it, we can say LSTM is a Step-by-step model with memory management, uses STRUCTURE OF LSTM
gates to decide what to remember or forget like a person jotting down key info.
 Transformer doesn’t pass memory step by step.

Instead, it lets every word look at every other word directly

using self-attention.

Structure Of Transformer Input sentence: ["I","wanted",...,"meeting",...,"last"]

Each word is turned into a vector xi , with positional encoding.

To summarise it , we can say all words interact simultaneously,

using attention to decide what's important — like a group of
experts cross-checking notes in one meeting.
 Transformer : A Transformer is a type of model that helps computers understand
and generate human language by paying attention to all the words in a sentence at
once, rather than one by one.

Unlike earlier models, Transformers uses self-attention mechanism which can see
the whole sentence at once, making them better at understanding meaning.

It follows Encoder–Decoder architecture :

Encoder: Reads and understands the input (e.g., a sentence).
Decoder: Generates the output (e.g., translated sentence).

TRANSFORMERS Transformers led to Birth of PLMs (Pretrained Language Models) transformers

powered models which were pretrained on large text data to learn general
language patterns, such as:
BERT (2018) – Read whole sentences for understanding.
GPT (2018) – Generated text left-to-right with transformer decoders.

As transformer-based models were scaled up with more data and powerful

hardware, they evolved into GPT-2, GPT-3, and GPT-4. Models trained on
massive text corpora with billions of parameters. This scaling marked the
transition from PLMs to Large Language Models (LLMs), laying the foundation for
today's Generative AI.
GEN AI
Generative AI (Gen AI) is essentially built on Large Language Models
(LLMs) or similar foundational models.

While LLMs are the underlying engine trained to predict and generate
language, Gen AI refers to the broader suite of tools and systems using
these models to create new, original outputs.

Gen AI systems can be multimodal, combining text, images, audio, or

video generation, but the underlying principle remains the use of large,
pretrained models similar to LLMs.

In summary, Gen AI = LLM (or related foundational model) + task-specific

tuning + generation capabilities to produce creative, usable content.

Some examples of Gen AI are :

ChatGPT
DALL·E
GitHub Copilot
Music generation models
Text-to-Speech (TTS) systems
LLM
LLM : An LLM is a type of artificial intelligence (AI) that
has been trained on a massive amount of text (like books,
websites, articles) so that it can understand and generate
human-like language.

Examples of LLM are :

GPT-3,GPT-3.5,GPT-4,
LLaMA
Claude
Gemini

Parameters are the learnable values (weights) in the

model that help it understand patterns in data — they
decide how the model processes and generates text.

Every input is embedded into vectors, and these vectors

are multiplied by weight matrices , and other matrices ,
through which we fine tune our model to generate
desired text accordingly.
WHAT IS
AGENTIC
AI
INTRODUCTION TO
AGENTIC AI
Agentic AI refers to AI systems that go beyond just completing tasks or
answering prompts but actually behave like an autonomous agents, capable
of planning, decision making and acting over time to achieve a goal.

AGENTIC AI :- LLM + Memory + Goals + Autonomy

Instead of passive assistants like siri or ChatGpt, It doesn’t just

answer questions like a chatbot, it plans, takes action, uses
tools and adapts autonomously to achive a given objective.
 Endorsements from Tech Leaders
Mark Zuckerberg (CEO, Meta) - Predicts that AI agents will
replace mid-level engineers by 2025, allowing human
engineers to focus on higher-level problem-solving and
creativity.

Jensen Huang (CEO, NVIDIA) - Stated that "IT departments

will become the HR of AI agents," indicating a paradigm
shift in organizational structures to accommodate AI agents
as digital employees.

Industry Transformation
Automating Complex Tasks: AI agents can autonomously
perform tasks that traditionally required human

WHATS THE intervention, increasing efficiency and reducing errors.

HYPE ?
Enhancing Productivity: By handling routine and time-
consuming activities, AI agents free up human workers to
focus on strategic and creative endeavors.

Driving Innovation: The integration of AI agents fosters

innovation by enabling rapid prototyping, data analysis, and
decision-making processes.
 WORKFLOW
There are phases of Agentic AI:
a. Goal Understanding
b. Planning
c. Tool usage & Execution
d. Memory + Adaptation

It takes a human written prompt/task and understands

the desired outcome.

Breaks the tasks into substeps ( e.g. :- Search → Read →

Extract → Draft).

Uses API’s, Webtools or software to complete those

substeps.

Remembers what it did, checks for failures, and adjusts

the plan if needed for reiteration.
CONCEPT
BREAKDOWN
Agent :- A system that can plan, reason, and act on its
own to accomplish goals.

It's not just answering questions like a chatbot — it's

deciding what to do, in what order, and how
LLM Model Tools
LangChain, AutoGen are some popular open-source
library to orchestrate agents.

LLM (Large Language Model) — it’s the brain of the agent.

Used for planning and creating substeps.

Such as to Understand what user wants, decide what

steps to take, interpret results from tools, generate code,
summaries, decisions, etc. A gent
LLMs like GPT-3.5/4, Claude, Gemini, LLaMA, Mistral,
Zephyr are used.
CONCEPT
BREAKDOWN
Tools :- Anything the agent can "use" to get work done,
they are like the arms of the agent, they let it take action.

Tools are used for tasks like Searching, Coding, File

parser, Browser, Databases, & APIs.

DuckDuckGo, Tavily, Python REPL, code interpreter, LLM Model Tools

SQL, MongoDB, PDF reader, are used as a tool plugged
into the agent so it can act in the real world.

Agent Workflow example:- Prompt: User says

"Summarize this PDF and send it to my email.", LLM
(within agent) interprets the request.

Agent decides: → Use PDF tool → summarize → call

email API. Tools are used one by one and final output is
A gent
sent to the user.

The LLM reasons, the Agent coordinates, and the Tools

act.
 Trigger – "When chat message received"
WORKING FOR This is how it starts. The system listens for a chat (like from a user
or chatbot).

AGENTIC AI AI Agent : This is the core "brain."

It decides what to do next , just like a virtual assistant.
It doesn’t just follow steps, it thinks, plans, and uses tools to act.

Chat Model (OpenAI) : This is the "language understanding" (LLM)

part.
It understands user intent (e.g., "Schedule a meeting with team").

Memory (Simple Memory) : Helps the AI remember context across

chats.
For example, it can recall who “John” is or what the last task was.

Tools:
Google Sheets (read) – Agent reads data (e.g., get tasks or meeting
list)
Google Sheets1 (update) – Agent writes new info (e.g., log a
“This diagram represents an AI agent that acts like a smart digital assistant. When someone completed task)
sends a message, this assistant understands it, remembers useful information, and uses
external tools like Google Sheets or Google Calendar to take actions — automatically.” Google Calendar (create event) – Agent creates meetings
automatically
 REAL LIFE ADAPTATION
& EXAMPLE

Goose.ai (Agent-Driven Automation Platform) Microsoft AutoGen – Enterprise Email Assistants

A platform by Jack Dorsey's Block that lets businesses build and Microsoft uses AutoGen (multi-agent) setup internally
deploy AI agents to automate tasks like customer support, and externally.
document analysis, and workflow management. One agent reads incoming email, another plans the task
Agents are powered by LLMs and integrated with tools (email, (e.g., schedule meeting, approve document).
web search, databases). Another executes the task or generates a reply and
Example: A "Legal AI Agent" that reads 20-page contracts, finally, a checker agent reviews and sends.
highlights key clauses, and sends compliance reports Enterprise Use: Auto-pilot mode for inboxes, especially
automatically. for executives and managers.
 DnT Infotech

THANK YOU
for your time and attention

Presented by Atharva Virkar