Forem: عبدالله عياد | Abdullah Ayad

LLM on AWS with Bedrock and Understanding Non-Determinism in Generative Models

عبدالله عياد | Abdullah Ayad — Sat, 22 Nov 2025 18:11:22 +0000

Foundation Models on AWS

AWS provides access to several foundation models through Amazon Bedrock, offering users a selection of state-of-the-art models from providers such as Anthropic, AI21 Labs, Cohere, Meta, Mistral, Stability AI, and Amazon Titan. These models can be easily integrated into applications for tasks like text summarization, code generation, chatbots, and more.

Large Language Models (LLMs) on AWS

A Large Language Model (LLM) is a type of foundation model specifically trained to understand, generate, and manipulate human language at scale. According to AWS, LLMs are built using deep learning techniques, particularly transformer architectures, and are pre-trained on massive datasets containing text from books, articles, websites, and other sources. These models are capable of performing a wide range of natural language processing (NLP) tasks, including text generation, translation, summarization, question-answering, and conversational AI.

AWS enables businesses and developers to leverage LLMs through services such as Amazon Bedrock, which provides access to state-of-the-art models from multiple providers, and Amazon SageMaker, which supports custom model training and deployment.

LLMs on AWS

AWS provides access to multiple LLMs through Amazon Bedrock, offering models from providers like Anthropic (Claude), AI21 Labs (Jurassic), Cohere (Command R), Meta (Llama), Mistral, and Amazon Titan. These models can be integrated into applications for various business needs, including automated content generation, AI-driven chat interfaces, and knowledge retrieval systems.

Use Cases of LLMs

Chatbots and Virtual Assistants – Powering customer support agents and interactive AI assistants.
Text Summarization – Condensing lengthy articles, reports, and emails into concise summaries.
Code Generation and Assistance – Helping developers write, review, and debug code more efficiently.
Search and Knowledge Retrieval – Enhancing search engines and enterprise knowledge bases with AI-driven understanding.
Language Translation – Providing real-time and accurate multilingual translations.

Understanding Non-Determinism in Generative Models

Generative language models, such as Large Language Models (LLMs), are inherently non-deterministic, meaning they do not always produce the same output given the same input.

Why Are Generative Language Models Non-Deterministic?

Probability-Based Predictions
Generative models use probabilistic token selection when generating text. Instead of selecting a single "correct" next word, they compute a probability distribution over possible next words and select one based on a sampling strategy.
For example, given the prompt "The sky is", a model might assign probabilities to different words:
"blue" (85%)
"clear" (10%)
"cloudy" (5%)
Depending on the sampling method, the model may return "blue" most of the time but occasionally choose "clear" or "cloudy".
Temperature and Sampling Techniques (Important for Exam!)
Temperature: This parameter controls the randomness of the model’s outputs. A higher temperature (e.g., 1.0) increases randomness, making the responses more creative and diverse. A lower temperature (e.g., 0.1) makes the model more deterministic and focused on high-probability outputs.

Top-k Sampling: Limits the selection to the top-k most probable words, reducing randomness but maintaining variability.
Top-p (Nucleus) Sampling: Dynamically selects from the smallest set of words whose cumulative probability meets a certain threshold, balancing creativity and coherence.
Context Dependence and Variability
Generative models consider the context of the input but may weigh words differently in each run, leading to slight variations.
Small changes in punctuation, word choice, or sentence length in the input can lead to significantly different outputs.
Fine-Tuning and Model Updates
When LLMs are fine-tuned with new data or updated with improved training methodologies, their internal probability distributions shift, making previously common responses less likely and introducing new variations.

Implications of Non-Determinism

Creativity & Diversity – Non-determinism allows AI-generated text to feel more natural, avoiding repetitive or overly predictable responses.
Challenges in Consistency – For applications requiring precise, repeatable outputs (e.g., legal documents or financial reports), fine-tuning model parameters and using lower temperatures can help achieve more deterministic behavior.
Bias & Uncertainty – Since responses are probabilistic, slight variations in prompts can lead to different biases or interpretations, requiring careful prompt engineering and model adjustments.

GitHub
LinkedIn
Medium

What is Generative AI For AWS and What is Amazon Bedrock Service?

عبدالله عياد | Abdullah Ayad — Fri, 17 Oct 2025 19:36:49 +0000

Generative AI definition for AWS

AWS defines Generative AI (Gen AI) as a subset of artificial intelligence that focuses on creating new content, including text, images, code, audio, and more. It is powered by foundation models (FMs), which are large-scale machine learning models trained on vast datasets and capable of performing a wide range of tasks with minimal fine-tuning.

AWS provides Amazon Bedrock, a managed service that enables users to build and scale Gen AI applications using foundation models from leading AI providers without needing to manage infrastructure. Additionally, AWS Trainium and AWS Inferentia offer high-performance, cost-efficient hardware for training and deploying large AI models.

Foundation Models: The Backbone of Generative AI

A foundation model (FM) is a large-scale machine learning model trained on vast amounts of data, enabling it to perform a wide range of tasks with minimal fine-tuning. According to AWS, foundation models serve as the core technology behind generative AI, allowing users to generate text, images, code, and other types of content by leveraging advanced deep learning techniques.

These models are designed to be general-purpose, meaning they can be adapted for multiple applications, such as natural language understanding, text summarization, image generation, and chatbot interactions. Unlike traditional machine learning models, which are trained for specific tasks, foundation models are pre-trained on diverse datasets and can be fine-tuned to suit specialized needs.

Key Characteristics of Foundation Models

Pre-Trained at Scale – Foundation models are trained on massive datasets using self-supervised learning techniques. This allows them to develop a deep understanding of language, images, or other structured data without the need for explicit labels.
Multi-Task Capability – Unlike traditional AI models that are trained for a single task, foundation models can be used for a variety of applications with minimal modifications. For example, the same model can be used for sentiment analysis, machine translation, and content generation.
Fine-Tuning and Adaptability – While foundation models are powerful in their raw form, they can be further optimized for specific tasks through fine-tuning. This process involves training the model on a smaller, domain-specific dataset to improve accuracy and performance in a particular field.
Efficient Deployment via APIs – AWS provides access to foundation models through services such as Amazon Bedrock, which allows businesses to integrate these models into their applications without needing to build or maintain the underlying infrastructure.
Optimized for Performance and Cost – AWS offers specialized hardware such as AWS Trainium and AWS Inferentia, which are designed to accelerate the training and inference of foundation models while optimizing cost efficiency.

GitHub
LinkedIn
Medium

Generate-Blogs-using-Amazon-Bedrock-Service

عبدالله عياد | Abdullah Ayad — Fri, 03 Jan 2025 11:48:32 +0000

Before you start, here are some prerequisites:

Python
AWS Background
AWS Lambda Function
Machine Learning Skills

So, let's get started...

Steps:

Here is the code repo we need https://github.com/a00ayad00/Generate-Blogs-using-Amazon-Bedrock-Service

Go to AWS Lamda Function and Create a function with Python 3.11
Create a layer in the function: Go to the repo and download the zip file boto3_layer.zip or:

Go to any terminal and run the following command line:
pip install boto3 -t python/
Add this python folder to another folder and zip the folder
Go to the layer section and start to create one

. Upload this zip file when you create the layer

Put the code that is in the repo (app.py) in the function
You can test it by creating an event that its body looks like

{
    'blog_topic': any_topic_you_want
}

Next, you can make an API from this app (API Gateway). Please tell me if you want me to write another blog explaining how to create an API from this function.

GitHub
LinkedIn
Medium

How did I pass the AWS Machine Learning Speciality Exam?

عبدالله عياد | Abdullah Ayad — Fri, 26 Jul 2024 18:40:15 +0000

Intro

On June 24, 2024, I passed the AWS Machine Learning Speciality Exam and became Amazon certified.
You can see my certification from here and my badge on credly from here.
Book your exam from https://www.aws.training/certification?banner=mlexam

1- Official Certification Resource

Certification Home Page — https://training.resources.awscloud.com/get-certified-machine-learning-specialty
Certification Prep Page — https://aws.amazon.com/certification/certified-machine-learning-specialty/

2- Udemy Courses

For AWS Machine Learning Specialty Certification, I chose the Udemy course taught by Frank Kane. — https://www.udemy.com/course/aws-machine-learning/?couponCode=ST3MT72524
It's very beautiful to see this course too taught by Neal Davis https://www.udemy.com/course/aws-certified-machine-learning-specialty-mls/?couponCode=ST3MT72524

3- Tutorials Dojo — AWS Certified Machine Learning Specialty Practice Exams

They have the best and the cheapest à la carte option to buy the practice exam for the certification you're interested in — https://portal.tutorialsdojo.com/courses/aws-certified-machine-learning-specialty-practice-exams/

4- Exam Topics

One of the best collections for many international exams you want and their exam materials are up to date — https://www.examtopics.com/

5- More Practice Questions

Questions and answers for the year 2021 — https://drive.google.com/file/d/1X25zgOl_MWSM83LRlJShlabghmyKRQCK/view?usp=drive_link
Sample 10 Questions — https://d1.awsstatic.com/training-and-certification/docs-ml/AWS-Certified-Machine-Learning-Specialty_Sample-Questions.pdf

Each AWS Certification exam has 10 sample questions as a standard.

Practice 20 Questions — https://explore.skillbuilder.aws/learn/course/external/view/elearning/12469/aws-certified-machine-learning-specialty-practice-question-set-mls-c01-english

Each AWS Certification exam also has 20 practice questions as a standard.

Exam Readiness Course with almost 60 Questions https://explore.skillbuilder.aws/learn/course/external/view/elearning/27/exam-readiness-aws-certified-machine-learning-specialty

6- My Drive and GitHub for more

You can see many resources, books, and courses on my GitHub Repo and my collections on Google Drive

Good Luck with your preparation 💖❤️👌

GitHub
LinkedIn
Medium

How did I pass the AWS Machine Learning Speciality Exam?

عبدالله عياد | Abdullah Ayad — Fri, 26 Jul 2024 18:40:15 +0000

Intro

1- Official Certification Resource

Certification Home Page — https://training.resources.awscloud.com/get-certified-machine-learning-specialty
Certification Prep Page — https://aws.amazon.com/certification/certified-machine-learning-specialty/

2- Udemy Courses

3- Tutorials Dojo — AWS Certified Machine Learning Specialty Practice Exams

4- Exam Topics

One of the best collections for many international exams you want and their exam materials are up to date — https://www.examtopics.com/

5- More Practice Questions

Questions and answers for the year 2021 — https://drive.google.com/file/d/1X25zgOl_MWSM83LRlJShlabghmyKRQCK/view?usp=drive_link
Sample 10 Questions — https://d1.awsstatic.com/training-and-certification/docs-ml/AWS-Certified-Machine-Learning-Specialty_Sample-Questions.pdf

Each AWS Certification exam has 10 sample questions as a standard.

Practice 20 Questions — https://explore.skillbuilder.aws/learn/course/external/view/elearning/12469/aws-certified-machine-learning-specialty-practice-question-set-mls-c01-english

Each AWS Certification exam also has 20 practice questions as a standard.

Exam Readiness Course with almost 60 Questions https://explore.skillbuilder.aws/learn/course/external/view/elearning/27/exam-readiness-aws-certified-machine-learning-specialty

6- My Drive and GitHub for more

You can see many resources, books, and courses on my GitHub Repo and my collections on Google Drive

Good Luck with your preparation 💖❤️👌

GitHub
LinkedIn
Medium

High Level Machine Learning services on AWS

عبدالله عياد | Abdullah Ayad — Sun, 17 Mar 2024 20:31:00 +0000

Recognition is a way for you to do face detections, to do labeling, and celebrity recognition.

Transcribe is for you to get, a way to get subtitles, for example, to convert your audio into text.

Polly is the opposite. It allows you to get, to use your text and create audio out of it.

Translate is for you to get translations.

Lex is to build conversational robots or chatbots.

Connect if you bundle Lex with Connect, then you can create a cloud contact center.

Comprehend is a way for you to do natural language processing.

SageMaker is a fully featured machine learning service that is accessible to developer and data scientist.

Forecast allows you to build highly accurate forecast.

Kendra is going to be an ML-powered document search engine.

Personalize is used for real-time personalized recommendations for your customers.

Textract is used to detect text and data, and extract them from various documents.

GitHub
LinkedIn
Facebook
Medium

Amazon Textracts Overview

عبدالله عياد | Abdullah Ayad — Mon, 11 Mar 2024 20:45:00 +0000

Amazon Textract is used to extract texts, so, hence the name.
So you extract text, handwriting, or data from any scanned document and behind the scenes, of course, uses AI or machine learning.

So we have, for example, a driver license, and then we upload it into Amazon Textract, and then, automatically, will be analyzed, and the results will be given to you as a data file, and so you'll be able to, for example, extract the date of birth, document ID, and so on.

So you can extract any data, even from forms and tables, and you can read PDFs, images, and so on.

The use cases for extracting texts are multiple, but you could be for financial services to process invoices or financial reports, could be for healthcare, for medical records, and insurance claims, or for the public sector, for example, for tax forms, ID documents, and passports.

GitHub
LinkedIn
Facebook
Medium

Amazon Personalize Overview

عبدالله عياد | Abdullah Ayad — Tue, 05 Mar 2024 20:37:00 +0000

It is a fully machine learning service to build apps with real-time personalized recommendations.

So what could be a recommendation?
For example, a personalized product recommendation, or re-ranking, or customized direct marketing.

For example, a user has bought a lot of gardening tools, and you want to provide recommendations on the next one to buy based on a personalized service.

So this is the same technology used by Amazon.com.
When you go and shop on Amazon.com and after buying a few products, what you will see is that Amazon.com start recommending products in the same category or in completely different categories based on how you've been searching and how you've been buying and user interest and that kind of things.

So, you read your input data from Amazon S3, for example, it could be user interactions, those kind of things.

Also, you can use the Amazon Personalize API to have real-time data integration into the Amazon Personalize service, and then this will expose a customized personalized API for your websites and applications, your mobile applications.
Also, you can send SMS or emails for personalization as well.

You have all these integrations.
It takes days to build this model, not months, and you don't need to build, train, and deploy ML solutions, you can just use this bundled as is.

Some use cases is going to be retail stores, and media, and entertainment.

GitHub
LinkedIn
Facebook
Medium

Amazon Kendra Overview

عبدالله عياد | Abdullah Ayad — Wed, 28 Feb 2024 20:29:00 +0000

This one is a fully-managed document search service that is powered by machine learning and it allows you to extract answers from within a document.

That document could be text, PDF, HTML, PowerPoints, Microsoft Word, FAQs, etc.

So you have a lot of data sources where these documents may be like Google Drive, and they're going to be indexed by Amazon Kendra which is going to build internally a knowledge index powered by machine learning.

And how does it help from an end-user perspective?
Well, we get natural language search capabilities just like you go on Google.
So for example, if a user says, where is the IT support desk into Amazon Kendra?
Kendra can reply, 1st floor.
And this could be due to the fact that Kendra knows from all the resources that it took that the IT support desk was on the 1st floor, which is quite awesome.

And also, you can just do a normal search it will learn from the user interaction and feedback to promote preferred search results which is called incremental learning.

Finally, you can fine tune the search results, for example, based on the importance of data, the freshness, or whatever custom filters you have, okay?
Okay.

GitHub
LinkedIn
Facebook
Medium

Amazon Forecast Overview

عبدالله عياد | Abdullah Ayad — Thu, 22 Feb 2024 20:28:00 +0000

What is Amazon Forecast

It allows you to do forecasts.
So this is a fully managed services that will use machine learning to deliver a highly accurate forecast.
For example, you want to predict the future sales of a raincoat.

Some features of using it

The idea is that it's going to be 50% more accurate, than looking at the data itself, and you reduce forecasting time from months to hours by using a managed service.

use cases

For example, product demand planning, financial planning, resource planning, and so on.

how does it work?

Well, you take your historical time-series data, for example, and you also add your product features, prices, discounts, website traffic, store locations, basically, any kind of data you can to then enhance your model.
Then you upload this into Amazon S3.
You then start the Amazon Forecast service which will create a forecasting model.
And you can use that forecasting model, for example, to say that your future sales of raincoats are going to be $500,000 next year.

GitHub
LinkedIn
Facebook
Medium

Amazon SageMaker Overview

عبدالله عياد | Abdullah Ayad — Fri, 16 Feb 2024 20:31:00 +0000

SageMaker is a fully managed service for developers and data scientists to build machine learning model.

So all of the services we've seen so far like Amazon Comprehend or Amzon Rekognition are managed machine learning service with a very specific purpose, for example translates some text, transcribes some audio, or convert text into audio or analyze parts of a text, but SageMaker is a higher level machine learning service where you have your actual developers or your data scientists within your organization create and build machine learning model.

So it is a lot more involved and a lot more difficult to use.
Now, when you want to do this kind of processes to build a machine learning model, you have to do a bunch of steps that you have to follow, and all these are quite difficult to do in one place, plus you need to provision some servers to perform these competitions to create these models, and that can be cumbersome as well.

So this is where SageMaker comes in, it will try to help you all along the way for the process.

to build your ML model, you can use SageMaker to do the following steps:

Gather as much data as possible and label it.
Build a machine learning model, so, you have to train it and tune it.
Deploy your machine learning model
Apply the machine learning model that you have created on any new data coming in to predict the labels.

GitHub
LinkedIn
Facebook
Medium

Amazon Comprehend Medical Overview

عبدالله عياد | Abdullah Ayad — Sat, 10 Feb 2024 20:22:00 +0000

Comprehend Medical detects and returns useful information in unstructured clinical text.

For example, if your doctor is taking notes, or if you have discharge summaries, or test results, or case notes, then it will use NLP.

So it will use natural language processing to detect all the protected health information within the document itself as well with a DetectPHI API.

So from an architecture perspective you would store whatever documents you have in Amazon S3, and then you would invoke the Comprehend Medical API.

Or you could have Kinesis Data Firehose, and then analyze that in real time.

Or you can use Amazon Transcribe to first transcribe the voice into text.

Okay, as we understand from before, and then once it's in text form we can pass it to the Amazon Comprehend Medical service.

GitHub
LinkedIn
Facebook
Medium