AI Powered event-driven translation bot

Editor's Notes

• #2: Building a serverless AI powered stranslation bot, or service. This talk is actually based on a solution that I have built for my tech blog. I have created a serverless AI pipeline that does some really nice things. We will go in a bit further on that later. Todays session will NOT be slide heavy, instead I will focus a lot more on building here on stage.
• #3: AI is basically all around us these days, from translastions, to text to speech, personlized ads and so many other use cases. With generative AI getting a major boost the interest for AI has also skyrocketed. When I submitted this talk genertive AI was already a hot topic, but during the last 8– 10 months since I submitted this it has litterly exploded in interest, with so many services being released. Today however, we will not be speaking about Generative AI, we’ll focus on some traditional AI service from AWS and we’ll be focusing on building something. All images in this presentation has been made with Genartive AI and Dall-e
• #8: What is generative AI? The answer actually lies in that first word “Generative” it can create new content, everything from music, text, stories, images and much much more. Like all AI it’s powered by ML models, and in this case very large models refered to as ”Foundation models”…. https://youtu.be/5EDOTtYmkmI?si=EJOUbBetEM-wy9P6&t=169
• #9: Foundation models are trained huge amount of data and is the latest advancement in Machine Learning, but like with everything else they are not magic. With RAG “Retreival Augmented Generation” you can use the foundation model on you domain specific data, with that you could build your own highly advances support chat bot. Imagine being an airline and where you can train a model on your data and then let your customers book, cancel, rebook their trips in a very efficienr manner that normal chat bots today cant. The use for Generative AI great and it can be in so many places, I think we are now on a very interesting journey where we have just started to take off.
• #10: Amazon has their own Service, Bedrock, around Generative AI where you only pay for what you use. It come with support for several foundation models from different vendors, like Hugging face, but also their own model Titan. It’s possible to customise bedrock on your own organisations data, which is very important I think. Enough about generative AI. Today we are going to build things with good old fashin traditional AI.
• #12: So what is the difference? As models in Generative AI can do many things, the models used in traditional AI rae tailored towards one single use case, like translations or text to speech. They are expert systems with a narrow focus. The tarditional models can also be faster and less expensive than the huge generative versions.
• #13: Modern AI services leverage a diverse range of machine learning models, each tailored for specific types of tasks. In supervised learning, models are trained on labeled data, with classic models like Linear Regression and more advanced ones like Neural Networks. Unsupervised learning, on the other hand, finds hidden patterns in data, with techniques like K-Means Clustering. Reinforcement learning revolves around agents learning by interacting with their environment, as seen in Q-learning. Deep Learning, a subset of ML, employs deep neural structures for complex tasks, with models like CNNs popular for image recognition. "
• #16: This is an image from AWS that show their different layers from the ML frameworks and infrastructure to the specialized AI services in the top. Where we have AI service for different use-cases, like image recognition, text translations, text to speech and many many more. In this picture we are missing the bits an pices about generative AI, as Bedrock was release just a few weeks ago.
• #17: What I think is very good about the services that AWS offer is that you don’t need to be an expert in AI to be using AI. You can use the services that AWS provide to build som really interesting and cool things. And if you combine Generative AI with these traditional services you can build something really powerful. One good example of this is from a blog post from AWS, where you use Generative AI to generate images, and then use AWS image services like recognize to moderate the generated image.
• #24: Now!! What are we building ??? This is what this session is all about! Building something!
• #25: Arch overview
• #28: But before we do any changes, just let us look at what an event can be defined as
• #30: Amazon Translate leverages the latest in neural machine translation technology to offer precise and fluent translations. Whether you need instantaneous translations or bulk text conversions, Amazon Translate caters to various demands. And with support for over 70 languages and their variants, its reach is truly global.
• #31: he secret sauce of Amazon Translate is its neural network architecture. By utilizing sequence-to-sequence models, it maintains context throughout translations. The attention mechanisms further fine-tune this by weighing the significance of each word or phrase. And, with continuous training, the service keeps evolving for better accuracy
• #32: Amazon Polly is AWS's premier Text-to-Speech service. At its core is a deep learning model, trained to transform text into lifelike speech. With a variety of voice options and broad language support, Polly paves the way for myriad applications, from audiobooks to virtual assistants.
• #33: The neural foundation of Amazon Polly ensures the speech it generates isn't robotic or artificial. By mimicking human neural patterns, it processes text in real-time and emphasizes prosody—the rhythm, stress, and intonation of speech. This is crucial to produce speech that sounds natural and relatable. Polly offers a wealth of advanced features tailored to developers' needs. With Speech Marks, one can synchronize speech output with visual animations or displays. Lexicons allow customization of pronunciations to align with specific requirements. And, as an AWS cornerstone, Polly seamlessly meshes with other services, be it storing audio files in S3 or leveraging Lambda for serverless computation.
• #34: Amazon comprehend is a Natural Language Service (NLP). That will help you find insights and relation ships in any unstrcutured data It requires no ML expertise, it’s just a set of APIs that you can use easily,
• #35: Comprehend is robust and flexible. It comes equipped with a suite of pre-trained models for tasks like sentiment analysis or entity recognition. But businesses aren't limited to these: Comprehend allows for custom models tailored to specific needs. And, as expected of an AWS service, it integrates smoothly with the broader AWS ecosystem for comprehensive data solutions." When designing the presentation, ensure consistent visual themes and high-quality visuals to maximize engagement and understanding. And there is no servers to provision