Code Conversations

Deploying Generative AI applications at production scale demands careful attention to architecture and security, starting with the realization that large language models are entirely stateless and state must be constructed and passed through (e.g., via a database) to avoid losing conversation context and enable proper scaling. To achieve production readiness and control costs, developers should implement basic patterns like rate limiting for tokens and messages, restrict maximum payload size to prevent exhaustion attacks, and proactively utilize message analytics to monitor abuse and understand user behavior.

Ref: https://www.youtube.com/watch?v=hn2Dn3fLIfg&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=23

HTML is not just the foundation we build on, its vital in making our websites accessible usable and performant.We'll explore how we can make the most of our HTML elements and attributes to improve the performance and accessibility of our website and applications as well as boosting the efficiency of our development process.All by using a technology we are already use day to day, but just using it better than we were before. At the end of the talk you'll be able to make things more useful and useable, not just for performance and accessibility but for our users and different technologies, likes bots, applications and AI as well.

Ref: https://www.youtube.com/watch?v=zA4QzRGIP_w&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=22

Everyone has at some point wished they could clone themselves – to do the dishes, or work more efficiently. With advancements and improved accessibility of AI, this becomes more of a reality...This session will explore how to use Azure Custom Neural Voice to create a synthesised version of our own voice for use in a range of fun and practical applications – telephony, voice overs, or even attending meetings on your behalf...!We'll cover the basics of training and tweaking a model of our own voice, with minimal code, that can be used to turn text into natural-sounding speech.Advancing further, we'll learn some ways to train our model with enhanced training data. As an example we'll take a larger amount of audio, such as a podcast, use widely available tools and APIs to identify the individual speakers and the words being spoken – including highly technical language – to feed into our model.You will leave with the knowledge and inspiration to give "cloning" your voice a try!

Ref: https://www.youtube.com/watch?v=XkbvfSLO3yE&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=21

This session covers the ethical use of AI, detailing how to identify, understand, and proactively counter potential risks while sharing examples of impactful solutions built for the nonprofit and humanitarian sectors. The discussion emphasizes the necessity of implementing responsible AI principles, utilizing mitigation strategies like prompt engineering and grounding, and acknowledging that AI will always make mistakes.

Ref: https://www.youtube.com/watch?v=odWIkRcqEAU&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=20

Large Language Models (LLMs), including GPT, operate at their simplest level by attempting to produce a reasonable continuation of the text they are given, basing their predictions on patterns observed across a massive corpus of information like billions of web pages. Prompt engineering is an iterative process that employs various techniques—such as role prompting, few-shot learning, and Chain of Thought prompting—to increase the accuracy, reliability, and personalization of the output, which helps minimize uncertainty and build trust in the generative AI technology.

Ref: https://www.youtube.com/watch?v=1XrgOK-Ydl8&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=19

Recent advances in generative AI, exemplified by LLMs like Stable Diffusion and ChatGPT, have created significant industry hype. Generative AI involves creating new media (such as text or images) by analyzing massive datasets to deduce and mimic existing patterns, a process driven by probabilistic and stochastic modeling. While models like GPT can produce humanlike text, they operate as language prediction models rather than utilizing true reasoning (AGI), which means they often "stumble over facts," produce inconsistent results, and struggle with basic tasks like multiplication, leading to "hallucinations". To leverage these tools effectively, prompt engineering is necessary—this "subtle art" involves providing clear, specific instructions, setting a system context or persona, and potentially using examples to coax a useful result from the AI. When integrating AI via the stateless Completions API, developers must manually maintain conversation state by sending the entire history with each request, often summarizing older messages to manage token costs. More robust applications can utilize GPT Functions (Tools) to allow the model to intelligently call external functions—avoiding expensive model retraining—to access live or proprietary data. Alternatively, to query custom data using natural language, facts can be converted into high-dimensional vectors called embeddings and compared using cosine similarity against user queries, often managed in a database like Postgress with PG Vector. Finally, the newer Assistants API simplifies the development of domain-specific helpers by automatically managing message history and context compaction, and uniquely, when referencing uploaded knowledge files (like a lease document), it provides specific references or footnotes detailing where the answer was found.

Ref: https://www.youtube.com/watch?v=OxHw_u45h7M&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=18

Enterprise software development is currently facing immense executive pressure, driven by boards and CEOs demanding rapid innovation, especially utilizing AI, to increase productivity, save costs, and gain a competitive advantage, a significant shift from previous executive disinterest in issues like integration modernizations. Developers frequently encounter frustrations and delays, including the "age old disconnect" where their pursuit of new tools clashes with IT and Security's focus on uptime, reliability, and avoiding security breaches, leading to delays of months in realizing business value after writing working code. This inefficiency is exacerbated by the extensive time spent on environment setup, sometimes weeks, which the speaker suggests "should be illegal" given the cost of developer time. To address these challenges, modern tooling focuses on standardizing environments: GitHub Codespaces provides an ephemeral, standardized development environment (VS Code in the browser connected to backend compute) where mean time to onboard can be reduced to minutes, defined by a devcontainer.json file that specifies necessary dependencies; complementarily, Dev Box offers full, on-demand virtual machines based on team-defined templates, which are highly supported by security teams because they are built on existing tooling (like Windows 365 and InTune) and allow for customized security profiles that can exclude productivity apps (a common attack vector), helping one customer reduce environment setup time for .NET applications from over two weeks to three hours. Further accelerating delivery involves codifying Deployable Architectural Archetypes using the Azure Developer CLI (azd), a principle that dictates repositories contain source code alongside infrastructure-as-code (Terraform or Bicep) and CI/CD pipelines, ensuring critical elements like Key Vault and Azure Monitor are "baked in" from the start. This approach is key to engaging security, architecture, and infrastructure champions early, transforming them into innovation accelerators by incorporating their requirements into the blueprint templates, allowing deployment of full infrastructure and applications with simple commands like azd up. Ultimately, while tools like these, along with GitHub Co-pilot for Business (which securely boosts performance by 30% to 50%), help "grease the wheels," successful acceleration relies on ensuring the people and process are right, including having security champions on the team and deploying applications to secure, compliant Landing Zones.

Ref: https://www.youtube.com/watch?v=mTozV_eV4jQ&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=17

This presentation by Mandy Michael, a Staff Software Engineer and Google Developer Expert, makes a compelling case for using HTML meaningfully to improve web performance and accessibility, arguing that these concerns should be addressed from the foundational HTML stage, not left until the end of development. The core message is that HTML elements should be chosen based on the content type—similar to using types in TypeScript—rather than relying excessively on generic div elements, because correct semantic HTML creates a crucial Document Object Model (DOM) and Accessibility Tree that is consumed by assistive technologies and search engines. Michael provides practical HTML tips, such as favoring native element functionality (like using  instead of a styled ) for built-in features and using attributes like fetchpriority, loading="lazy", and resource hints (preload, preconnect) to effectively optimize resource loading and improve performance metrics like Cumulative Layout Shift (CLS) and Largest Contentful Paint (LCP).

https://youtu.be/zA4QzRGIP_w?list=TLGGh5u1cryB-zYwNzEwMjAyNQ

This speech synthesis service allows you to train your own model based on existing base models, utilizing a neural Voder to generate speech from text input. Crucially, Microsoft promotes responsible use: every custom voice must provide a talent statement (consent) that expires after 90 days.

https://youtu.be/XkbvfSLO3yE?list=TLGGiewzlim6miMwNzEwMjAyNQ

This talk was recorded at NDC Sydney in Sydney, Australia.

Ref: https://www.youtube.com/watch?v=odWIkRcqEAU&list=PL03Lrmd9CiGey6VY_mGu_N8uI10FrTtXZ&index=20