AI/ML Meetup Summary

Introduction

In May of 2024, at the headquarters of Solution Street, the NOVA Software Architecture Roundtable Meetup group held an AI/ML focused meetup. Illuminating the ever-evolving landscape of artificial intelligence and machine learning, this was the largest Meetup we’ve had since the pandemic, with over 40 individuals. Over the course of the evening, five distinct presentations offered compelling perspectives, practical applications, and thought-provoking ideas. 

Below are brief summaries of each talk. Enjoy!

The full meetup can be viewed here.

Who’s Who in the AI Zoo  – Ryan Ashcraft

In his presentation “Who’s Who in the AI Zoo,” Ryan Ashcraft discusses the various roles involved in AI projects and the challenges associated with role confusion. He emphasizes the importance of understanding these roles to improve resource engagement and expectation setting. Ryan draws on his extensive experience across numerous roles in the software industry, highlighting the benefits of having team members with versatile skills who can adapt to different project needs. However, he also points out the potential for stress and underperformance when roles and expectations are unclear.

Ryan uses personal anecdotes, such as his experience with his dog and a professional dog trainer, to illustrate the necessity of recognizing when specialized expertise is needed. He notes the difficulties in defining roles in AI due to the evolving nature of the field and differing interpretations of roles like data scientists and machine learning engineers. Ryan advocates for applying traditional software delivery principles to AI projects, ensuring clear role definitions and structured processes to manage the complexity and rapid changes inherent in AI development.

Just the Tip of the Iceberg… – Greg Hodum

In his presentation “Just the Tip of the Iceberg,” Greg Hodum provides an overview of using Azure AI Machine Learning Studio for training machine learning models. He highlights the platform’s capabilities, including data ingestion, Jupyter notebook support, drag-and-drop pipeline creation, and AutoML for automating the machine learning process. Greg demonstrates the machine learning process using the Titanic dataset, showcasing how AutoML rapidly tests various algorithms and hyperparameters to identify the best model.

Greg also emphasizes the importance of understanding domain-specific data to effectively analyze and clean it. He showcases the Azure platform’s ability to deploy models as REST endpoints for inference. He concludes by discussing the pros and cons of Azure AI Machine Learning Studio, noting its ease of use, scalability, and comprehensive MLOps features, while also acknowledging the potential for high costs and vendor lock-in.

Taming Python for Repeatable AI Delivery – Eric Konieczny

In “Taming Python for Repeatable AI Delivery,” Eric Konieczny addresses the challenges of using Python for AI/ML projects in large, multi-language (polyglot) enterprise systems. He critiques the frequent reliance on notebooks as a substitute for repeatable delivery lifecycle tasks, such as virtual environment management or automated test execution. As a result, teams often face challenges when operationalizing notebooks or even when simply reproducing notebook-generated results in a different environment. Further complicating the production delivery of AI/ML, monorepos are an increasingly popular approach to structure a polyglot codebase, but the build tooling that FAANG-class firms use and promote may not be appropriate for your project.

To address the gap, Eric introduces “Habushu,” an open source Maven plugin designed to enforce best practices in Python projects and blend Python, Java, and other Maven plugin supported frameworks (i.e., Node.js, Docker) into the same polyglot build at a lower maintenance cost than typical monorepo tooling. Habushu standardizes the execution of an opinionated set of lifecycle activities for Python projects that facilitate maintainability and consistency without sacrificing Python delivery velocity. Take a look at this link for more details!

A Sight to Behold: Using Glasses to Explore Embeddings with Open AI’s CLIP Model – Ryan Gehl

In his presentation, “A Sight to Behold: Using Glasses to Explore Embeddings with Open AI’s CLIP Model,” Ryan Gehl explores the practical applications of embeddings using OpenAI’s CLIP model. He demonstrates how embeddings can map content into multi-dimensional space, enabling powerful similarity searches and clustering. By calculating embeddings for 30,000 pairs of glasses, Ryan showcases how similar pairs of glasses can be identified based on their positions in embedding space.

Ryan further explains the use of principal component analysis (PCA) to visualize these embeddings, showing clustering of similar items. He also illustrates a blending technique to find intermediary styles between two pairs of glasses. Ryan closes by reminding the audience that embeddings, the fundamental technique which enables similarity search, is the same technology that underpins Retrieval Augmented Generation (RAG).

AI Hallucinations – Joel Nylund

Joel Nylund, CEO of Solution Street, presents on the topic of AI hallucinations—incorrect or misleading results generated by AI models that are presented as facts. The phenomenon, gaining prominence with the rise of generative AI in the 2020s, stems from issues like insufficient training data, overfitting, and the AI’s inability to understand idioms or adversarial prompts. These hallucinations pose significant challenges, especially in applications requiring high accuracy and reliability.

To mitigate the risks associated with AI hallucinations, Joel suggests treating AI/ML as “Human Assist” tools, emphasizing the importance of human review in critical applications. He also recommends using multiple AI models, combining AI with business rules, and employing Retrieval Augmented Generation (RAG) for better context. The presentation underscores the need for continuous improvement and careful implementation of AI technologies to ensure consistent and reliable outcomes in software development and other fields.

Solution Street was proud to host the AI/ML Meetup. The event was instructional, entertaining, inspiring and fostered networking opportunities within the local tech community. We look forward to hosting more tech meetups in the future and encourage everyone to join the NOVA Software Architecture Roundtable meetup group for updates and future opportunities for engagement.

At Solution Street we are AI/ML experts and have developed our own process to help our customers overcome internal barriers and get the most out of AI. If you are interested in hearing more about Solution Street and what we can offer our customers – especially in the ever-changing and expanding AI universe, reach out!