I’m Doron, a rising senior at Emory University from Berlin, Germany. At Emory, I’m pursuing a double major in Computer Science and Economics and am passionate about applying technology to solve real-world problems and create value for businesses and society. This summer, I’m thrilled to join Liberty Advisor Group as an AI Solutions intern, where I’ll have the opportunity to learn from experienced professionals and tackle challenging AI-related projects.
Before joining Liberty Advisor Group, I gained valuable experience in technology consulting and software engineering through various internships and research projects. During my freshman summer, I worked at ALDI NORD, a leading retail company in Germany, where I developed automation solutions for repetitive employee tasks using UiPath. The following summer I interned at Deloitte, a global consulting firm, where I was able to gain firsthand consulting experience and worked with large global financial clients.
Currently, I’m conducting research with Dr. Matt Lyle, a professor of Accounting at Emory University’s Goizueta Business School. His research involves combining machine learning and fundamental analysis for stock selection and portfolio construction.
At Liberty Adviser Group, my mission is to work closely with senior technical consultants on their efforts to explore AI offerings and capabilities.
- AI Blog #1 – Exploring Copilot Studio
- AI Blog #2: – Exploring the AI Landscape and the AI of Common Software Solutions
- AI Blog #3 – Creating a Copilot Studio model for internal use
- AI Blog #4 – Advancing the Copilot Studio Model and Exploring New Tools
- AI Blog #5 – Delving into GitHub Copilot, Perplexity AI and Amazon Q
- AI Blog #6 – Diving Deeper into Amazon Q and Presenting Copilot Studio to the C-Level
AI Blog #1
Exploring Copilot Studio
One of the projects I’m working on this summer is to explore the AI landscape and the use cases of Copilot Studio, a new tool that can enable organizations to build their own tailored Copilots. Copilot Studio is an innovation stemming from the relationship between Microsoft and OpenAI (the company behind Chat GPT). Copilot studio works similarly to Chat GPT models in interpreting user prompts but uses company specific data to enhance employee productivity by providing relevant responses.
As part of my internship, I will be sharing my findings and insights about how Copilot Studio can help clients across different industries and domains to enhance their productivity, efficiency, and innovation in a series of blog posts. In these blog posts, I will be sharing use cases, potential opportunities, and challenges that I’ve learned. I hope that you’ll find them informative and engaging, and that you’ll join me in this journey of discovering the potential and the pitfalls of Copilot Studio.
AI Blog #2
Exploring the AI Landscape and the AI of Common Software Solutions
This week at Liberty Advisor Group my primary task was to explore the AI landscape and start identifying the use cases of Microsoft Copilot Studio. Marketed as a solution for seamlessly integrating a chatbot with e.g. your company’s data, Copilot Studio aims to enable users to ask questions about company policies and retrieve relevant information. This integration is designed to enhance overall efficiency and improve knowledge accessibility within the organization.
As part of my first objective, I researched the AI of common solutions, such as ERP systems, chatbots, and recommendation systems, and how they are used in different industries and domains. For example, SAP recently introduced their new GenAI, Joule, which can answer analytical questions and enhance data modeling, among other features. The goal of this was to create an extensive knowledge base for internal use to enable a clear overview of the seemingly overwhelming range of AI products.
Regarding the second objective, navigating the Copilot Studio platform presented some initial challenges. As a computer science student, I found it difficult to adapt to the low-code/no-code environment that Copilot Studio offers. The platform’s interface and functionality differed significantly from the traditional coding environments I am accustomed to.
The platform includes an AI wizard designed to assist users in setting up and deploying chatbots. However, I am still uncertain about the AI wizard’s effectiveness at this stage and its true utility in simplifying the process. There have been several instances where the AI wizard either informed me that it couldn’t assist with a task (which I was able to complete manually), made changes that it didn’t actually implement, or, most frustratingly, claimed there were no issues in my process when there clearly were as running the process caused errors. Unfortunately, it didn’t provide any way to identify where the problems occurred.
It’s important to note that these challenges occurred within the first couple of days of using the tool. I’m confident that Copilot Studio has the potential to be a valuable product, but it certainly comes with a learning curve. I will continue to explore and utilize Copilot Studio in the coming weeks.
I hope you found this blog post informative and interesting, and I look forward to hearing your feedback and comments.
AI Blog #3
Creating a Copilot Studio model for internal use
This week, I continued working on my project of creating a Copilot studio model for internal use at Liberty Advisor Group. My goal is to build a model that can answer questions and provide relevant documents based on the knowledge that I feed it. I learned a lot this week through a variety of challenges, focusing on exploring the capabilities and how to add knowledge effectively.
To start, I added 10-15 test documents to see if the model could function correctly before connecting it with all the data. However, the AI struggled significantly, often hallucinating and failing to understand which document to use in specific situations. To address this, I created manual ‘topics’ that specified which triggers should retrieve which documents. This approach minimized hallucinations and greatly improved the model’s accuracy.
Despite these improvements, I encountered two main unsolvable issues. One of the documents I uploaded was an Excel sheet containing a list of all employees’ birthdays. This would ideally allow the Copilot to respond to prompts like, “Who has a birthday this month?” by returning a list of names and dates for the current month. However, the first issue is that Copilot has no understanding of time, making this prompt not possible. Adding a source of knowledge for the world clock didn’t seem to work, and when attempted, it parsed the time incorrectly. The second issue was that even when the AI correctly parsed the document, it still omitted some information. For example, when asked, “Who has a birthday in June?” it correctly understood the query but only listed three names instead of four.
On a positive note, I became much more familiar with the software and discovered a feature called ‘actions’. As the name suggests, actions are tasks that the Copilot can perform. There are some library actions available, but there’s also the option to connect with Microsoft Power Automate, a tool used to automate workflows. The idea is to enable the Copilot to perform tasks such as scheduling appointments without having to manually check the calendar for available times. The Copilot would call the action created in Power Automate and execute the ‘flow’.
Overall, this week, I successfully set up environments on both platforms, made significant progress in developing the Copilot, and began familiarizing myself with Power Automate.
In the next week, I will attempt building such flows. I am excited about the progress and look forward to refining the model further.
I hope you found this blog post informative and interesting. I look forward to hearing your feedback and comments.
AI Blog #4
Advancing the Copilot Studio Model and Exploring New Tools
This week, I continued developing the Copilot Studio model for internal use at Liberty Advisor Group. Building on the progress from last week, my focus was on using Microsoft Power Automate to create actions for scheduling appointments. The idea is to enable Copilot to perform tasks such as scheduling appointments without manually checking the calendar for available times. The Copilot would call the action created in Power Automate and execute the ‘flow’.
Despite its potential, Power Automate presented several challenges. The AI wizard within Power Automate, much like the one in Copilot Studio, proved to be relatively unhelpful. Initially, I struggled with the software, encountering a bug that prevented testing, but the AI wizard said there was no bug, and my flow should work. Additionally, I found the user interface to be less intuitive than expected. Navigating through the features and setting up actions had a steep learning curve and I often faced frequent setbacks followed by periods of progress.
When I faced these bugs, I sometimes discovered that I was the first to encounter them, leading to suggestions that I create new posts on the Microsoft Community forums. This was somewhat frustrating, as it highlighted the newness of the tool and the lack of robust support and troubleshooting resources. The absence of comprehensive video documentation and being a first-time user further compounded the difficulty, resulting in a lot of trial and error. I spent considerable time experimenting with different configurations and settings to achieve the desired functionality.
However, after overcoming these hurdles, I successfully managed to schedule appointments through Copilot Studio. This achievement was significant, as it showcased the potential of Copilot Studio and Power Automate to streamline and automate routine tasks within the organization. This success highlights just one of the many potential use cases, demonstrating the opportunity to save a large amount of time and enhance efficiency by automating workflows.
Overall, Copilot Studio is still in its early phases but shows promise. The platform requires more work than initially advertised, as all knowledge paths must be explicitly defined. This requirement can be quite labor-intensive, especially for complex workflows. Additionally, being a new product, Copilot Studio suffers from limited documentation. The lack of detailed guides and tutorials means that users must often figure things out on their own, which can be time-consuming and frustrating.
Despite these challenges, I believe that Copilot Studio has the potential to be a useful tool. The ability to integrate chatbots with company data and automate workflows can enhance efficiency and productivity within the organization. However, it is important to manage expectations and recognize that achieving these benefits requires a significant investment of time and effort.
With the testing of Copilot Studio now complete, I will shift my focus to testing GitHub Copilot and Amazon Q, Amazon’s GenAI product. GitHub Copilot is designed to assist with coding by suggesting lines of code and completing functions, which could be useful for developers. Amazon Q, on the other hand, aims to provide advanced analytical capabilities, answering complex business questions using natural language processing.
I am excited about the next phase of my project and look forward to exploring these new tools. Each of these products offers unique capabilities, and I am eager to see how they can be leveraged to further enhance productivity and innovation at Liberty Advisor Group and our clients.
I hope you found this blog post informative and interesting. I look forward to hearing your feedback and comments.
AI Blog #5
Delving into GitHub Copilot, Perplexity AI and Amazon Q
This week, I spent most of my time working with GitHub Copilot, Perplexity AI and briefly Amazon Q business. GitHub Copilot is an AI-powered code assistant that offers a range of functionalities including conversational generative AI, code suggestions, and error fixing. I worked with the individual version of GitHub Copilot which was easy to integrate into my coding environment (VS code).
One of the features that stood out to me was GitHub Copilot’s ability to suggest code snippets as I typed. This capability proved to be useful for repetitive or simple coding tasks, allowing me to focus on more complex aspects of the project. For example, I could comment, “create an array (collection of objects) with all capitals of Europe” and instead of having to look it up and write it out manually, Copilot generated it for me. Additionally, GitHub Copilot could complete entire functions based on my comments with an understanding of my previous code, saving me a significant amount of time. The tool’s context-aware assistance ensured that its suggestions were relevant and accurate most of the time.
However, it is important to note the limitations of Copilot as well. For more complicated or lengthy coding projects, Copilot is more likely to introduce new errors. Microsoft claims that GitHub Copilot is powered by GPT-4.0, but interestingly, I noticed that on some occasions, it seemed to use GPT-3.5 instead of GPT-4.0 for code suggestions. This discrepancy led to inaccuracies that could have been resolved using ChatGPT instead. Despite these challenges, the tool significantly accelerates development and is something I will continue using when writing code.
Before diving into Amazon Q, I also spent some time testing Perplexity AI, an AI-powered research tool designed to provide reliable information on any topic. Perplexity AI surprised me with its ability to cite sources accurately and generally refrains from confidently providing misinformation. Additionally, Perplexity AI offers further prompts to extend your search, guiding your research by suggesting related questions to explore next. Another useful feature is its ”focus” mode, which allows you to select specific sources for Perplexity AI to search from. Compared to other AI tools I’ve used, Perplexity AI stood out for its ability to rapidly retrieve information and provide data to the user in a highly structured manner, providing a more reliable way to conduct research.
After testing Perplexity AI, I shifted my focus to Amazon Q, Amazon’s new GenAI product. Amazon Q business primarily aims to answer user questions using company data, similar to Copilot Studio. However, the experience is integrated solely into Amazon Web Services (AWS) and cannot be used without it. This week, my goal was to become familiar with the AWS interface and understand how Amazon Q can simplify workflows and provide actionable insights for different companies. My initial exploration of Amazon Q has been promising, and I am eager to delve deeper into its features next week.
Overall, this week has been productive and filled with valuable learning experiences. GitHub Copilot has demonstrated how AI can streamline coding tasks, while Perplexity AI offers a more reliable way to conduct research. I am excited to see how these products can drive innovation at Liberty Advisor Group and our clients.
I look forward to sharing more insights and discoveries in my next blog post. As always, I welcome your feedback and comments.
AI Blog #6
Diving Deeper into Amazon Q and Presenting Copilot Studio to the C-Level
This week has been truly exciting. My main tasks involved working with Amazon Q and presenting Copilot Studio to C-level executives of a multi-billion-dollar company.
This week, I connected Amazon Q with approximately 50 documents and a handful of websites. The process was seamless, with Amazon Q indexing the items by itself efficiently. This automated indexing feature significantly simplifies data management, making it easier to retrieve relevant information quickly in contrast to Copilot Studio, where you have to construct the conversation flow manually.
Once you have a foundational understanding of AWS, Amazon Q proves to be an easy-to-use application. My goal for the coming weeks is to connect Amazon Q to all company files and various internally used applications to test its enterprise readiness. By doing this, I aim to evaluate how well Amazon Q performs on a large scale, dealing with thousands of documents. Currently, I am in the process of obtaining the necessary permissions as admin status is required for these connections.
Another major highlight of the week was presenting Copilot Studio to C-level executives. During the presentation, we demoed Copilot Studio and showcased its potential to enhance productivity.
We addressed various company-specific use cases and answered potential questions from participants. Demonstrating the complexity of implementing Copilot Studio and how it can be tailored to meet specific business needs was a key focus.
The main goal for next week will be setting up Amazon Q with enterprise-wide data. This involves connecting it to all relevant company files and applications to test its performance on a large scale.
Overall, this week has been incredibly productive and filled with valuable insights. The hands-on experience with Amazon Q and the opportunity to present Copilot Studio to top executives have been significant milestones in my internship with Liberty Advisor Group.
