<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Padraig Moran on Medium]]></title>
        <description><![CDATA[Stories by Padraig Moran on Medium]]></description>
        <link>https://medium.com/@padraig_36878?source=rss-377f5f6682cb------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*h3oOlyTkWWm4DXsiL4PNYg@2x.jpeg</url>
            <title>Stories by Padraig Moran on Medium</title>
            <link>https://medium.com/@padraig_36878?source=rss-377f5f6682cb------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 07:40:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@padraig_36878/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Now It’s Getting Serious! — A Practical Assessment of 100% AI based Software Development]]></title>
            <link>https://medium.com/@padraig_36878/now-its-getting-serious-a-practical-assessment-of-100-ai-based-software-development-43d164a3692f?source=rss-377f5f6682cb------2</link>
            <guid isPermaLink="false">https://medium.com/p/43d164a3692f</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[development]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Padraig Moran]]></dc:creator>
            <pubDate>Wed, 28 May 2025 08:45:15 GMT</pubDate>
            <atom:updated>2025-05-28T08:45:15.030Z</atom:updated>
            <content:encoded><![CDATA[<h3>Now It’s Getting Serious! — A Practical Assessment of 100% AI based Software Development</h3><h3>Introduction</h3><p>The recent developments in generative AI, with the easy accessibility and broad applicability has been creating new opportunities all over the place. This is enhanced with improvements coming virtually every week, in terms of capability, usability and availability. The whole area of support for developers, with tools like Github CoPilot plugging into VS Code, has helped make developers more productive by removing mundane tasks, providing boilerplate code, solutions to problems and finding and fixing bugs that arise. In most situations, it has required developers to be involved, to design the solution, and insert code and assess the appropriateness of the solutions provided, and guide them forward, based on some software development knowledge.</p><p>Continually we read expert opinions saying that software developers will be replaced by AI, and we’re nearly there. While I think there is a role for a software engineer going forward, to design the architectures and solutions, and critically assess the suitability of any automatically generated solution to the target requirements, I have been interested to understand how effective the current batch of tools are.</p><p>With this in mind, I saw that Github CoPilot with VS Code Insider Edition was supporting agentic code creation with a range of LLM models, moving the coding support to a new level. No longer was the developer needed to create things, but allow CoPilot do it’s thing, creating files, code, making changes, testing things, following up on errors, etc.</p><p>To dig further, I decided to set the following challenge for Agentic CoPilot:</p><ul><li>Develop a fully functional application, with a clear purpose, which is tested, working and can be deployed as-is on a target platform.</li><li>Do this without writing a single line of code myself.</li><li>I only interact with VS Code, via CoPilot prompts, should I want improvements, or resolution of error messages arising and where relevant accept recommendations on changes to be made by CoPilot, and let it do it all.</li></ul><p>I started with a single clear application which I need myself, but as you will see, the success with that drove me to see what else it could do, and where the limitations were… Read on!</p><h3>Introducing the ‘PDF Joiner’</h3><p>The first application was a very practical one. I want a tool which will provide a simple mechanism for a user to take a directory of PDF files and join them together, creating a new PDF. There are loads of services online which do this, but I wanted something simple and local which I could use on my Mac.</p><p>I started VS Code Insider and selected the Agent mode and Claude 3.7 Sonnet as the model I wanted to use. There was no specific reason for this, but previous experiences of using it were positive.</p><p>Then off I went …</p><p>The initial prompt I provided was the following, setting the general parameters for the development, in terms of python tools I wanted it to use. Being specific on what to use was primarily so I could understand the produced code, and critique it.</p><pre>User:<br>You are an expert python programmer who likes to create innovative solutions<br>where you leverage streamlit for GUIs, FastAPI for APIs and provide code <br>which is well documented, includes extensive error testing, logging and <br>implements automated testing.<br><br>Create a program which will take a number of pdfs which are located in <br>a specified direction, indicated in the GUI. It will join these pdf documents<br>together into a new pdf document that combines them. The name for the new one<br>is specied in the GUI. When the input directory, and output name is specified,<br>pressing a button will support the creation of the combined document.</pre><p>With this provided, off it went on its merry way. After a few minutes thinking, it came back with a list of things it would do, and proceeded to create a directory structure, files, code, configuration files, etc. It created a test framework using pytest and implemented a batch of tests to run. With the code written, It asked me if I wanted to run the tests, which I did, and it proceeded to find a few errors. It went through a number of cycles of testing and fixing errors until it was working well.</p><p>The GUI it created initially was based on the Streamlit web UI framework which was following my original instructions, however CoPilot had challenges getting it to present a good way of selecting the directory for the input files, and the output file, and despite many efforts, which I tested, and commented on, it really was not having much success in creating a good solution.</p><p>Eventually it came back with the following recommendation, as it felt I needed to have a more standalone approach to get the best user experience:</p><pre>CoPilot:<br>Alternative Solution: Creating a Truly Standalone Application<br>- For a more user-friendly experience that doesn&#39;t require users to install <br>Python manually, you could:<br>• Use PyInstaller instead of py2app: PyInstaller can bundle Python and all <br>  dependencies into a truly standalone executable.<br>• Bundle a complete Python environment: Here&#39;s how you could modify your <br>  build approach:<br>• Use a proper native GUI framework: Consider switching from Streamlit <br>  (which runs in a web browser) to a native GUI framework like PyQt or <br>  Tkinter for a more integrated macOS experience.</pre><p>Based on this, I then chose option 3:</p><pre>User:<br>Can you update the solution to use PyQt instead of streamlit?</pre><p>With this in hand, it proceeded to change everything, on on the frontend, and came up with a new GUI which was really impressive, and user friendly. I don’t think I needed to change anything with it, as it was that good! As it was still a python based application, I asked CoPilot to package it into a MacOS application, and then ship it in a DMG file so I could easily distribute it.</p><p>This is where it seemed to have challenges, having gone through numerous attempts at getting things to work, putting in debug code, simpler approaches, different approaches, and more, but really could not get it to work effectively. It was creating the MacOS application, but had problems with getting python to work correctly or including libraries at the right location. In fact, it came back with alternative recommendations, and tried these, but still did not succeed.</p><p>CoPilot seemed to be realising that it was having problems too, so provided the following option.</p><pre>CoPilot:<br>Alternative: PySide2 Instead of PyQt5<br>PySide2 is another Python binding for the Qt framework and is often easier <br>to package with PyInstaller on macOS. Let&#39;s update your requirements and <br>build script to use PySide2 instead.</pre><p>I was on the verge of giving up, when I thought that maybe another LLM would do better, so I switched to OpenAI o4-mini and posed the packaging question to it. After it did a survey of what was in place already, it took a completely different approach and got an application built and executable on the first attempt.</p><p>It then updated all the documentation in the README file with architectural details, class details, user guidance with everything I needed.</p><p>All seemed to work fine on my Mac, so I took the DMG it created, and moved it to my wife’s Mac, and dragged the application to the Applications folder, and bingo — worked like a charm, with a cool icon it created for the application.</p><p>While there was some interaction with CoPilot during the process, it was really to recommend improvements, or slight changes and otherwise it was pretty straight forward. The whole process took about 2 hours total, from start to finish. I think that my knowledge of how it works, providing input to it related to bugs, or issues I saw, did make things simpler, and I’m not sure if a total novice would not have had challenges, but I did not write any code, and I got what I wanted, a tested, working, packaged solution to a well defined problem.</p><p>As you can see in the screenshot of the installed application, it has a very user friendly and clear interface.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/1*FExyISapwGfAZmkEe9_jaQ.png" /><figcaption>Figure 1- Native application for joining PDFs to create a new file.</figcaption></figure><p>This went way better than I expected.</p><p>Armed with that success, and after a chat with my university student son interested in managing his monthly budget, I decided to challenge it again with a new use-case. Could it be as good?</p><h3>Personal Budgeting Application</h3><p>The intention with this application is to provide a website (local or remotely accessible) allowing the creation of a personal budget, including income, fixed and variable costs amongst others, to allow basic control of monthly personal finances. Providing budgeted figures for each month, combined with actual figures for the month, the system should create graphs, offer comparisons and provides a simple way of moving figures from month to month, without needing to start from scratch.</p><p>To provide some more meat to this application, and see if it could handle some more complexity, I decided that I wanted a frontend/backend solution, with the data to be stored in a MySQL database. A REST API would be provided to a backend which would manage the data. The frontend would use the Streamlit framework and provide a web interface, with a simple and graphical interaction approach.</p><p>I went back to using Claude Sonnet 3.7 again, as it was pretty good the first time out. I started with the following prompt, and applied some minor adjustments and evolution during the session. As you can see I provided a lot more detail for this project, and it generally helped, as the initial shot at delivery was better. There were of course some issues which needed to be addressed, and some bugs which arose, which I fed back to it and asked it to fix.</p><pre>User:<br>You are a highly experienced python software developer in front- back-end <br>projects. You are experienced at developing API-first business applications, <br>with user friendly user interfaces, and simple clear work flows. Automated <br>testing is very important to you, and you ensure that as much functionality <br>as possible can be automatically tested as part of the development process.<br><br>Create a personal budgetting application, which provides a web GUI where a <br>person can enter, update and view their budget parameters and details for a <br>month. For each month, the user can create a Budget and then, with a copy of <br>this budget, update it during the month with the actual costs . <br>Having both allows the user to assess how good their budgeting was, and <br>how different the reality was.<br><br>The tool shall allow the creation of a monthly budget for a month a year. <br>The user can enter and edit budget items such as:<br>-- Primary Income: Monthly Gross Amount (pre-tax) and the amount paid in <br>   taxation<br>-- Other Incomes: Monthly Net Amount (after tax) received.<br>-- Fixed expenses: monthly amounts which are are regular, such as rent, <br>   loan repayments, etc. These typically don&#39;t change from month to month.<br>-- Variable expenses: These can be infrequent costs, or annual payments, like <br>   gym subscription, or vacation, etc. These will likely be different from <br>   month to month.<br>-- Saving Plans: Indicate amount that the user intends to save that month.<br><br>I shall refer to this stored information for a month as the Budget Profile.<br>When a budget is created for a new month, it is related to a month and year <br>- e.g. April, 2025.<br>The budget will, by default include the budget items, as shown above, as the <br>starting template values, and these can be updated or removed with additions <br>provided as needed. For each monthly budget created, there will be a similar <br>&#39;Actual Profile&#39; for the month, which will have the same information initially <br>as the Budget Profile for that month. This can be updated based on the actual <br>expenses by the user during or after the month to reflect the real income and <br>costs. Having the actuals for each month allows better analysis to see if <br>the budget was met, and get a perspective over time how accurate the <br>budgeting process is.<br><br>Based on these inputs above for the budget profile, a monthly budget for <br>the month shall be displayed in a user friendly format. <br>A bar chart will also be shown, showing what amount of the income is taken <br>by tax, fixed and variable costs, savings etc.<br><br>The GUI shall provide a menu allowing the user to: <br>-- edit the values  associated with the budget profile as well as the actual <br>   profile,<br>-- calculate the budget for this month, <br>-- view the history of the previous months, with a line chart including the <br>   same figures as provided in the barchart, but for each month being displayed. <br>   This history shall support providing graphs for the budget vs actual <br>   profiles for one or multiple months.<br>-- Provide a dashboard with a clear up to date view of the budget, actual, etc.<br><br>Some system-wide settings should be possible to configure, such as:<br>-- Currency which is shown - default is SEK<br>-- Name of the User to be displayed in titles for reports<br>-- Exporting the budget and actual profiles for one or many months <br>   in CSV or Excel format.<br><br>Implement the solution with the following technology:<br>-- The solution is implemented in python, with a frontend backend approach.<br>-- The frontend shall use streamlit for the GUI.<br>-- A FastAPI REST API shall be provided by the backend and used by the <br>   frontend. All functionality shall be accessible via the API, if needed and <br>   versioning of the API shall be supported.<br>-- MySQL shall be used for storing the data. The credentials for accessing the<br>   database are username: root, password: pass1234<br>-- Automated test cases shall be included which provide good coverage of the <br>   functionality.<br>-- Extensive commenting of the code, and error checking shall be included in <br>   the code.<br>-- Best practice associated with frontend-backend python development shall be <br>   adhered to, as far as possible.<br>-- A focus on a user friendly UX is important, with help information provided <br>   where needed. The UI should be clear and logical to start using without <br>   guidance. It should be quite graphical, in terms of charts, graphs and <br>   other pictorial mechanisms of presenting data, comparisons and results.<br>-- The solution shall provide appropriate configuration files to be deployed <br>   in Docker, with clear instructions on how this is to be done, and created <br>   docker compose files or other configuration files provided.<br><br>You shall create the code, and all appropriate files and documentation.<br>You shall install the system, incl. Creation of the database<br>You shall run and correct the automated test cases until all working correctly.<br>You shall run the final solution.</pre><p>The process took, again, 2–3 hours, and resulted in a really effective application with 40+ test cases created and executed successfully by CoPilot along the way. It provided a frontend and backend solution, as requested, with FastAPI providing the REST API to the backend. It also provided the docker-compose file to allow me to deploy frontend, backend and database in containers.</p><p>At the end, I realised I would like to have a clear view on the API, so asked it to create swagger documentation for the API. This it promptly did, updating the various files, and providing me with the URL for version1 of the API, providing the following details:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r-l7mfx8NRoExrNClwYr5A.png" /><figcaption>Figure 2 - REST API provided by the backend</figcaption></figure><p>To run the application in the local environment, I need to run the backend server which created the database schema if not existing and provided the API URL. I then ran the frontend UI using streamlit which provided a URL to get access to all the functionality.</p><p>As shown in the screenshots below, the user is presented with a simple, accessible UI to create and update the various parameters associated with a budget and the actual figures for any month.</p><p>With these in place, comparisons are displayed between budget and actual for any month, and it is possible to track the overall changes, over time, for actuals or budgets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C_NWuLYNjsvx-4M9nILuxw.png" /><figcaption>Figure 3 — Editing the parameters making up the budget</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IcAL1Rn_4w2eEeTOpwu0bQ.png" /><figcaption>Figure 4 - Comparison between budget &amp; actual for a selected month</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BmWWjGhzBbszZh_mSI-gIA.png" /><figcaption>Figure 5 — Tracking figures for budget or actuals over time.</figcaption></figure><p>Some design aspects were really good, that I would have likely taken time to come around to. For example, the representation of all stored values being the same, just being categorised differently, was great, and would make it very simple to extend with more types of expenses or incomes. How these categories are processed determines what they are. On GUI side, the possibility to add or remove parameters from the graph is great and very user friendly, but would definitely have taken many revisions, and likely frustrated users, to come up with it myself.</p><p>This deploys in docker too, with the docker-compose file provided. This substantially simplifies the distribution of the solution, with docker being the only target system requirement, hiding all the implementation and import library details.</p><p>OK, so CoPilot Agent can do native apps which work, and now can do FE-BE apps with databases and REST APIs deployable in a containerised environment.</p><p>It’s good, and clearly provides some real value which even at this stages seems to be valuable for non programmers, or does a huge amount of work to simplify and speed up development for even minimally experienced programmer.</p><p>Let’s try one final use-case and throw something more ‘in-vogue’ at it, and see how it behaves.</p><h3>Stock Investment Assessment System</h3><p>I have one more idea. Can the CoPilot Agent create an agentic solution to leverage AI to predict stock prices taking into account a company details and relevant news feeds. Can it make recommendations on selling or holding and provide up to date and clear reasoning behind its decision. Of course, I want it to be simple to use, graphical and visual pleasing. And of course, I would like it to create printable reports which I can refer to later. This sounds complex!</p><p>This will make calls to external systems, such as OpenAI for the LLM as well as some sites to get stock price and information as well as business news sites to get access to related company news to use in the process. These may require IDs and/or subscriptions, if I am accessing the services over APIs, so I am aware of this, should I need to consider this further on.</p><pre>User:<br>You are an experienced python developer with experience of implementing GenAI <br>and specifically multi agent approach to problem resolution. <br>In your projects, you focus on implementing APIs where appropriate, <br>implementing automated testing for all functionality, including clear <br>comments and documentation, including error checking and data validation <br>for entered values or parameters. You provide solutions which are deployable <br>directly on a machine (MacOS) but can be deployed in a docker environment. <br>In relation to GenAI, you are comfortable leveraging the latest technique <br>such as MCP or RAG where appropriate. Where a UI is needed, you implement <br>this using Streamlit as the framework.<br><br>You will create an agentic solution to address the following requirements <br>associated with an investment assessment solution. The solution shall do the <br>following:<br>-- Provide a GUI to allow the user to enter a stock ticker for a company <br>   which will be used for the analysis.<br>-- Run an agentic workflow to address the following aspects:<br>   -- Search the internet for the latest stock price, and get the historic <br>      stock price andn volume of trade from the past year.<br>   -- Search the internet for the latest news associated which canhelp <br>      guide on investment decisions. <br>      This information should include revenue and profit figures, cashflow, <br>      debt, sales projections or warnings etc.<br>-- Analyse all the input and create graph of the performance of the company <br>   over the past year, with some prediction for the coming 6 months on the <br>   stock price. Some indication on the level of certainty should be provided, <br>   and some bullet points explaining the reasoning behind the forecast.<br>-- Present a report on the company, including a summary of the company, <br>   including the key figures, business, etc. Include in this a risk assessment <br>   of the company, and a recommendation on whether one should buy, hold or <br>   sell the stock. Where it is buy or sell, the recommended target price to <br>   execute this transaction at should be indicated.<br>-- The reporting, including the graph, shall be saved as a PDF or a location <br>   that the user selects.<br>-- The details of the report, including the graph shall be presented on <br>   the GUI in a clear and informative manner.<br><br>A left-hand menu should be provided to allow:<br>-- Running a search<br>-- List and download previously created reports<br><br>OpenAI shall be used with GPT-04-mini</pre><p>This was again Claude Sonett 3.7 in action, and it seemed to do a pretty well. I subsequently asked it to provide a technical architecture document of the solution, and it produced, amongst lots of other good info the following diagram which reflects the agents and tools implemented:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2sIgrQ8KiN32zzaXcN0siw.png" /><figcaption>Figure 6 — Architecture implemented by CoPilot</figcaption></figure><p>The main agent classes are:</p><ul><li>Provides common functionality for all agents, including OpenAI API interaction.</li><li>StockDataAgent: Retrieves stock market data, historical prices, and related financial information.</li><li>NewsAgent: Gathers and analyses news articles related to a company.</li><li>AnalysisAgent: Analyses stock performance and generates price predictions.</li><li>ReportAgent: Creates visualisations and generates PDF reports.</li><li>Orchestrator: Coordinates the workflow between all agents and manages the overall assessment process.</li></ul><p>As I mentioned above, I realised that I would need to interact with services to get up to date information, so used the Yahoo Finance API and a site called Marketaux. These both have APIs, and I needed to provide keys, which I asked CoPilot to configure in the system. This all worked fine, however, I decided I needed more news information, so provided it with my Google API key and it issued search request to get more extensive financial news via Google. In this case, it added this source on, as another one with Marketaux, and provided all the output to OpenAI to assess it as part of its recommendation process.</p><p>With this done, and the test cases working after some problems which CoPilot identified and resolved, I ran the solution. It provides a local website allowing the entry of the ticker or company name. It then goes through the various agents, co-ordinated by the orchestrator, and eventually presents a stock report and predictions. It gives a recommendation on sell, buy or hold for the stock chosen and sell or buy a target price to aim for to carry out the action. It provides clear motivations for its recommendation based on the combined figures and news information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w8KcZIvqVynjAeQAh8DmpA.png" /><figcaption>Figure 7 - AI generated report on a company</figcaption></figure><p>It offers the possibility to save the specific report to a pdf file which is more distributable and printable.</p><p>On the left hand menu, it also provides an option to list all previous reports that have been created and these can be read online or downloaded. These each contain the graphs, news items read, recommendations, and lots of financial documentation on the company itself. It is pretty comprehensive. Whether it is correct or not is a completely different matter … to be seen!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J2Al7WUkNAK2RxMpkK6UCg.png" /><figcaption>Figure 8 — Accessing, viewing and downloading previous reports</figcaption></figure><p>Like the other two use-cases, this took 2–3 hours of work for CoPilot to think, create, test, fix etc. All along my only input was adding more details in the chat, answering questions or providing error messages that appeared on the browser when I was initially running it, and letting CoPilot figure out what to do.</p><p>While not covered here, I did go with the more structured approach afterwards using the initial prompt to create a proposal, which I reviewed, and getting a to-do list from that to get it to implement. This resulted in a much more advanced solution, and likely more effective. This approach is clearly the way to go for more complex requirements, I would think, and I know various experts out there recommend this way of working with it.</p><p>So …. I am 3 for 3 here, with my tests! So what do I think about the status of agentic code creation with VS Code and CoPilot?</p><p>Having previously used the code completion, creation of snippets, and asking questions when developing code, I can say that the agentic approach takes things to a new level, and we are very much at the start of this.</p><p>I understand that agentic software creation is the next attack target for Open AI with Codex, so I can only imagine what improvements we’ll see. Claude Code similarly brings even more capability and automation.</p><p>Understanding how it works and understanding how the software might address the problem space is more important than it might appear, as it currently needs some guidance. No doubt this will evolve.</p><p>I think that this will be a great tool for creating systems faster, but I still feel that some level of software engineering will be needed to assess the proposal and guide the tool. It is not unlike a system architect working with a group of programmers, guiding the direction and taking the decision on the technical choices to be made.</p><p>Whatever is coming, it is going to be interesting to see! Keep in mind that we are not even 3 years from the initial emergence of ChatGPT and the LLM explosion onto the world, where will we be in another 3???</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=43d164a3692f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>