<?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 Sumita Palanisamy on Medium]]></title>
        <description><![CDATA[Stories by Sumita Palanisamy on Medium]]></description>
        <link>https://medium.com/@sumita.palanisamy6?source=rss-442abd5d70f9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*MfnMTlKfeGBzCsDU.</url>
            <title>Stories by Sumita Palanisamy on Medium</title>
            <link>https://medium.com/@sumita.palanisamy6?source=rss-442abd5d70f9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 22:55:18 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sumita.palanisamy6/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[Sagamore Bridge : 77 Minutes vs 58 Minutes]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@sumita.palanisamy6/sagamore-bridge-77-minutes-vs-58-minutes-a9c0f3cc3ade?source=rss-442abd5d70f9------2"><img src="https://cdn-images-1.medium.com/max/1400/1*n6GWOnzm4CcMY8F9VlepHw.png" width="1400"></a></p><p class="medium-feed-snippet">A stress test for the word &#x201C;agentic,&#x201D; run on the most utilitarian problem I could find.</p><p class="medium-feed-link"><a href="https://medium.com/@sumita.palanisamy6/sagamore-bridge-77-minutes-vs-58-minutes-a9c0f3cc3ade?source=rss-442abd5d70f9------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@sumita.palanisamy6/sagamore-bridge-77-minutes-vs-58-minutes-a9c0f3cc3ade?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9c0f3cc3ade</guid>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Tue, 07 Jul 2026 23:25:01 GMT</pubDate>
            <atom:updated>2026-07-07T23:50:27.810Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[From user-demonstrated workflows to Reusable Skills: What Agentic Automation Adds to the Automation…]]></title>
            <link>https://medium.com/@sumita.palanisamy6/from-demonstration-to-automation-423c9fcd2c04?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/423c9fcd2c04</guid>
            <category><![CDATA[no-code-automation]]></category>
            <category><![CDATA[record-and-play]]></category>
            <category><![CDATA[codex-plugin]]></category>
            <category><![CDATA[xcode]]></category>
            <category><![CDATA[workflow-automation]]></category>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Sun, 21 Jun 2026 19:08:11 GMT</pubDate>
            <atom:updated>2026-06-22T19:27:13.073Z</atom:updated>
            <content:encoded><![CDATA[<h3>From user-demonstrated workflows to Reusable Skills: What Agentic Automation Adds to the Automation Landscape</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3iMyvnAN2youNQ9_BkZudA.png" /></figure><p><strong>I taught an agent a task, turned it into a reusable skill, and built a recurring SharePoint monitor — all in one evening.</strong></p><p>One task began with a demonstration. The other began with a destination and a schedule.</p><p>I showed an AI agent how to copy text between documents, and the demonstration became a reusable skill. Then I built a SharePoint monitor that checks for new files and responds only when something changes.</p><p>Different starting points. Different mechanisms. Same destination: automation.</p><p>That difference matters — especially for organizations already using robotic process automation (RPA), workflow platforms, APIs, and other enterprise automation tools. These experiments do not prove that one technology will replace another. They do, however, show how AI-assisted automation can package a demonstrated task as a reusable capability and then place capabilities inside recurring, outcome-oriented operations.</p><p>The strategic question is no longer only, “What process can we automate?” It is also, “What is the fastest, safest, and most governable way to teach an agent to perform the work?”</p><h3>Experiment one: teaching a desktop workflow by demonstration</h3><p>The first experiment was intentionally simple: copy text from one TextEdit document into another and save the destination document.</p><p>The value was not in the complexity of the task. It was in how the workflow was created.</p><p>Using Record &amp; Replay, I demonstrated the process on my Mac. The recording captured the sequence of actions involved — identifying the source and destination, selecting the content, copying it, placing it in the correct document, and saving the result. That demonstration was then turned into a reusable Codex skill.</p><p>The result was more than a recording that could simply be played back. It became a written, reusable set of instructions that Codex could invoke when asked to perform the same kind of task again.</p><p>This resembles traditional user-interface automation because it operates at the application layer. It is useful when a system does not provide an API or when the real process spans desktop tools, menus, windows, and files.</p><p>It also inherits some familiar challenges. Interfaces change. Window state matters. Ambiguous filenames or multiple open documents can create risk. A production version would need validation, exception handling, access controls, and clear confirmation points.</p><p>Still, the speed of going from demonstration to reusable workflow was striking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XlHc3lqEoAVz_1oudSRCqw.png" /></figure><p><em>A publication-safe recreation using sample data: a demonstrated desktop task became a reusable skill that copies, saves, and verifies a destination document.</em></p><h3>Experiment two: monitoring SharePoint on a schedule</h3><p>The SharePoint experiment used a different model. It was not a direct SharePoint connector or API integration. It was a recurring browser-based monitor operating through an authenticated SharePoint session.</p><p>The workflow checked a specific document library against an established baseline. If it found a genuinely new file, it recorded the visible filename and time, opened the document read-only, extracted readable content when possible, and prepared a concise summary with key facts and action items. If nothing changed, it remained quiet.</p><p>This was an automation in the operational sense: it had a trigger, a schedule, state, decision rules, and a notification policy. It was also intentionally constrained. The monitor was instructed never to modify, move, rename, share, or delete files.</p><p>The browser layer made it possible to work with the signed-in application even without building a new API integration. But it also means the workflow depends on the browser session and the visible SharePoint interface. Authentication expiry, interface changes, or unavailable reading views can interrupt it.</p><p>In other words, the first experiment created a reusable <strong>skill</strong>. The second used automation logic to run a <strong>workflow repeatedly over time</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-iZcc3gw4nqsDUIR8N8_WA.png" /></figure><p><em>A publication-safe recreation of the read-only SharePoint monitor. Institutional tenant information, account details, names, URLs, filenames, and document content have been replaced with generic sample data.</em></p><h3>So, are both of these automation?</h3><p>Yes — but at different layers.</p><p>The Record &amp; Replay example produced a <strong>reusable skill</strong>. It captures how a person completes a particular task through an application interface and converts that process into instructions the agent can invoke again.</p><p>The SharePoint example is a <strong>recurring automation</strong>. It combines a schedule, browser interaction, a baseline, read-only guardrails, change detection, content interpretation, and a notification decision.</p><p>A useful way to think about the distinction is:</p><ul><li><strong>Record &amp; Replay skill:</strong> “Here is how I perform this task. Learn it so you can use it again.”</li><li><strong>SharePoint monitor:</strong> “Check this location on a schedule, remember the baseline, and notify me only when something changes.”</li></ul><p>Neither approach is universally better. The right choice depends on the system, the task, the available interfaces, and the level of control required.</p><h3>Where this may fit alongside RPA</h3><p>It is tempting to frame every new automation technology as a replacement for the previous one. I do not think these small experiments justify that conclusion.</p><p>Traditional RPA platforms have mature capabilities for orchestration, scheduling, credentials, monitoring, audit trails, deterministic execution, and enterprise support. Those capabilities matter — particularly for regulated, high-volume, or business-critical processes.</p><p>AI-assisted automation introduces a different set of possibilities. It may be useful when work is document-heavy, when inputs vary, when interpretation is required, or when a process crosses both integrated and non-integrated systems.</p><p>The emerging pattern may be complementary:</p><ol><li>Use a direct integration when the target system exposes supported capabilities.</li><li>Use interface-level automation to bridge applications or steps without suitable integrations.</li><li>Use reasoning to interpret content, select the next action, and manage variation.</li><li>Keep deterministic controls, human approvals, and auditability around consequential actions.</li></ol><p>That could make AI-assisted automation a potential alternative for selected workflows — not a blanket replacement for an established automation program.</p><h3>My takeaway</h3><p>What interested me most was not that I automated two tasks. It was that the two tasks revealed two different paths to automation.</p><p>One path begins with human demonstration and turns a familiar sequence of actions into a reusable skill. The other wraps browser interaction, memory, rules, and a schedule around an ongoing outcome.</p><p>The more interesting future may not be one automation technology replacing another. It may be intelligent orchestration choosing the right method for each part of the work: direct integration when possible, interface automation when necessary, and human oversight where judgment or risk demands it.</p><p>These were small experiments. But they made the changing shape of automation feel much more tangible — and gave me several ideas for what to test next.</p><p><em>Author’s note: This article describes exploratory prototypes. Any enterprise implementation would require appropriate security, privacy, compliance, governance, and technology review.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=423c9fcd2c04" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Do KPop Demon Hunters Sound Different from Real Idols?]]></title>
            <link>https://medium.com/@sumita.palanisamy6/do-kpop-demon-hunters-sound-different-from-real-idols-30d94a1aa194?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/30d94a1aa194</guid>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 01:04:22 GMT</pubDate>
            <atom:updated>2025-12-02T15:29:02.378Z</atom:updated>
            <content:encoded><![CDATA[<h3>Do <em>K-Pop Demon Hunters</em> Sound Different from Real Idols?</h3><p>Using simple music features to see how <em>KPop Demon Hunters</em> stacks up against BTS and BLACKPINK.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jhCqgW7FycSn2nRR5XjCsQ.png" /></figure><p>When I finished watching <em>KPop Demon Hunters</em>, my brain was stuck on one thing ( other than the Soda Pop song) :</p><blockquote><strong><em>Do these fictional idol groups actually </em>sound<em> different from real K-pop idols OR is it all the same pop DNA with different faces?</em></strong></blockquote><p>I decided to treat that question like a data-science project. I curated a small set of songs, some lightweight musical features, and a few visualizations.</p><h3>What was I actually trying to answer using data science?</h3><p>I framed the project around three questions:</p><ol><li><strong>Fiction vs reality:</strong><br> Do fictional idol songs (Huntrix, Saja Boys) differ in tempo, key, and perceived energy from real-world idols (BTS, BLACKPINK)?</li><li><strong>Inside the KDH universe:</strong><br> Within <em>KPop Demon Hunters</em>, how do hero songs (Huntrix), villain songs (Saja Boys), and OST tracks differ in tempo, energy, and major/minor mode?</li><li><strong>Tempo vs vibes:</strong><br> How strongly is perceived energy related to tempo .Does faster always mean more intense?</li></ol><p>Everything else : plots, tables, little stats ,was in service of those three.</p><h3>Data Collection</h3><p>I collected a small curated set of 33 songs from five sources: the fictional idol groups Huntrix and Saja Boys, three KDH OST tracks (“Free”, “Jinu’s Lament”, “Prologue”), and real idols BTS and BLACKPINK.</p><p>For each song I collected lightweight musical and perceptual features: tempo (BPM), key and mode (major/minor), duration in ms and seconds, popularity and a manual subjective energy rating (1–5) and role. I was able to get the perceptual features such as Tempo , Key and mode using Tunebat.</p><p>Each song info- such as track id , artist name ,along with duration and track id was obtained using Spotify APIs.</p><p>Tracks were also manually tagged by role (hero anthem, villain song, OST cue, etc.) and source type (fictional idol, real idol, OST).</p><p>The dataframe.head() looks like this</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yjnZgl7TARHgw2AzNaEjWA.png" /></figure><h3>EDA</h3><p>For each song I collected lightweight musical and perceptual features: tempo (BPM), key and mode (major/minor), duration, and a manual subjective energy rating (1–5).</p><p>Each song info along with duration and track id was obtained using Spotify APIs. I was able to get the perceptual features using Tunebat.</p><p>Tracks were also tagged by role (hero anthem, villain song, OST cue, etc.) and source type (fictional idol, real idol, OST).</p><h3>Tempo</h3><p>I plotted <strong>tempo (BPM)</strong> for fictional vs real idols using boxplots.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jHaAhk07qkiCvN7ns8OrgQ.png" /></figure><h3>Energy vs Idol Type</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cMHoUL_knleAgSsg1PIw3Q.png" /></figure><p>Result : It looks like the Tempo is much higher for fictional idols and the so is the perceived energy. In other words, the Demon Hunters universe keeps its fictional idols “always on,” while BTS and BLACKPINK mix in more mellow material. This supports the idea that the fictional soundtrack is optimized for constant intensity, whereas the real groups balance hype singles with slower, more reflective songs.</p><h3>Hero vs villain vs OST (inside the KDH world) by Tempo</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ez7bXr63bpUHHpNohkE81w.png" /></figure><p>Result : Within the KPop Demon Hunters soundtrack, Saja Boys tracks exhibit the highest median tempo and the broadest tempo range (≈90–180 BPM), making them, on average, faster and more aggressive than Huntrix songs, which cluster in a narrower dance-tempo band around 120–140 BPM.</p><p>OST cues occupy a separate space, mixing one very slow lament with more mid-tempo tracks, resulting in the widest overall spread.</p><h3>Idol singles vs OST — Tempo vs Duration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yCHcMDJKMmVojC-XMm4JpQ.png" /></figure><p>Result : The tempo–duration scatterplot above shows that both fictional and real idol songs cluster tightly around 2.5–3 minutes with tempos in the 120–150 BPM range, reflecting a shared “pop single” structure.</p><p>In contrast, KDH OST tracks deviate from this pattern: one cue (“Jinu’s Lament”) is extremely short and slow, while the other OST tracks vary more in both tempo and duration. Overall, there is no strong relationship between tempo and duration, but OST tracks clearly occupy a more flexible, cinematic space compared to the tightly structured idol singles.</p><h3>Does faster = more energetic?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MT9uc6HEC7jlRDOk_Gvqpw.png" /></figure><p>Result : Subjective energy ratings show that both fictional and real idol songs are generally high-energy, but real idols more frequently reach the maximum rating of 5 and span the full range from 1–5.</p><p>Fictional idols avoid very low-energy songs and cluster around 4–5, while KDH OST tracks sit closer to mid-energy with one notably low-energy lament.</p><h3>Keys Difference in Fake Idols vs Real Idols</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X_J9f39jzDPmnlZi77Q1MQ.png" /></figure><p>Result : Looks like the Huntrix have songs in the major key wheras the Saja boys have majority in the minor key. This may show the hero vs villan divide.</p><p>Of course, the sample size for each group is small, so these patterns are more suggestive than definitive, but they illustrate how musical key can be used to reinforce character roles in a fictional idol universe.</p><p>The real idols are absolutely mixed — playing in both fields equally.</p><h3>Are Saja Boys not just faster, but also felt as more intense?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FD7NYgZ-1WX0-MX--S3mdA.png" /></figure><p>Result : Saja Boys are often faster than Huntrix ,but Saja Boys are not the most intense .They are strictly in the middle.</p><p>Within the KPop Demon Hunters soundtrack, Huntrix songs are consistently rated as very high-energy (4–5), while Saja Boys tracks sit in the mid-range (2–4) and OST cues are lower still.</p><p>Combined with the tempo analysis, this suggests that villains are written as fast but not maximally intense, whereas hero tracks are mixed to feel more relentlessly “turned up,” independent of BPM.</p><h3>Are more popular songs also more energetic?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FvU3W13I12vbLpCBxVNyIQ.png" /></figure><p>Result : There’s a moderate positive correlation (r ≈ 0.37) between my 1–5 energy ratings and Spotify popularity: higher-energy songs tend to be more popular in this dataset, but the relationship is far from deterministic. Plenty of mid-energy tracks do well, and some very energetic songs aren’t the most popular, which suggests that factors beyond “hype level” — like fandom, promotion, and emotional tone — also drive listenership.</p><h3>Summary</h3><p>Overall, fictional and real idols live in the same K-pop single zone: most songs are 120–150 BPM, ~2.5–3 minutes long, and have high energy around 4/5. The difference is that fictional idols are more “locked in” at high energy, while real idols span a wider emotional range from soft ballads to maximum-hype tracks.</p><p>Inside the KPop Demon Hunters world, Huntrix come out as the most intense (all songs rated 4–5), Saja Boys sit in the middle, and KDH OST cues are the calmest, including one very short, low-energy lament.</p><p>Key usage supports the story. Huntrix lean toward major keys, Saja Boys toward minor, while BTS and BLACKPINK are evenly split.</p><p>The data suggests that hero music is written and produced to feel relentlessly bright and powerful, villains are fast but slightly less overwhelming, and OST tracks occupy a separate, more cinematic space.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30d94a1aa194" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Real-World Use Cases with NanoBanana]]></title>
            <link>https://medium.com/@sumita.palanisamy6/using-nano-banana-in-irl-b88b99ef0c42?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/b88b99ef0c42</guid>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Tue, 16 Sep 2025 22:09:18 GMT</pubDate>
            <atom:updated>2025-09-23T01:05:07.975Z</atom:updated>
            <content:encoded><![CDATA[<p>I’ve been experimenting with <strong>Nano Banana</strong> in a few real-world scenarios, and I’m genuinely impressed. The standout differentiator for me is how well the model <strong>preserves input images</strong>, instead of generating entirely new ones — something that sets it apart from models like DALL·E 3.</p><p>A few use cases I’ve tested:</p><ol><li><strong>Real estate staging &amp; listings</strong> →I am using for staging and listing purposes. This helps in not waiting for the property to be cleaned before listing it. For example I was able to add furniture to the living room below</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IA5pPyhmWISWjTv94pOXkw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lFmuQzpks089pcZviROfig.png" /></figure><p>2. <strong>Style-based staging</strong> →I am also able to specify a style for the rooms I want to stage- I used victorian style for the bedroom.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*06j9CL7Uv_zTK3bStXVeEQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yF2mDJjh0_kcteJ-4e5IGw.png" /></figure><p>3. <strong>Cleaning </strong>-&gt; It even allowed me to stage the porch without having to fully prepare the space ahead of time!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5sTE3deOOM4C8YizOg3PEQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7rzPoi0Ir8fmgWZBa5tvLQ.png" /></figure><p>4. <strong>Remodel removing the patios-&gt;</strong> I was able to imagine what my remodel will look like when i removed the patios of floor 2 and 3</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XamslgNWUitv7HfGj3t7xw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XyzLVw-2hkE_OMT3ISeEFA.png" /></figure><p>5<strong>. Remodel adding Juliet Windows-</strong>&gt; Was able to generate a look as well so i can point my vision to a contractors</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pOPDIMuPUhbCeRV4lQjKrA.png" /></figure><p>6. <strong>Text accuracy in images</strong> →I am able to create new images with fewer obvious spelling mistakes with a single prompt. For example the prompt I gave was “can you create an Indian store that sells bangles and name it as “Sp’s bangles”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I0p9rqNHIlLAb2XopDi2EQ.png" /></figure><p>I couldn’t be more excited about the potential of this model — it’s practical, accurate, and already saving me time in workflows that usually involve a lot of waiting or manual clean up!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b88b99ef0c42" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Life-Saving Multi-Agent AI Solution using Crew AI]]></title>
            <link>https://medium.com/@sumita.palanisamy6/building-a-life-saving-multi-agent-ai-solution-using-crew-ai-7cad424ddb6b?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/7cad424ddb6b</guid>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Mon, 11 Aug 2025 10:45:23 GMT</pubDate>
            <atom:updated>2025-08-11T12:30:00.910Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TrorYR-516lv3x8DORVzPg.png" /></figure><p>In today’s world, where climate change is leading to increasingly unpredictable and extreme weather events, flood emergency response systems are more critical than ever.</p><p>I wanted to a explore a solution using cutting-edge solution using AI Multi agents to solve for this exact problem : a Flood Alert System built using crewAI, a framework that enables the creation of collaborative AI agent teams to tackle complex problems and inform folks who live in flood frone areas a plan ahead of time</p><h3>System Architecture: A Team of Specialized AI Agents</h3><p>The Flood Alert System employs six specialized AI agents, each with distinct responsibilities:</p><h3>1. Geographic Research Specialist</h3><p>This agent analyzes geographic information for specific zip codes, identifying nearby rivers, FEMA flood zones, and elevation data. It assesses flood risk by examining topographic data and historical flood events, providing the foundation for all subsequent decision-making.</p><h3>2. Daily Flood Monitor</h3><p>Working tirelessly in the background, this agent performs daily checks on flood conditions, weather patterns, and emergency alerts. It compares current data with baseline information to determine if risk levels have changed and whether notifications are necessary.</p><h3>3. Hotel Research Analyst</h3><p>When evacuation may be necessary, this agent researches hotels within a 50-mile radius that are outside flood zones and on higher ground. It ranks options based on safety level and proximity to the target area.</p><h3>4. Hotel Booking Coordinator</h3><p>This agent verifies hotel availability for families, checks booking options, and provides detailed reservation information. It ensures that in an emergency, people know exactly where they can go for safety.</p><h3>5. Flood Alert Coordinator</h3><p>The system’s integrator, this agent compiles all the flood risk analysis, safe hotel recommendations, and booking information into a comprehensive emergency response plan with clear, actionable steps.</p><h3>6. Gmail to Zapier SMS Coordinator</h3><p>The communication specialist, this agent sends urgent flood alerts via email, which then triggers Zapier to send SMS messages to users’ phones. This creates a reliable notification chain that works even when internet connectivity may be unreliable.</p><h3>The Workflow in Action</h3><p>Let’s walk through how the system operates:</p><ol><li>A user inputs a zip code they want to monitor</li><li>The Geographic Research Specialist analyzes flood risk for this location</li><li>The Daily Flood Monitor checks current conditions and alert levels</li><li>If risk is detected, the Hotel Research Analyst identifies safe accommodation options</li><li>The Hotel Booking Coordinator verifies availability and booking procedures</li><li>The Flood Alert Coordinator compiles all information into an action plan</li><li>If necessary, the Gmail to Zapier SMS Coordinator sends alerts to the user</li></ol><h3>Output and Next Steps</h3><pre>Subject: FLOOD ALERT  <br><br>Dear Resident,  <br><br>**Current Flood Status for 70112:**  <br>- **Alert Level:** Orange (Warning)  <br>- **Current Weather Conditions:** Flood Warning in effect with heavy rainfall expected (60% chance of precipitation).  <br>- **Flood Warnings:** Active flood warning issued by NWS.  <br>- **River Levels:** The Mississippi River is rising due to forecasted rain.  <br>- **FEMA Flood Zone Status:** You are in a designated flood zone.  <br><br>**Recommended Actions:**  <br>1. **Immediate Steps:**  <br>   - Monitor local weather updates and the National Weather Service advisories.  <br>   - Prepare an emergency kit with essentials (water, food, medications).  <br>   - Charge all electronic devices and have backup power ready.  <br><br>2. **Evacuation Recommendations:**  <br>   - Plan your evacuation routes, avoiding low-lying areas.  <br>   - Stay informed on local authority instructions regarding evacuations.  <br><br>**Primary Hotel Recommendation:**  <br>- **Hotel Name:** Drury Inn &amp; Suites Baton Rouge  <br>- **Address:** 7939 Essen Park Ave, Baton Rouge, LA 70809  <br>- **Contact:** +1 225-766-2022  <br>- **Booking Instructions:**  <br>   1. Call to inquire about room availability.  <br>   2. Confirm early check-in and cancellation policies.  <br><br>**Alternative Hotel Options:**  <br>- Holiday Inn Express &amp; Suites - Gonzales: +1 225-743-8000  <br>- Hampton Inn &amp; Suites Baton Rouge - I-10 East: +1 225-751-4600  <br>- La Quinta Inn &amp; Suites by Wyndham - Slidell: +1 985-641-8800  <br>- Hilton Garden Inn Covington/Mandeville: +1 985-327-7118  <br><br>**Emergency Contacts:**  <br>- Local Emergency Services: 911  <br>- New Orleans Office of Homeland Security: +1 504-658-8700  <br>- FEMA: 1-800-621-3362  <br>- American Red Cross: 1-800-733-2767  <br><br>Stay safe and prepared. Remain vigilant as conditions may change.  <br><br>Best regards,  <br>Emergency Notification Specialist</pre><p>Scheduling this to run daily and to get alerts via SMS would be the next logical step in this workflow.</p><h3>Future of AI in Emergency Response</h3><p>This Flood Alert System represents just the beginning of what’s possible with multi-agent AI systems in emergency response.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7cad424ddb6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building Voice-Interactive Experiences: Lessons from a Family Tech Project That Scaled]]></title>
            <link>https://medium.com/@sumita.palanisamy6/building-voice-interactive-experiences-lessons-from-a-family-tech-project-that-scaled-59ed6a6a2b64?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/59ed6a6a2b64</guid>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Thu, 03 Jul 2025 02:55:19 GMT</pubDate>
            <atom:updated>2025-07-03T14:57:24.567Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How a weekend bedtime story project revealed strategic insights about voice UX, and technical accessibility— plus the AI-assisted development workflow that made it possible</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uHdPFeQzDeglzBGzkp30qA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2sjt9kQwJ2QLHP-S6JGAEQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zgb2f8txQMcsI2y74OQylg.png" /></figure><h3>The Strategic Context</h3><p>As a Director of Engineering, I spend my days thinking about scalable architectures, team productivity, and technology roadmaps. But sometimes the most valuable technical insights come from the smallest experiments.</p><p>When my child asked if they could “hear more stories about Moana,” I saw an opportunity to explore voice interfaces ! What started as a weekend family project became a masterclass in voice UX design, progressive enhancement, and the strategic implications of accessible AI.</p><p>But perhaps more importantly, it became a case study in <strong>AI-assisted development workflows</strong> that has profound implications for how engineering teams will work in the coming years.</p><h3>The AI Development Partner Revolution</h3><p>This project was built entirely using Claude Artifacts — an AI-assisted development environment that fundamentally changed my development workflow. As engineering leaders, we need to understand not just what we’re building, but how AI is reshaping how we build.</p><h3>Traditional Development Flow:</h3><ol><li>Concept → Requirements → Architecture Design</li><li>Environment Setup → Framework Selection → Boilerplate Creation</li><li>Component Development → Integration → Testing</li><li>Debugging → Iteration → Deployment</li></ol><h3>AI-Assisted Development Flow:</h3><ol><li>Concept → Direct Conversation with AI</li><li>Immediate Interactive Prototype → Real-time Iteration</li><li>Refinement Through Natural Language → Deployment</li></ol><p><strong>Strategic insight:</strong> The time from idea to working prototype has collapsed from days to minutes. This isn’t just about individual productivity — it’s about fundamental changes in how we validate product concepts.</p><h3>The Artifacts Advantage: Rapid Prototyping</h3><p>Using Claude Artifacts, I could:</p><pre>Me: &quot;Add voice recognition so kids can ask questions about the story&quot;<br>Claude: [Instantly updates the working prototype with voice features]<br>Me: &quot;The microphone permission isn&#39;t working in embedded mode&quot;<br>Claude: [Immediately provides debugging insights and alternative approaches]</pre><p>This conversational debugging and feature development has strategic implications:</p><h3>1. Concept Validation Speed</h3><p>Ideas can be prototyped and tested within a single conversation. For product teams, this means we can validate 10 concepts in the time it previously took to build one.</p><h3>2. Cross-Functional Collaboration</h3><p>Non-technical stakeholders can see working prototypes immediately, not wireframes or mockups. This fundamentally changes product discussions.</p><h3>3. Architecture Exploration</h3><p>I could experiment with different technical approaches in real-time:</p><ul><li>“What if we used templates instead of AI for story generation?”</li><li>“How would this work as a standalone HTML file versus a React app?”</li></ul><p>Each question resulted in working code, not theoretical discussions.</p><h3>The Technical Architecture Decision</h3><p>The first strategic choice was <strong>build vs. buy vs. integrate</strong>. The obvious path would have been:</p><ul><li>OpenAI API for story generation</li><li>AWS Polly for text-to-speech</li><li>Google Speech-to-Text for voice recognition</li></ul><p>Instead, working with Claude, I explored a different approach: <strong>leverage browser-native capabilities</strong> with intelligent fallbacks.</p><pre>// Zero external dependencies, maximum compatibility<br>const recognition = new webkitSpeechRecognition();<br>const synthesis = window.speechSynthesis;</pre><p><strong>Strategic insight:</strong> AI-assisted development made it easy to explore the “simple but sophisticated” approach that I might have dismissed in a traditional planning meeting.</p><h3>Lessons in Product and Engineering Strategy</h3><h3>1. Progressive Enhancement as Competitive Advantage</h3><p>Rather than building a voice-first experience that fails without perfect conditions, I designed (with AI assistance) for <strong>graceful degradation</strong>:</p><ul><li><strong>Tier 1:</strong> Full voice interaction (Chrome/Edge with mic access)</li><li><strong>Tier 2:</strong> Text-to-speech playback only (most browsers)</li><li><strong>Tier 3:</strong> Text-based Q&amp;A (universal compatibility)</li></ul><p><strong>AI Development Advantage:</strong> I could rapidly test each tier and ensure smooth fallbacks, something that would have taken much longer with traditional development.</p><h3>2. The Template Strategy vs. AI-First Approach</h3><p>Instead of generating stories with GPT, I built a template system that combines user inputs into coherent narratives. Working with Claude, I could explore both approaches quickly:</p><p><strong>Template Approach Wins:</strong></p><ul><li>Predictable content quality</li><li>Zero API costs at scale</li><li>Complete content control</li><li>Sub-second response times</li></ul><p><strong>AI Approach Limitations:</strong></p><ul><li>Unpredictable outputs with children’s content</li><li>API dependency and costs</li><li>Content moderation complexity</li></ul><p><strong>Strategic takeaway:</strong> AI-assisted development lets us rapidly compare architectural approaches with working prototypes, not just theoretical analysis.</p><h3>3. Voice UX: Strategic Insights for Enterprise</h3><p>Building voice interactions for children taught me principles that apply directly to enterprise voice interfaces:</p><h3>a. Confirmation Loops Are Critical</h3><pre>// Always show what was heard<br>recognition.onresult = (event) =&gt; {<br>  const transcript = event.results[0][0].transcript;<br>  displayUserSaid(transcript); // Visual confirmation<br>  processCommand(transcript);<br>};</pre><p><strong>AI Development Insight:</strong> Claude could instantly implement and iterate on these UX patterns, allowing me to test different confirmation approaches rapidly.</p><h3>b. Multimodal Redundancy</h3><p>Every voice interaction had a text equivalent. With AI assistance, implementing these redundant pathways was trivial rather than time-consuming.</p><p><strong>Strategic insight:</strong> When AI reduces implementation friction, we can afford to build more robust, accessible interfaces.</p><h3>Scalability and Architecture Lessons</h3><h3>Browser APIs as Infrastructure</h3><p>Using native browser capabilities meant zero backend infrastructure for core functionality. AI assistance made it easy to explore and implement browser APIs that I might not have considered:</p><ul><li><strong>Development velocity:</strong> No API integration, no deployment pipelines for core features</li><li><strong>Scaling costs:</strong> Fixed costs regardless of usage</li><li><strong>Geographic performance:</strong> No API latency, works globally</li><li><strong>Offline capability:</strong> Core functionality works without connectivity</li></ul><h3>The Composable Frontend Strategy</h3><p>With AI assistance, I could build a sophisticated single-file application without the usual complexity trade-offs:</p><ul><li><strong>Zero build complexity:</strong> No webpack, no transpilation, no dependencies</li><li><strong>Maximum portability:</strong> Runs anywhere HTML works</li><li><strong>Easy distribution:</strong> Single file sharing, email attachment, GitHub gist</li><li><strong>Developer onboarding:</strong> Any developer can read and understand the entire codebase in 30 minutes</li></ul><p><strong>Strategic insight:</strong> AI-assisted development makes sophisticated functionality accessible without traditional complexity overhead.</p><h3>Voice Technology Accessibility</h3><p>This project demonstrated that sophisticated voice interfaces are now accessible to any web developer, especially with AI assistance lowering the learning curve.</p><h3>Looking Forward: Strategic Recommendations</h3><p>This project revealed that we’re experiencing a fundamental shift in how software gets built. AI assistance doesn’t just make development faster — it makes different types of solutions accessible and economical.</p><p>As technical leaders, we need to understand both the technical capabilities AI unlocks and the workflow changes it enables. The teams that adapt their development processes to leverage AI partnership will have significant competitive advantages in both speed and quality.</p><p><strong>Disclaimer</strong></p><p>This is a fan-made project inspired by Disney’s Moana. All character names, settings, and references are property of Disney. This project is for educational and entertainment purposes only.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=59ed6a6a2b64" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using AutoGen to analyze Blue Bikes routes in Boston]]></title>
            <link>https://medium.com/@sumita.palanisamy6/using-autogen-to-analyze-blue-bikes-routes-in-boston-88276854df42?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/88276854df42</guid>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 18:04:38 GMT</pubDate>
            <atom:updated>2025-01-07T18:04:38.067Z</atom:updated>
            <content:encoded><![CDATA[<p>I wanted to experiment with AutoGen, an open-source framework that allows developers to build LLM applications by composing multiple agents to converse with each other to accomplish tasks.</p><p>AutoGen agents are customizable, conversable, and can operate in various modes that employ combinations of LLMs, human inputs, and tools.To make these agents suitable for multi-agent conversation, every agent is made conversable — they can receive, react, and respond to messages. This was key for choosing AutoGen for this visualization project.</p><p>My goal for the project was simple:-</p><ol><li>Test out the code generator and the code executor part of the AutoGen framework.</li><li>Are the conversable agents able to collaborate and course correct if there are any issues found while execution of the generated code?</li><li>Find out any interesting insights about the Blue Bikes routes in Boston.</li></ol><h4>Implementation</h4><p>I started by downloading the <a href="https://bluebikes.com/system-data">Blue Bikes data</a> .The data looks like this for the month of December 2024.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x619ijVYF_JLLNkdGq2rYA.png" /></figure><p>I proceeded to define the task at hand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wN2QMayWR-eYqEYNjiMMpA.png" /></figure><h4>Build a group chat</h4><p>This group chat will include these agents:</p><ol><li><strong>User_proxy</strong> or <strong>Admin</strong>: to allow the user to comment on the report and ask the writer to refine it.</li><li><strong>Planner</strong>: to determine relevant information needed to complete the task.</li><li><strong>Engineer</strong>: to write code using the defined plan by the planner.</li><li><strong>Executor</strong>: to execute the code written by the engineer.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yd7KgLGyjjBEyRNFtTu9bA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zYLVK4-oS96J8ruqumimXA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y91vN-PMCSTAaPLUUBGlzA.png" /></figure><p>Then I defined the group chat and spicifed the number of max rounds and the agents themselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7zu_SWPDuhpcYBO8MMuOTA.png" /></figure><p>Finally I kicked off the group chat</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*azZtoXI43jBlgYrHVswhtg.png" /></figure><p>The Admin talking to the Planner first as shown below:-</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8m4aHDURCurXIEnr_bsJww.png" /></figure><p>and then the Engineer takes over</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3hfoXJLdT64PVesu9g3nWg.png" /></figure><p>and the executor follows</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mQOe_8redckf4IF3UgMEOw.png" /></figure><p>When an issue was identified by the engineer — the agents can be seen going back and forth to fix the bug</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*myzr-ejaMAix0O_ppMjKkw.png" /></figure><p>The executor gets the updated script and it is able to generate the correct results!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nril3k57-cTTrTQ7rpK-Pw.png" /></figure><p>Finally the engineer generates the code for the barchart and the executor runs it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zCbJLmbrEElPgJscWsxU9Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LqG1re1gWoqVmaeapDghpQ.png" /></figure><p>The plot didn’t look right with all the route names cut off as shown above. So I modified the task to be more specific leading to much better results as shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ALvTlkaMb-GJ42dnJ8MPTQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VxiICPKmMzexWE0CbWzIrg.png" /></figure><h4>Lessons Learnt</h4><p>1) I tried to get the data from Blue bikes via an agent hoping it could scrape the data from the site, but it was not able to identify what month i was looking for by just scraping the data, which was the reason I download the data and added it as a csv file to the project directory.</p><p>2) There is a need to be specific for what is needed in the task — as we saw in the wrapping text example. Like IRL debugging, coding with agents is also an iterative process.</p><p>3) The collaborative work performed by the agents is truly inspiring and its also a reminder of real worls projects!</p><p>Conversable agents are here to stay and I hope they will get better throughout this year and beyond! Happy Coding!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=88276854df42" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Making Restaurant Reservations using Crew AI and GPT]]></title>
            <link>https://medium.com/@sumita.palanisamy6/making-restaurant-reservations-using-crew-ai-and-gpt-9a2bc966dee3?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/9a2bc966dee3</guid>
            <category><![CDATA[crew-ai]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[gpt]]></category>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Tue, 31 Dec 2024 04:24:29 GMT</pubDate>
            <atom:updated>2024-12-31T04:54:34.199Z</atom:updated>
            <content:encoded><![CDATA[<p>As I am preparing to go to the NADA 2025 road show, the teams at work are preparing to meet with partners and clients while we are at New Orleans. However looking for ideal locations for dinner/lunch that are close to the NADA location is a struggle because</p><p>1) A large group size is not always ideal for booking directly via OpenTable. For parties of more than 10 , restaurants ask you to call them directly</p><p>2) Its good to choose a restaurant that’s close to the NADA event location, so that folks can return back to the event space if needed</p><p>I was able to resolve this by using CrewAI and GPT 3.5-turbo.</p><p>As per CrewAI’s documentation, I created Agents, Tasks , Set the constraints and Ran the Crew to accomplish the restaurant reservation use case.</p><h4>Creating Agents</h4><p>I started by creating a restaurant co-ordinator that identifies 10 restaurants that fit the criteria for an official meal with clients and partners. I was also able to specify the time and location of the conference itself to make the code more configurable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kBWBE_lsfUZQ8Oe5mjvq2g.png" /><figcaption>Agent #1</figcaption></figure><p>The second agent created is making sure the specified logistics are agreeable — such as rating and ambience</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X2FXHjj25qGKnVowOE7hlA.png" /><figcaption>Agent #2</figcaption></figure><h4><strong>Creating Tasks</strong></h4><p>The first task was to way to list the restaurants that satisfy the criteria given for people and availability</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dLh5LreZFaEYY2RUwN222A.png" /></figure><p>The second task is to make sure the rating of the restaurant is at a minimum at 4.5</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2cIA61cYUg5vAwv35T1zow.png" /></figure><h4>Constraints and KickOff</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hjUJK6g7sdI7a3VUHlgx4w.png" /><figcaption>Setting up the constraints</figcaption></figure><p>The final step after the constraints is the kick off.</p><p>The list returned below shows the various restaurants and their address and phone numbers along with distance from the conference location. Having all this information in one spot prevents the need to look up additional information, thereby creating an efficient reservation process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hqR14kBLOKifpF16pSIayg.png" /><figcaption>Top 5 of the 10 results returned by the Crew</figcaption></figure><h4>Conclusion</h4><p>Using #crewai and the multi-agent set up has been an absolute pleasure! Can’t wait to see what 2025 holds for us in terms of all things AI — Happy coding!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9a2bc966dee3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Creating a Karaoke song suggestion app using #awsbedrock]]></title>
            <link>https://medium.com/@sumita.palanisamy6/creating-a-karaoke-song-suggestion-app-using-awsbedrock-e25914dd9de4?source=rss-442abd5d70f9------2</link>
            <guid isPermaLink="false">https://medium.com/p/e25914dd9de4</guid>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[party-rock]]></category>
            <category><![CDATA[aws-bedrock]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Sumita Palanisamy]]></dc:creator>
            <pubDate>Sun, 07 Jan 2024 00:30:16 GMT</pubDate>
            <atom:updated>2024-01-07T02:25:45.645Z</atom:updated>
            <content:encoded><![CDATA[<p>AWS announced PartyRock, an Amazon Bedrock Playground in November ‘23. PartyRock is an intuitive hands-on, generative AI app-building playground. In just a few steps, you could create a variety of apps to experiment with generative AI!</p><p><a href="https://medium.com/u/1d1fa7d2f998">Daniel Ansted</a> and I built a #generativeai <a href="https://partyrock.aws/u/noob1701/BhqtCOeU-/FindMyJam">app</a> for suggesting songs to Karaoke singers based on their current mood, singer, or genre using #partyrock.</p><p>We used the Claude Instant LLM from Anthropic for our app — the reason for this being PartyRock Credits. PartyRock credit usage varies by model to help users build intuition for cost consideration when using generative AI and this model was a cost effective model for us.</p><p><strong>Pros</strong></p><ul><li>The app was really easy to build and fine-tune</li><li>Truly no-code ( we built the app in under half an hour)</li><li>Easy way to play with prompt-engineering and #generative-ai</li><li>Free to develop ( for now). Generous credits — with our testing we had used less than 1% of our allocated usage. When users use the app , the credits are used from the user’s account not the app builder’s account.</li><li>Way to use the foundation models from AWS Bedrock</li><li>Learn how credits associated models work — being aware of the cost associated with each AWS service is crucial for anyone working with AWS</li><li>Ability to mix apps — so you need not build an app from scratch.</li><li>The LLM did perform really well for songs in the “pop” category or while choosing “Taylor swift” as an artist.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Eb7lvJGMisxgElIKEhVnzg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OqMEki-2mUywAos7dTAuOQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SFModNQXdFdsYdZniO-X2g.png" /></figure><ul><li>The LLM performed Ok when tested out for non english songs as well! We used “ar.rahman” for testing out this use case as shown below:-</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PblDO2KFRi_mWFegcm_h4A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qg4gJqI8Rf1gjvOnixbdqQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ffNigINxDwebAuMZwVyGIg.png" /></figure><p>We also did a test for a Spanish song as well as shown below:-</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LkxrB1C5t2RJNhRtHj94qA.png" /></figure><p><strong>Cons</strong></p><ul><li>The LLM did not do well while choosing “Weird AL” as shown below. The lyrics to the chosen song were incorrect.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WBJ25gU3K3qWe1Wj1dOOog.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lWoo8Rn8wJXejlf73Cd0iQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8HeIiVfnVcrTvbGRn6ZRbw.png" /></figure><ul><li>The linking of the widgets was clunky</li></ul><p><strong>Conclusion</strong></p><p>Using #partyrock , we were also able to generate an image for this medium article — check out our <a href="https://partyrock.aws/u/Spala/4cQGF0HQY/Medium-article-Image-Generator">image generator app</a> to generate images for your articles!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*xnZ-SXC-wwzRH23cpEnTYw.png" /></figure><p>Overall, we really enjoyed the app(s) building process!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e25914dd9de4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>