<?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 Akshata K on Medium]]></title>
        <description><![CDATA[Stories by Akshata K on Medium]]></description>
        <link>https://medium.com/@akshatakini?source=rss-f7237c9e16f2------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*BbItu_px8r9hr26w6tu4qA.jpeg</url>
            <title>Stories by Akshata K on Medium</title>
            <link>https://medium.com/@akshatakini?source=rss-f7237c9e16f2------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 11:48:28 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@akshatakini/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[2026 Delta Fellowship Wrap Up: Building for Social Good!]]></title>
            <link>https://medium.com/deltanalytics/2026-delta-fellowship-wrap-up-building-for-social-good-005ee749ffee?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/005ee749ffee</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[nonprofit]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[data-for-good]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 19:21:52 GMT</pubDate>
            <atom:updated>2026-11-06T19:21:52.871Z</atom:updated>
            <content:encoded><![CDATA[<p>This year, we have 21 fellows joining us to deliver four data analytics and tech solutions projects for nonprofits working in financial inclusion and fraud prevention, mental health and economic mobility &amp; entrepreneurship. The solutions spanned across data engineering, AI engineering, dashboards and data visualizations, and AI-assisted analytics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KCtE2aY24dCSLe9Xam-OHQ.png" /></figure><p>Our 2026 fellowship kept the same format like last year, hybrid meetings and hackathons. This year, we have introduced the Product/Project Manager and Owner role. This role drove the project end-to-end and served as the primary point of contact between the partner non-profits and our project operations team. More information on the need for this role can be found <a href="https://medium.com/deltanalytics/adapting-data-fellowship-teams-project-ownership-860cb77080fa">here</a>.</p><h4><strong>2026 projects recap</strong></h4><p>Let’s dive into what we built this year!</p><p><strong>Give Us The Floor</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/250/1*PTd3RL_8U9YFr8FUm5gf6w.png" /></figure><p><a href="https://giveusthefloor.org/">Give Us The Floor (GUTF)</a> is a non-profit that fosters peer support for LGBTQ+ youth through online group chats led by trained youth facilitators. These safe, confidential spaces help alleviate mental health challenges like depression, isolation, and anxiety, with guidance from experienced mental health professionals.</p><p><strong>What we built:</strong> The fellows built an LLM powered solution that fully automated the sentiment analysis needed by the GUTF team to better understand how user engagement, conversation patterns, and content relate to emotional well-being, so they can strengthen what works and intervene earlier when support is needed.</p><p><strong>Tech stack/methods: </strong>The team prompt-engineered the Phi-4-mini model to score more than 50,000 user messages on a scale from very positive to very negative, then productionized the entire pipeline on Google Cloud Platform. Because high message volume kept average scores clustered near the middle, they defined Net Sentiment Score (NSS), which the organization adopted for its clearer, more actionable signal. They also built a Looker Studio dashboard with sentiment cohort analysis and a state-level heatmap.</p><p><strong>Impact: </strong>This project enabled the organization to measure user sentiment at scale and turn unstructured data into a quantifiable signal. It made it easier for stakeholders to identify meaningful trends in sentiment across cohorts and states and prioritize areas of improvements.</p><p><strong>Team:</strong> Akshata Kini, Shobana Siranjeevilu, Shounak, Sree Nair, Ridha, Jason Kushner</p><p><strong>Video Presentation: </strong><a href="https://www.youtube.com/watch?v=beYjk7fSWbI">https://www.youtube.com/watch?v=beYjk7fSWbI</a></p><p><strong>The Tazama Project</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*69dwXIE4JoWRKiYTAV3rSg.png" /></figure><p><a href="https://www.tazama.org/">Tazama</a> is an open-source, real-time transaction monitoring platform that helps financial institutions and payment providers detect fraud and money-laundering patterns. Billed as the first platform of its kind, it was designed with emerging markets in mind, where costly proprietary tools can be out of reach. Hosted by Linux Foundation Charities and backed by the Gates Foundation, it has been piloted and evaluated by institutions in Africa and elsewhere, including South Africa and West Africa’s central bank (BCEAO)</p><p>Delta worked on two projects with Tazama, one related to GenAI chatbot/agent to answer end user questions and the second to build dashboards and visualizations for a catalogue of reports aligned to their ISO20022 message/transaction data.</p><p>Project 1 — AI Solution</p><p><strong>What we built:</strong> The team built an AI-Assistant called FraudClarity that integrates with the Jupyter environment to answer end user questions. This chatbot is useful to help onboard users to Tazama’s open-source fraud monitoring platform.</p><p><strong>Tech stack/methods: </strong>Fellows built a five-layer pipeline in which source documents were chunked and embedded into Chroma DB, a local vector database, and Ollama, a local model, was used to retrieve the answers. LangChain and LangGraph were used to orchestrate the agent workflow and state management from user query to answer retrieval, and Gradio was used for the interactive interface. The team delivered a fully functional local demo that is ready to scale into production.</p><p><strong>Impact: </strong>Investigators are now able to use this chatbot to ask questions in natural language about the definition of a Tazama rule, rule query measures, the transaction attributes each rule reads, and get guidance on how rule outputs should be interpreted.</p><p><strong>Team:</strong> Yvette Kondoh, Lily Chebotarova, Dhanashree Pokale, Danish Kamal Azmi, Phani Karripedi, Shashank Singhal</p><p>Project 2 — BI Solution</p><p><strong>What we built:</strong> For the BI solution, the fellows team built visualizations that allow analysts to investigate patterns in fraud case management. The Alert Navigator eliminates the manual process investigators used to analyze fraud alerts, and the Rule Discovery Trigger lets investigators immediately identify the specific behavior (constituent transactions or metadata) that triggered an individual rule, without manually searching a participant’s entire transaction history.</p><p><strong>Tech stack/methods: </strong>The tools were built within the Jupyter ecosystem to help fraud analysts examine each alert beyond the initial trigger. PySpark 3.4.2, Apache Hudi (a data lakehouse framework), and JSON parsing handled data pre-processing. Visualizations were built with Plotly and ipywidgets in Jupyter Notebooks, and Voilà, an open-source tool that converts notebooks into web applications, was used to create the final front end.</p><p><strong>Impact: </strong>These two tools provided instant visibility into fraud alerts without requiring any manual efforts for the analysts and the solution would be embedded inside the Case &amp; Investigation Management System available to every operator who deploys Tazama 4.0 worldwide. This program currently includes more than 12 participating organizations.</p><p><strong>Team:</strong> Shikha Pamnani, Emily Ward, Rajesh Vayyala, Husandeep Toor</p><p><strong>Village Capital</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/347/1*Bq9NPNI3N6ldGeO5McO3NA.png" /></figure><p><a href="https://vilcap.com/">Village Capital</a> helps entrepreneurs tackle climate, economic mobility, and access to essential services by providing accelerator programs, partnering with local support organizations, and connecting startups to the right kind of capital. It also builds tools to make the entrepreneurship support system work better for everyone.</p><p><strong>What we built:</strong> The team built an Intelligent Process Automation (IPA) solution, utilizing AI/NLP models, that automatically processes, categorizes, and analyzes qualitative and quantitative data from our entrepreneur surveys to identify recurring themes, sentiment patterns and priority improvement areas.</p><p><strong>Tech stack/methods:</strong> Raw survey data from 200+ companies across 44 countries, spanning the past 10 years, was standardized and processed in Google Colab. BERTopic was used to identify recurring themes in the qualitative data, VADER to score response sentiment, and facebook/bart-large-cnn to generate short survey summaries. Results were visualized in an interactive HTML page with ipywidgets for exploration.</p><p><strong>Impact:</strong> The solution enables VilCap to detect recurring concerns across programs, monitor participant sentiment in near real time, and prioritize enhancements based on evidence. By cutting survey analysis time by a projected 75%, it turns raw feedback into timely insights, allowing the team to respond to participant needs sooner and improve the entrepreneur experience.</p><p><strong>Team: </strong>Naman Singh, Sydney Moseley, Manasvi Malepati, Merlissa Alexandre, Ayushi Arora</p><p><strong>Video Presentation: </strong><a href="https://www.youtube.com/watch?v=cBEYisP0pb4">https://www.youtube.com/watch?v=cBEYisP0pb4</a></p><p><strong>A Heartful Thank You</strong></p><p>2026 Delta Analytics Fellowship wrapped up, and what a ride! Thank you to our volunteers and clients for giving your all and making it count. We’re grateful for everything we accomplished together!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=005ee749ffee" width="1" height="1" alt=""><hr><p><a href="https://medium.com/deltanalytics/2026-delta-fellowship-wrap-up-building-for-social-good-005ee749ffee">2026 Delta Fellowship Wrap Up: Building for Social Good!</a> was originally published in <a href="https://medium.com/deltanalytics">deltanalytics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build an LLM-Powered User Feedback Analysis App With Me]]></title>
            <link>https://medium.com/@akshatakini/build-an-llm-powered-user-feedback-analysis-app-with-me-00aeae73a447?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/00aeae73a447</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 04:09:19 GMT</pubDate>
            <atom:updated>2026-10-04T18:55:05.514Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A hands-on guide to building an AI-powered data analysis workflow, from collecting and classifying user feedback to validating AI output and deploying the app.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FSNxPu8RmlkeAy6MFZH4_w.jpeg" /><figcaption>Created using Microsoft’s Copilot</figcaption></figure><p><strong>What we’re building</strong></p><p>Do you want to know what people are saying about AI apps? Let’s find out by building a user feedback analysis app.</p><p>Starting with a pipeline that pulls real user feedback for an AI app from three live sources (Google Play, Apple App Store, Hacker News), classifies each item by category, sentiment, and urgency using an LLM, and surfaces what needs attention first. Live demo: [<a href="https://user-feedback-triage-pipeline-mbkyaeqefkrdzu9sjxqx8e.streamlit.app/">link</a>]. Code: [<a href="https://github.com/AkshataKini1/user-feedback-triage-pipeline">repo link</a>].</p><p><strong>Setting Up the Gemini API</strong></p><p>We’ll use Google Gemini’s API, specifically the free Flash-Lite model, since this project only needs classification, not heavy reasoning, and the free tier covers everything here at zero cost.</p><p>You’ll need:</p><ul><li>A free Gemini API key from <a href="http://aistudio.google.com/">aistudio.google.com</a> — no credit card required</li><li>A Google Colab notebook (or any Python environment) — no local setup needed</li></ul><p>One thing worth knowing upfront: Gemini’s free tier may use your prompts and responses to improve Google’s products. That’s fine for the public review data I’m using here, but it’s something to keep in mind if you’re working with private or sensitive data.</p><p><strong>Building the Data Pipeline</strong></p><p>Three very different APIs.</p><p>Google Play was the easiest: a Python library (google-play-scraper) pulls reviews with no authentication needed.</p><p>Hacker News uses Algolia’s public search API, also no auth required, but the text comes back full of raw HTML, &amp;#x27; instead of apostrophes, stray &lt;p&gt; tags, so it needs cleaning before it’s usable.</p><pre>import html, re<br><br>def clean_html(text):<br>    text = html.unescape(text)  # &amp;#x27; → &#39;<br>    text = re.sub(&#39;&lt;[^&lt;]+?&gt;&#39;, &#39;&#39;, text)  # strip tags like &lt;p&gt;<br>    return text.strip()</pre><p>Apple App Store doesn’t have an official reviews API, but there’s an undocumented public RSS feed that works well for recent reviews, also no auth.</p><p>I originally planned a fourth source, Reddit, but Reddit closed self-serve API access. New apps now require manual approval with no fixed timeline, not practical for a project like this. I swapped it for Hacker News and Apple instead.</p><pre># Without this, combining sources crashes when rendering the table<br>df[&#39;rating&#39;] = df[&#39;rating&#39;].astype(str)</pre><p><strong>Designing the LLM Classification</strong></p><p>The classification needed a fixed schema- category, sentiment, urgency, and a one-line summary, so results could be filtered and sorted consistently across all items.</p><pre>prompt = f&quot;&quot;&quot;Classify the following user feedback into structured JSON with these exact fields:<br>- category: one of [bug, feature_request, complaint, praise, question, other]<br>- sentiment: one of [positive, negative, neutral, mixed]<br>- urgency: one of [low, medium, high]<br>- summary: a single sentence summarizing the core point<br><br>Respond with ONLY valid JSON, no markdown formatting, no explanation.<br><br>Feedback: &quot;{test_text}&quot;<br>&quot;&quot;&quot;<br><br>response = client.models.generate_content(<br>    model=&quot;gemini-flash-lite-latest&quot;,<br>    contents=prompt<br>)<br><br>print(response.text)</pre><p>Before running it on the full dataset, I tested the prompt on a single Hacker News comment to make sure the output looked right. Here’s what came back:</p><pre>{<br>  &quot;category&quot;: &quot;complaint&quot;,<br>  &quot;sentiment&quot;: &quot;negative&quot;,<br>  &quot;urgency&quot;: &quot;medium&quot;,<br>  &quot;summary&quot;: &quot;The user is expressing extreme frustration over alleged silent regressions and an overwhelming amount of AI-generated content flooding open-source repositories.&quot;<br>}</pre><p>The response was valid JSON, and the classification looked right for the feedback. I tested it on one item first to make sure the format and classification were working before sending the rest of the data through.</p><p><strong>Batching for Performance</strong></p><p>One call per item worked, but it was slow, about 0.59 seconds per item on average. At that rate, processing a few thousand items would take quite a while.</p><p>So I tried batching instead. I put multiple feedback items into a single prompt and had Gemini return a JSON array with one result for each item. I tested batch sizes from 5 to 100.</p><p>All of the batch sizes worked in my tests. I got valid JSON back each time, with a result for every feedback item, and nothing failed.</p><pre>batch_sizes = [1, 5, 10, 20, 40, 60, 80, 100]<br>sec_per_item = [0.59, 0.226, 0.181, 0.154, 0.142, 0.156, 0.154, 0.144]</pre><p>Going from 1 to 20 cut the time per item by nearly 4x. After 20, the gains basically flattened out, the numbers stayed between 0.14 and 0.16 seconds per item, with no clear improvement from using larger batches.</p><p>I settled on 20 because it gave me almost all of the speed improvement without making each batch unnecessarily large. If a batch failed to parse, I’d have to rerun 20 items instead of 100.</p><p><strong>Output Validation: Making the Pipeline Reliable</strong></p><p>During the batch-size benchmark, a Gemini call returned 429 RESOURCE_EXHAUSTED. The loop had no delay or retry. Rerunning the same code three times did not reproduce the error, so it appears to be transient, and I can&#39;t say which limit was hit. Because it can recur, I added retry with backoff on 429s and a short pause between batches.</p><p>Separately, even with an explicit schema in the prompt, the model didn’t always follow it. On one item out of 147, it returned a category, &quot;concern&quot;, that wasn&#39;t one of the six I&#39;d defined. Nothing errored. I only caught it by reading through the output by hand.</p><pre>ALLOWED_CATEGORIES = {&quot;bug&quot;, &quot;feature_request&quot;, &quot;complaint&quot;, &quot;praise&quot;, &quot;question&quot;, &quot;other&quot;}<br><br>def normalize_item(item):<br>    category = item.get(&#39;category&#39;, &#39;other&#39;)<br>    if category not in ALLOWED_CATEGORIES:<br>        category = &#39;other&#39;<br>    return category</pre><p>A prompt-level schema is a strong suggestion to the model, not a guarantee. Validating the output after the fact is what actually makes it safe to trust.</p><p><strong>Turning It Into a Real App:</strong></p><p>The Colab notebook was enough to test everything, but I wanted something people could actually try. I moved the code into Streamlit and built a simple interface around it.</p><p>A few things came up once I started building the UI.</p><p><strong>Dropdown instead of free text for app selection.</strong> I could have supported any app, but that would mean figuring out how to map an app name to the correct Google Play and Apple App Store IDs. Instead, I limited the app to three options with verified IDs for now.</p><pre>APP_CONFIGS = {<br>    &quot;Claude&quot;: {&quot;google_play_id&quot;: &quot;com.anthropic.claude&quot;, &quot;apple_app_id&quot;: &quot;6473753684&quot;, &quot;hn_query&quot;: &quot;Claude Anthropic&quot;},<br>    &quot;ChatGPT&quot;: {&quot;google_play_id&quot;: &quot;com.openai.chatgpt&quot;, &quot;apple_app_id&quot;: &quot;6448311069&quot;, &quot;hn_query&quot;: &quot;ChatGPT OpenAI app&quot;},<br>    &quot;Gemini&quot;: {&quot;google_play_id&quot;: &quot;com.google.android.apps.bard&quot;, &quot;apple_app_id&quot;: &quot;6477489729&quot;, &quot;hn_query&quot;: &quot;Gemini Google app&quot;}<br>}</pre><p>Each user enters their own API key. I didn’t want to put my Gemini key in the app, so the key is entered through a password-masked field and only used for that session. It’s not stored or shared with other users. I also added a progress bar to show the classification progress as the batches run.</p><p>For deployment, I used Streamlit Community Cloud and connected it to my GitHub repo. Now when I push a change, the app redeploys automatically.</p><p>I started this project because I wanted to see how far I could get building an app with AI. I didn’t expect to get everything right, and I definitely didn’t. But I got something working, shipped it, and learned a lot about the parts of software development I usually don’t get to spend much time on.</p><p>It’s not perfect. Things I’d still want to improve:</p><ul><li>Add automated tests</li><li>Support arbitrary apps instead of three hardcoded app IDs</li><li>Replace the fixed pauses with rate-aware throttling</li><li>Check that the model returns every item in a batch, not just valid-looking ones</li></ul><p>One other thing worth knowing: Hacker News comments are discussion, not reviews, and the ratings aren’t directly comparable across sources. The app stores have star ratings, while Hacker News has comment points.</p><p>Live app: [<a href="https://user-feedback-triage-pipeline-mbkyaeqefkrdzu9sjxqx8e.streamlit.app/">link</a>] (bring your own free Gemini API key, it isn’t stored)<br> Code: [<a href="https://github.com/AkshataKini1/user-feedback-triage-pipeline">GitHub link</a>]</p><p>If you try it out or build something on top of it, I’d like to hear what you do with it.</p><p>I enjoy sharing the things I build and the things I learn along the way. Hopefully some of this is useful to you too. If you have questions or want to share how you’ve approached something differently, drop a comment. You can also find me on <a href="https://www.linkedin.com/in/akshatakini/">LinkedIn</a> if you want to stay in touch.</p><p>Happy Building!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=00aeae73a447" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[4 Things I Learned Building and Shipping an AI-powered App]]></title>
            <link>https://medium.com/@akshatakini/4-things-i-learned-building-and-shipping-an-ai-powered-app-660d70e76fc7?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/660d70e76fc7</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 01:58:01 GMT</pubDate>
            <atom:updated>2026-10-13T01:00:42.258Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i7ZQjmpZron37g8Ad31ijw.png" /><figcaption>Image created using OpenAI’s ChatGPT</figcaption></figure><p>I’ve always liked building things. Even over the past six years as a data scientist, I’ve enjoyed prototyping, messing around with ideas, and seeing if I could turn them into something that actually worked.</p><p>A few weeks ago, User Feedback Triage, a tool for AI apps that pulls in user feedback, uses an LLM to classify it, and helps surface the things that need attention first. You can find it live at this <a href="https://user-feedback-triage-pipeline-mbkyaeqefkrdzu9sjxqx8e.streamlit.app/">link</a>. I’ll get into how it works below.</p><p>Prototyping was straightforward. The more interesting challenge was taking the project from prototype to a finished, deployed application. I used Claude throughout the process, mostly as a way to move faster through parts of the stack I wasn’t as familiar with. I still made the decisions and tested everything myself, but I could spend less time getting stuck on implementation details and more time actually building the thing.</p><p>This app pulls real, live user feedback for a chosen popular AI app (Claude, ChatGPT, or Gemini) from three sources: Google Play reviews, Apple App Store reviews, and Hacker News discussions. Then it classifies each item/review using an LLM (Google Gemini Flash-Lite) into category, sentiment, and urgency along with a one-line summary. It then surfaces a prioritized list of high-urgency bugs and complaints, the items worth looking at first.</p><p>While building the product, I started thinking beyond just getting the thing to work. I found myself asking questions like: What could go wrong? How would this hold up if more people actually used it? What could I improve next?</p><p>Building this made me curious about a lot of new things, and I ended up learning a ton along the way. AI made it much easier to explore them without getting stuck for days.</p><p><strong>Here are some lessons I learned along the way</strong></p><p><strong>Test it, don’t assume:</strong> I tested batch sizes from 1 to 100, <strong>grouping multiple feedback items into a single LLM request</strong> and measured time per item. Moving from single items to batches of 20 cut latency from 0.59s to 0.165s per item (roughly 3.5x faster). Beyond 20, gains plateaued (0.15s at batch 100).</p><p>Even though batch size 100 ran without failing, that didn’t automatically make it the right choice. A couple of things I considered were, if an API request fails, a batch of 100 means recovering 100 items instead of 20. Longer prompts also increase the risk of the model degrading or missing nuance on later items. That tradeoff matters more at scale, where a failed batch can mean reprocessing a much larger number of items. I chose 20 items. It captured nearly all of the speed improvement while keeping the prompt length and failure domain manageable.</p><p><strong><em>Quick takeaway:</em></strong><em> Batch size isn’t just about maximizing throughput. It’s also about controlling how much work has to be redone when something goes wrong. I chose 20 because it captured most of the speed gains while keeping the potential cost of a failed batch reasonable.</em></p><p><strong>APIs come with someone else’s rules:</strong> I originally planned to pull reviews from Reddit, but my first attempt was blocked. I looked into getting official access, only to find that I needed approval first. That would have meant waiting around for approval, which I didn’t want to do for a weekend project, so I switched to Hacker News and the Apple App Store, which both offered free public endpoints.</p><p><strong>Even then, getting the data into my dashboard wasn’t as straightforward as I expected</strong>. Hacker News returned text with HTML that I had to clean up, and Apple sent star ratings as text while Google Play sent them as numbers. That broke my results table until I standardized the formats. Getting the data is only half the job. I also have to figure out what I can access, how it’s formatted, and what limits I’ll run into before building around it.</p><p><strong><em>Quick takeaway: </em></strong><em>You don’t control external data sources, so you have to work within their rules. Before building around an API, check what you can access, what the data looks like when it comes back, and what limits you’ll run into. Otherwise, you can end up fixing problems in your app later that you could have caught earlier.</em></p><p><strong>Validate AI output, always:</strong> When setting up the LLM classification pipeline, I explicitly listed six allowed categories in the system prompt. However, when auditing the outputs across 147 test items, I found an <strong>unapproved label called “concern”</strong>, which slipped through. I later realized <strong>that LLMs use prompt instructions as suggestions</strong>, unless you use a stricter structured-output feature or guardrails.</p><p>So,<strong> I added an output validation step after every API response</strong>. Now, if the model returns something outside the categories I defined, my app maps it to “other” instead of inventing new categories.</p><p><strong><em>Quick takeaway:</em></strong><em> LLM outputs are not deterministic, even when the instructions are clearly laid out. If your application depends on the model returning one of a fixed set of values, validate the response before passing it downstream. That way, an unexpected model output doesn’t silently make its way into your data or break something later.</em></p><p><strong>Don’t let AI speed replace your curiosity:</strong> AI made this much faster to build, no doubt. But the bigger change was how I learned. Normally, I’d try to understand a concept deeply, get pulled into ten tabs of research, and lose the thread. Here, I could ask questions and follow a concept all the way through in one place.</p><p>I still read articles, looked through documentation, and did my own research, but AI helped me get unstuck much faster. Instead of spending hours trying to piece together an unfamiliar concept, I could ask about it, dig into the answer, and then go read the relevant sources when I needed more context.</p><p><strong><em>Quick takeaway:</em></strong><em> AI can remove a lot of the friction from learning something unfamiliar. Use LLMs to map out unfamiliar concepts rapidly, but always anchor your understanding in primary documentation and real-world testing.</em></p><p>Before this project, I mostly worked on the data science side of things. Building this app gave me a chance to get involved in the parts I hadn’t really worked on before, from figuring out how to structure the app to dealing with API limitations and making sure the results were actually usable.</p><p>Going through this process myself gave me a much better view of things from an engineering perspective. I got to deal with all the little things that you don’t necessarily see when you’re only thinking about the product, API limitations, messy data, edge cases, things breaking for reasons you didn’t expect.</p><p>I think that experience will be useful going forward. When I’m working on an idea or talking to engineers, I’ll have a better sense of what’s happening on the other side and what I need to think about upfront.</p><p>The code’s already up if you want to poke around before the next post: [<a href="https://github.com/AkshataKini1/user-feedback-triage-pipeline">GitHub link</a>]. I’ll post the full technical breakdown next, follow if you want it in your feed. And if you’ve built something similar, or are thinking about it, I’d like to hear your version of it.</p><p>I enjoy sharing the things I build and the things I learn along the way. Hopefully some of this is useful to you too. If you have questions or want to share how you’ve approached something differently, drop a comment. You can also find me on <a href="https://www.linkedin.com/in/akshatakini/">LinkedIn</a> if you want to stay in touch.</p><p>Happy Building!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=660d70e76fc7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 5 Things Business Leaders Get Wrong About Data]]></title>
            <link>https://medium.com/@akshatakini/the-5-things-business-leaders-get-wrong-about-data-974022f8a419?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/974022f8a419</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 21:47:42 GMT</pubDate>
            <atom:updated>2026-10-03T00:32:11.280Z</atom:updated>
            <content:encoded><![CDATA[<p>A simple guide breaking down common misunderstandings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*suezvLMt_5f1u9phPNu12g.jpeg" /><figcaption>Designed By <a href="https://www.freepik.com/free-vector/hand-drawn-flat-design-business-communication-concept_20548463.htm#fromView=search&amp;page=2&amp;position=0&amp;uuid=90166083-6164-4bf4-b872-1a83ad6c38a2&amp;query=business+leaders+data&quot;&gt;Image%20by%20pikisuperstar%20on%20Freepik">Freepik</a></figcaption></figure><p><strong>Misconception #1: More Data Means Better Insights</strong></p><p>This is a very common misconception that more data equals better insights. Not true!</p><p>I often seen leaders think their lack of actionable insights comes from not having enough data. Their solution? Add more sources of data, add more charts, collect more data or track everything under the sun.</p><p>But here’s the catch: More data does not automatically give you clarity. Without clear questions, right structured data and proper analysis you just get noise not insights.</p><p>Do you know, 80% of a data person’s job is defining the problem and cleaning the data? So, the next time you think “we need more data” ask yourself, are we asking the right questions? What are we actually hoping to find? Why is the existing data or chart not providing us what we want? or start from what problem do we want to solve?</p><p><strong>Misconception #2: Data Initiatives Don’t Require Strategy</strong></p><p>What? 😮</p><p>Yes, without strategy you’re just throwing darts in the dark. Just like every other part of the business, data initiatives need strategy so that they are aligned with the business goals.</p><p>Another dashboard will not solve your business problem. A common theme a couple of years back was “We need another dashboard…. we actually need a tool… no we need to build a machine learning model… just hire a data scientist…insights will follow”</p><p>But all of this rarely helps, too often leaders treat data as a side project and realize it very late that it has to be core to business strategy for it to actually provide value. It definitely does not strive without direction. When data work isn’t tied to a clear business strategy, teams end up producing dashboards nobody uses, reports with no audience, and models that solve problems no one actually cares about. It becomes activity without impact.</p><p>What do I mean by core to business strategy, you ask? See if what you’re building is tied to the business goal, why it matters and how will success be measured?</p><p>Without that alignment, even the most advanced tools or talented analysts can’t move the needle.</p><p><strong>Misconception #3: Data Ownership Is Not Important</strong></p><p>Many leaders don’t think data or metrics need ownership for example, who owns the data, ensures data quality, and interprets the results? Yes, down to the understanding of it’s relation to the business area that it belongs to. Having a single owner can bring in clarity, plus it frees up time of the data department to focus on more complex tasks, the responsibility is divided and the metric or data belongs to the actual true owners i.e. the business departments.</p><p>It all comes down to having proper data governance practices in the organization. Without it data initiative can be fragmented, confusing and ineffective.</p><p><strong>Misconception #4: Data Science means Machine Learning or Artificial Intelligence</strong></p><p>NO!! 🙂‍↔️</p><p>Data Science is broad field to understand and study ways to analyze data in small or large amounts to get meaningful insights. That involves applying statistical and mathematical methods to study the data. Whereas machine learning includes building algorithms that help machines or computers to learn from the data without explicit programming at each step. It is a subset of Artificial Intelligence (AI). Now AI, is a broader term again that is built to mimic human intelligence. Advanced AI in future will or can be used to perform physical tasks that traditionally humans would perform.</p><p>Now-a-days with vast amount of available content on this topic things are clearly defined and people know that these are separate fields but companies still mix these in when looking for a candidate. Understanding the difference will help in identifying the right skillset needed for the job and therefore find the right talent.</p><p><strong>Misconception #5: Technology Solves Everything</strong></p><p>There’s a common belief that buying the “right” technology an AI platform, a BI tool, a data warehouse will suddenly unlock better decisions. But technology is only a small part of the investment. I think business leaders know this very well.</p><p>Then what is the misconception, you ask? It comes back to the right problem, sometimes a solution seems to be the right fit for your problem until it isn’t. For example, in the current race to adapt GenAI without really analyzing current business needs is resulting into many projects being stalled.</p><p>People, processes, and culture are equally critical for data to drive business change. You need people who understand how to use the tool, ask the right questions, and translate insights into action. You need processes that ensure data is collected, cleaned, and shared in a way that is consistent. And you need a culture that values experimentation, accountability, and data-driven decision making.</p><p>These are some of the top 5 common misconceptions I have seen around and think can be easily tackled with the right approach and mindset. Is there anything I missed? or anything that you see and are tired of hearing? Let me know in the comments.</p><p>I enjoy sharing the things I build and the things I learn along the way. Hopefully some of this is useful to you too. If you have questions or want to share how you’ve approached something differently, drop a comment.</p><p>If you enjoyed reading this, feel free to give it a clap or follow along. You can also find me on <a href="https://www.linkedin.com/in/akshatakini/">LinkedIn</a> if you want to stay in touch!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=974022f8a419" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Google’s Hiring Assessment: here’s what you need to know]]></title>
            <link>https://medium.com/@akshatakini/passing-googles-hiring-assessment-here-s-what-i-learned-65d458d77040?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/65d458d77040</guid>
            <category><![CDATA[career-advice]]></category>
            <category><![CDATA[hiring]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[job-hunting]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Mon, 17 Nov 2025 05:34:29 GMT</pubDate>
            <atom:updated>2025-12-01T02:05:46.717Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently passed Google’s hiring assessment and wanted to share what the experience was like.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ow03j32BNbbML-Ex3tpJiw.jpeg" /><figcaption>Design by <a href="http://www.freepik.com/">Freepik</a></figcaption></figure><p>A few days after applying for a position at Google, I received an email inviting me to complete the hiring assessment. At first, I wasn’t sure it was real, since I hadn’t heard of this assessment before. Once I confirmed it was genuinely from Google, I was thrilled! 🥹</p><p>Then, I started to do a little digging about the assessment and what it is all about, this article covers things you need to know about Google’s Hiring Assessment (GHA).</p><p><strong>The assessment email…</strong></p><p>I received an email that said the next step in the hiring process is to complete the assessment and I was given four days to complete it. It included the link I can use to access the assessment that was unique to my candidacy. It also had some instructions on how long it will take to complete the assessment and they also mentioned that the candidate will only move forward in the hiring process if they complete the assessment by the given deadline.</p><p>They also had some accommodations support along with contact info for people who need accommodations based on their condition which was great to see.</p><p><strong>My research…</strong></p><p>After researching, I found out that Google had recently started this process around early 2024 to evaluate candidates on their ethics &amp; integrity, collaboration &amp; teamwork and factors like adaptability, accountability etc.</p><p>I read a few articles on Medium, LinkedIn, and Reddit. They gave me a good sense of what the assessment would be like, and everyone mentioned it was pretty straightforward and I found that to be true. Some shared tips to successfully pass the assessment is to have strong sense of what you would do, like choosing strongly agree or disagree. Which made sense because it shows how self-aware you are and what qualities you possess to be a good fit at Google.</p><p>Some opinions were criticizing the assessment as an extra step in the hiring process while others were speaking in support of it. I also found that everyone does not receive the hiring assessment.</p><p><strong>The assessment itself…</strong></p><p>I will not deny it but I was nervous to start the assessment because there were no actual examples of the questions that’ll were going to be included in the assessment. All I took from the research and reading I did was to be extremely focused during the assessment. And that is what I did, I cleared all distractions and was ready to jump right in. I ensured that I read each question very carefully and answered it based on what I would do in a given situation. All of the questions are subjective, so it is best to answer it based on how you would handle a certain situation.</p><p>I went in nervous but the assessment actually turned out to be fun, I found myself looking forward to more questions to be honest. I have to say that the answers you choose also reflects your level of experience and competence. There are two different style of questions, one included situation and then answer in scale (strongly agree to strongly disagree) to assess behavior. The other style included situation and select what you would do given certain conditions to assess judgement. Note you cannot and should not record or take screenshots of the questions to maintain the integrity of the assessment.</p><p><strong>The result…</strong></p><p>After completing the assessment, you will receive the result within 24–72 hours. I received my results, saying that I passed the assessment within 24 hours. That email also included what to expect next in the process.</p><p>Note that passing the assessment does not guarantee that you move forward in the process. It is just an evaluation that helps the hiring team narrow the pool of candidates that fit Google’s culture. Good thing is if you pass, the results of this assessment is good for up to a year / 24 months. So even if you do not move forward in the process, you can still keep applying to other positions and you do not have to retake the assessment again for a year!</p><p><strong>Things to keep in mind…</strong></p><ul><li>Staying focused and minimizing distractions during the assessment is absolutely crucial.</li><li>It is not about preparing but keeping clear mind and answering everything with extreme clarity about what you will do in a situation.</li><li>This assessment is not testing technical knowledge but your soft skills and behavior.</li></ul><p>Overall, I really enjoyed the assessment and think it’s a fair approach. From the recruiting team’s perspective, Google receives thousands of applications from talented people worldwide, and this seems like a thoughtful way to identify the qualities that matter most for success. It’s also a very thoughtfully designed assessment that covers a wide range of topics, giving a well-rounded view of each candidate.</p><p>If this post was useful, I’d love your support — <strong>clap and follow</strong> to stay connected! And feel free to connect with me or follow me on <a href="https://www.linkedin.com/in/akshatakini/"><strong>LinkedIn</strong></a>!</p><p>Got thoughts or questions or want me to cover anything else? Drop them in the comments!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=65d458d77040" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building SQL Intuition: How Analysts Decide Which Query to Use]]></title>
            <link>https://medium.com/data-science-collective/building-sql-intuition-how-analysts-decide-which-query-to-use-048cac5f50d5?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/048cac5f50d5</guid>
            <category><![CDATA[sql]]></category>
            <category><![CDATA[business-intelligence]]></category>
            <category><![CDATA[data-pipeline]]></category>
            <category><![CDATA[data-analyst]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Sun, 16 Nov 2025 08:28:16 GMT</pubDate>
            <atom:updated>2026-01-04T09:19:06.391Z</atom:updated>
            <content:encoded><![CDATA[<h3>Building SQL Intuition: How Data Analysts Decide Which Query to Use</h3><h4>Save this guide as you practice SQL — it’ll help you know what to use and when.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2TGsNDNLIHMWux7dWg9I4Q.jpeg" /><figcaption>Designed by <a href="http://www.freepik.com/">Freepik</a></figcaption></figure><p>Writing SQL queries is one thing, but choosing the right approach for different situations is a skill that develops with experience. When starting out, the focus is naturally on getting results. As the complexity of your work grows — whether in analytics, reporting, or building pipelines — you begin to consider other factors, such as readability, reusability, and query performance.</p><p>When I was new to SQL, I didn’t know there are multiple ways to you can query a table to get the same results. I always found myself asking why choose this over that and when do I use which query. This post breaks down how to build that intuition, so you can move from writing queries to designing them. The intent of this article is to demonstrate when to use each type of query, using analysis questions as context rather than the main focus.</p><p>To demonstrate, we will use the following datasets to run our queries:</p><ol><li>fact_sales</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8XszOZTDSRoR1naSaSIZyA.png" /></figure><p>2. dim_product</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ROTQlBmj40ioL24eA36O8Q.png" /></figure><p>3. dim_customer</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M07K3wGxjnODdVZcdthgWQ.png" /></figure><h4>1. CTE vs. Subquery vs. Temporary Table — When to Use Each</h4><p><strong>Common goal:</strong> All three queries can help you get intermediate results, it can either be used to get a single aggregated value, a single row or a temporary full dataset. The results only exists during a query session and not as a permanent table within the database. The query types above are mainly used to break down complex queries into smaller steps.</p><p><strong>Subquery -</strong> When filtering or aggregating based on another query’s result.</p><p>✅ <strong>Use when: </strong>Logic is short and isolated, result isn’t reused elsewhere and when a join is not possible.</p><p>Q: Get products whose price is above the average product price</p><pre>SELECT product_name, price<br>FROM dim_product<br>WHERE price &gt; (<br>    SELECT AVG(price)<br>    FROM dim_product<br>);</pre><p>❌ <strong>Avoid when: </strong>There is too much nesting as it can hurt the readability and query performance.</p><p>Q: Customers who spent more than the overall average spend.</p><pre>SELECT <br>    c.customer_name,<br>    SUM(s.total_revenue) AS total_spent<br>FROM fact_sales s<br>JOIN dim_customer c <br>    ON s.customer_key = c.customer_key<br>GROUP BY c.customer_name<br>HAVING SUM(s.total_revenue) &gt; (<br>        SELECT AVG(customer_total)<br>        FROM (<br>            SELECT customer_key, SUM(total_revenue) AS customer_total<br>            FROM fact_sales<br>            GROUP BY customer_key<br>        ) t<br>    );</pre><p>There is too much nesting in the second query and it is harder to follow the logic. On top of everything, debugging a complex nested query can be extremely time consuming because you cannot pinpoint which line of code or nested level is causing an error. Therefore, when handling complex queries or building data pipelines that require multiple calculations, use either a CTE or a temporary table. Let’s explore both options now.</p><p><strong>CTE - </strong>When you need to break complex logic into readable steps and you want to use the results of the query multiple times in the same query.</p><p>✅ <strong>Use when: </strong>You want to structure multi-stage queries clearly, and can be used as a substitute for view when you don’t need the table to be stored permanently.</p><p>Q: Rank customers by total spend.</p><pre>WITH revenue_per_customer AS (<br>    SELECT <br>        customer_key,<br>        SUM(total_revenue) AS total_spent<br>    FROM fact_sales<br>    GROUP BY customer_key<br>)<br>SELECT <br>    c.customer_name,<br>    r.total_spent,<br>    RANK() OVER (ORDER BY r.total_spent DESC) AS rank_by_revenue<br>FROM revenue_per_customer r<br>JOIN dim_customer c <br>    ON r.customer_key = c.customer_key;</pre><p>❌ <strong>Avoid when: </strong>You can get the results using a simple join or a subquery. Also, when you have a large dataset and you reference the same CTE multiple times, because the database may scan and compute that large dataset multiple times which can slow things down.</p><p>When I started to build complex data pipelines for reporting, I typically used CTE’s because that’s is what I learned while working with a data engineer. I eventually learned that there are other ways you can get the same result depending on what you want and the set up.</p><p><strong>Temporary Table -</strong>when you need to store intermediate results, index them, reuse across multiple queries, or when that data needs to be modified multiple times.</p><p>✅ <strong>Use when: </strong>Intermediate results are large or reused often. You want to index, update or filter again for performance.</p><p>Q: Rank customers by total spend. (same question as above)</p><pre>CREATE TEMP TABLE revenue_per_customer AS<br>SELECT <br>    customer_key,<br>    SUM(total_revenue) AS total_spent<br>FROM fact_sales<br>GROUP BY customer_key;</pre><p>This creates a temporary table stored only for the current session and is dropped once the session is closed or ended. Now you can query this table as many times as you’d like. Unlike in CTE’s, where you cannot query the same table twice in the same window.</p><pre>SELECT <br>    c.customer_name,<br>    r.total_spent,<br>    RANK() OVER (ORDER BY r.total_spent DESC) AS rank_by_revenue<br>FROM revenue_per_customer r<br>JOIN dim_customer c <br>    ON r.customer_key = c.customer_key;</pre><p>❌ <strong>Avoid when: </strong>The logic is simple or one-off, basically when you know that you will not need to reference this table again and again while you’re working on analyzing something or just doing a quick check. Stick to simpler queries like subquery or CTE or when your environment restricts temp tables (e.g., cloud warehouses like BigQuery).</p><h4>4. GROUP BY vs. Window Functions</h4><p><strong>Common Goal:</strong> Both Group by and window functions help you compute aggregated information (SUM, COUNT, AVG, etc.). But the difference is in how the result is presented. For this comparison, I’ll share examples of the resulting table to show exactly how.</p><p><strong>Group By -</strong> When you want one row per group (e.g., total sales per region).</p><p>✅ <strong>Use when: </strong>You only need aggregated results, and are creating KPI summaries, report roll-ups, totals etc.</p><p>Q: What is the total revenue per product?</p><pre>SELECT<br>    product_key,<br>    SUM(total_revenue) AS product_total_revenue<br>FROM fact_sales<br>GROUP BY product_key;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/359/1*YcMDAB7FelUPqwUUmzl3Xw.png" /><figcaption>Group By query result</figcaption></figure><p>❌<strong> Avoid when:</strong> You need both the original rows AND aggregated values together. Also avoid if you need running totals, ranks, moving averages, or percentiles or if you want partial detail within a group. Because a group by reduces the table or results to an aggregate value per row of dimensions. When group by is applied to the tables, it will isolate each row and then aggregate a value. For example, if I include sale_id, product_key, and total_revenue in GROUP BY, the query will treat each row as its own group. So instead of getting the total product revenue ($45,000), SUM(total_revenue) will return the row’s own value (i.e., $500).</p><p><strong>Window Functions -w</strong>hen you want both row-level and aggregated data together and want to keep all rows in the resulting table (e.g., each order + total sales for that region).</p><p>✅<strong> Use when: </strong>You want row-level data + aggregated metrics side by side or when you need ranking or time-based analysis. You want flexible grouping without collapsing the table.</p><p>Q: For each sale, what is the product’s total revenue?</p><pre>SELECT<br>  sale_id,<br>  product_key,<br>  total_revenue,<br>  SUM(total_revenue) OVER (PARTITION BY product_key) AS product_total_revenue<br>FROM fact_sales;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vwsMEKUtIc-Xbdb6FCTPTg.png" /><figcaption>Window Function query result</figcaption></figure><p><strong>❌ Avoid when: </strong>You need output collapsed into one row per group or when you need strict aggregations before joins or filters. Let me explain this, when you look at the Window function query result table you’ll see that product_key is repeating per sale_id, when you use this resulting table to join with another table you will get multiple duplicate rows. In such a case it is helpful to aggregate using group by first and then join other tables. Also, in cases where memory preservation is important to a situation window functions don’t help because they preserve all rows with details and don’t reduce the data size.</p><h4>5. Case When vs. JOINing Lookup Tables</h4><p><strong>Common goal:</strong> Both Case When and Joins are used to map values into categories, labels or business rules. For example, creating a table with customer segments, or adding promotion type labels depending on the business logic. The method you choose really depends on the complexity or the available information that lives in the database, exactly what the examples below help explain.</p><p><strong>CASE WHEN -</strong> For small conditional logic embedded in a query.</p><p>✅<strong> Use When: </strong>Logic is based only on columns inside original table that you’re working with in the case below it’s fact_sales. Also, when rules are simple and stable.</p><p>Q: Classify each sale as “High Discount” or “Low Discount” using only fact_sales.</p><pre>SELECT<br>    sale_id,<br>    product_key,<br>    discount_amount,<br>    CASE <br>        WHEN discount_amount &gt; 20 THEN &#39;High Discount&#39;<br>        ELSE &#39;Low Discount&#39;<br>    END AS discount_flag<br>FROM fact_sales;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R8-iW4cNOfMZxEcul0t5FA.png" /><figcaption>Case When query result</figcaption></figure><p><strong>❌ Avoid when: </strong>There are many case when statement needed for a lot of categories or segments, in that case it is best to create a new look up table instead of a lot of adding a lot of case when in the query and then join the tables to get the results. Also, if you realize that the categories are going to keep changing in the future go for creating a new look up table and use join function.</p><p><strong>JOIN to Lookup Table -</strong> When categories or mappings come from another table (or if the categories may change in future).</p><p>✅<strong> Use when: </strong>Your mapping values that exists in a dimension table or if categories change over time.</p><p>Q: Add product_name and category to each sale using dim_product.</p><pre>SELECT <br>    fs.sale_id,<br>    fs.total_revenue,<br>    dp.product_name,<br>    dp.category<br>FROM fact_sales fs<br>LEFT JOIN dim_product dp<br>    ON fs.product_key = dp.product_key;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/988/1*NURVArlOiIvYtqtujnqU2Q.png" /><figcaption>Join query results</figcaption></figure><p><strong>❌ Avoid When: </strong>The logic is small, and static. That means you’re only categorizing rows using same table and does not require another table to look up values.</p><h4>7. UNION vs. JOIN</h4><p><strong>Common goal:</strong> When you want to merge two tables together to get more robust or complete data for your analysis.</p><p><strong>UNION -</strong> When combining results of similar tables but vertically i.e. stack one table on another to merge tables.</p><p>✅<strong> Use when: </strong>You want to stack rows vertically and you want to consolidated list from unrelated tables. The tables you want to merge together have the same structure like same number of columns, and they’re in the same order plus columns have the same names.</p><p>Q: Create a single list of all product categories AND all customer segments.</p><pre>SELECT category AS label, &#39;Product Category&#39; AS source<br>FROM dim_product<br><br>UNION<br><br>SELECT customer_segment AS label, &#39;Customer Segment&#39; AS source<br>FROM dim_customer;</pre><p>These come from two different tables with the same-type column and do not have a common column that can be used as a key.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/769/1*w7K-l7U_guj9kDAuNGGvuQ.png" /><figcaption>Union query result</figcaption></figure><p><strong>❌ Avoid when: </strong>You need to combine data side-by-side (use JOIN instead) or when the tables are related by keys. Also, when you want to preserve duplicates rows using Union ALL is best because a Union will automatically drop duplicate rows.</p><p><strong>JOIN -</strong> when combining related tables horizontally using relationships within the tables.</p><p>✅<strong> Use when: </strong>Tables are related via foreign keys or you need to add attributes (category, customer name, product name) to the original table. And you want to combine the data horizontally.</p><p>Q: Show each sale along with product_name and customer_segment.</p><pre>SELECT <br>    fs.sale_id,<br>    fs.total_revenue,<br>    dp.product_name,<br>    dc.customer_segment<br>FROM fact_sales fs<br>LEFT JOIN dim_product dp <br>    ON fs.product_key = dp.product_key<br>LEFT JOIN dim_customer dc<br>    ON fs.customer_key = dc.customer_key;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yWUFQgKLp-bk6qqsffIEQw.png" /><figcaption>Join query result</figcaption></figure><p><strong>❌ Avoid when: </strong>The tables you’re trying to join are unrelated or you only need to stack lists or categories on top of one another.</p><p>I often questioned my choice of SQL functions and noted unique challenges to remember the rationale behind them. This inspired me to write this comparison article as a guide for others.</p><p>I love sharing what I learn, and writing is the most natural way for me to pass those insights on. Got thoughts or questions or want me to cover anything else? Drop them in the comments!</p><p>I’d love your support, <strong>clap and follow</strong> to stay connected! And feel free to connect or follow me on <a href="https://www.linkedin.com/in/akshatakini/"><strong>LinkedIn</strong></a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=048cac5f50d5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science-collective/building-sql-intuition-how-analysts-decide-which-query-to-use-048cac5f50d5">Building SQL Intuition: How Analysts Decide Which Query to Use</a> was originally published in <a href="https://medium.com/data-science-collective">Data Science Collective</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What I Learned About Space, Data, and Collaboration Through the Women in Data™️Datathon]]></title>
            <link>https://medium.com/@akshatakini/what-i-learned-about-space-data-and-collaboration-through-the-women-in-data-%EF%B8%8Fdatathon-478f61329eb9?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/478f61329eb9</guid>
            <category><![CDATA[hackathons]]></category>
            <category><![CDATA[space-exploration]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[women-in-tech]]></category>
            <category><![CDATA[nlp]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Sat, 11 Oct 2025 19:09:16 GMT</pubDate>
            <atom:updated>2025-11-16T08:23:26.695Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*anQviJgPUhjcRRtV-AJAdw.jpeg" /><figcaption>Image Source: Canva</figcaption></figure><p>🌌 Automating Space Operations with AI, ML, and Interactive Dashboards</p><h4>1. Introduction 🚀</h4><p>I got the chance to participate in the Women In Data™️Space Situational Awareness Datathon. I joined the Datathon mostly out of curiosity but it opened up an entirely new world for me, one that’s incredibly important and inspiring. Before the hackathon, I didn’t know much about how space is managed and it never occurred to me that keeping track of what’s up there is such a complex and critical task.</p><p>According to a <a href="https://www.iiss.org/globalassets/media-library---content--migration/files/research-papers/2025/04/1-enabling-responsible-space-behaviours-through-space-situational-awareness.pdf">report</a> by the International Institute for Strategic Studies, improving SSA and data sharing capabilities would encourage responsible behavior when it comes to space activities. And lack of knowledge of space can perpetuate the risk of collisions. I read <a href="http://satmagazine.com/story.php?number=1149658926">this article</a> and learned how space is a global responsibility because it has global implications.</p><p>To really understand the problem, we looked at it from multiple angles — explored stakeholder pain points, data governance challenges, current analysis methods in space situational awareness, and even country-specific approaches that revealed political complexities. As we dug into the research, we realized that -</p><ul><li>Space is filling up fast, making operations unsustainable; there are currently 15,000 active and inactive satellites (<a href="https://ts2.tech/en/mind-blowing-nearly-15000-satellites-are-whizzing-around-earth-right-now-find-out-why-it-matters/">TS2 Space, 2025</a>), 1.2 million debris objects large enough to to cause catastrophic damage (<a href="https://www.microwavejournal.com/articles/43891-space-situational-awareness-issues-and-challenges">Microwave Journal, 2025</a>). Also, 18,500 small satellites will launch by 2033, driven by mega-constellations (<a href="https://payloadspace.com/euroconsult-releases-smallsat-industry-report/">Euroconsult, 2022</a>).</li><li>A vast amount of data exists, but issues with data sharing hinder standardization.</li><li>Even with all the advances in analytics, approximately more than 90% of collision alerts turn out to be false alarms, leaving operators fatigued and overwhelmed. We saw the same trend in the reports/articles published by major space agencies like NASA, ISRO and ESA from various countries.</li></ul><p>So, as a team, we set out to design something that could make orbital risk analysis simpler and more scalable.</p><p>None of this would have been possible without the amazing people I got to work with. Let me introduce the Astrolytics team.</p><h4>2. Meet the Team 🤝</h4><p>Astrolytics was built on collaboration across disciplines, each member brought a unique strength that made our solution more complete.</p><ul><li><strong>Akarshi R. [AI Engineer]</strong> — went above and beyond to ensure we had access to clean data, built a <em>“what-if” simulator and weekly bulletin</em>, and even developed an <strong>AI NLP chatbot</strong> that answers collision questions in plain English.</li><li><strong>Anusha J. [Data Scientist]</strong> — transformed complex data into a <strong>Tableau dashboard</strong>, making our findings intuitive and visually accessible, while walking the team through each chart so we all understood the insights.</li><li><strong>Monique F. [Business Analyst]</strong>– served as the team’s <strong>stakeholder advocate</strong>, keeping us grounded and focused on real-world needs and operational impact.</li><li><strong>Ylyshume B. [Data Governance Specialist &amp; Presentation designer]</strong> — dove deep into <strong>space data governance methods</strong>, ensuring we covered the global space data governance best practices.</li><li><strong>Akshata K. [Lead &amp; Project Manager / Researcher]</strong> — gathered insights from research papers and organizational reports to study both <strong>country and solutions level challenges</strong>. Worked alongside the team to bring each discipline together.</li></ul><h4>3. Our Approach &amp; Solutions💡</h4><p><em>Our approach was built around three core ideas:</em></p><ul><li><strong>Data diversity:</strong> we utilized <em>five different datasets</em> to build out our solution — 1. Debris shape data from <a href="http://space-track.org/"><strong>Space-Track</strong></a> , Satellite conjunction data from <a href="http://celestrak.org/"><strong>CelesTrak </strong></a>showing how close satellites and objects come. Satellite General Information (SGI) contains key identifying and operational details about each satellite, State vectors contains exact position and velocity of a satellite at a specific moment in time, and ELSET data for deeper orbital analysis from the <a href="https://unifieddatalibrary.com/"><strong>Unified Data Library (UDL)</strong>.</a></li><li><strong>Governance-first thinking:</strong> since space data affects everyone, consistency and transparency had to come first.</li><li><strong>Automation with accessibility:</strong> we wanted tools that reduced manual work but still felt intuitive to use.</li></ul><p><em>Our Solution Highlights:</em></p><ul><li><strong>Weekly Collision Risk Bulletin:</strong> summarized key orbital risks and patterns.</li><li><strong>AI Chatbot:</strong> answered collision-related questions in plain English.</li><li><strong>What-if Simulator:</strong> helped explore trade-offs and scenario outcomes interactively.</li><li><strong>Tableau Dashboard:</strong> brought data, metrics, and visuals together in one view.</li></ul><h4>3. Encouraging Words from the Judges 🌠</h4><p>We were humbled by the judges’ feedback. One comment that stood out said -</p><blockquote>“Used several datasets, great stakeholder analysis, clear solution framing, absolutely loved the chatbot and considered data governance.”</blockquote><p>Another judge noted:</p><blockquote>“Superb use of team member expertise in each discipline, as evident in the professionalism and with advanced analytical skills displayed throughout.”</blockquote><p>Their recognition meant a lot because we intentionally designed Astrolytics to be multidisciplinary, to look at space safety from technical, operational, and governance angles. I’m grateful that this approach came through in our presentation.</p><h4>4. Reflections and Lessons Learned ✨</h4><p>The Space Situational Awareness Datathon was more than just a hackathon, it was a great learning experience. Here are some reflections from my experience -</p><ul><li>Our diverse backgrounds <strong>added layers to how we approached the problem</strong>, helping us see things we might’ve missed otherwise.</li><li>I learned that in Space Situational Awareness it is best to apply Governance methods to raw data itself, this creates a transparent, <strong>shareable, and trusted data</strong> source for analysis and exploration because there is a lack of standardized measurements and it varies from organization level to country level.</li><li>When preparing the presentation, we ended up <strong>trimming the content heavily </strong>a lot because there was a 5-minute time limit and five of us presenting. One minute per person seemed reasonable, but it goes by fast. So in future I would finalize the structure before building the deck to include key points.</li><li>For teams that built tools or models, I later learned that it’s helpful to <strong>include a GitHub link </strong>along with the final video demo or presentation. It allows judges to explore the work more deeply.</li></ul><p>Although we didn’t win, we brought together our team’s strengths and expertise to make a difference. I’m so grateful of the work our team produced within those four weeks. Coordinating across 4 different time zones to finalize this project was a huge effort, and I’m incredibly grateful for everyone’s hard work and contribution. Finally, a massive thanks to the organizers at Women In Data™️for putting on such a valuable event and letting us explore a whole new world in Space Situational Awareness!</p><p>For curious problem solvers, diving into this space (no pun intended) lets you dig into real-world challenges and play a part in something vital. Do explore this further and build. Let me know if you uncover something new!</p><p>If you enjoyed this post, I’d love your support — <strong>clap and follow</strong> to stay connected! And feel free to connect with me or follow me on <a href="https://www.linkedin.com/in/akshatakini/"><strong>LinkedIn</strong></a>!</p><p>Got thoughts or questions or want me to cover anything else? Drop them in the comments, I’d love to hear your perspective!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=478f61329eb9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Getting Started with Natural Language Processing (NLP): The Absolute Basics]]></title>
            <link>https://medium.com/@akshatakini/getting-started-with-natural-language-processing-nlp-the-absolute-basics-aa23a10257d9?source=rss-f7237c9e16f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa23a10257d9</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <dc:creator><![CDATA[Akshata K]]></dc:creator>
            <pubDate>Wed, 09 Apr 2025 03:32:28 GMT</pubDate>
            <atom:updated>2025-10-14T01:02:06.462Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f0Q29-rfBTKJSv4C6RiU3g.png" /><figcaption>Image Generated with OpenAI ChatGPT</figcaption></figure><p>Every business collects text data in some shape or form — emails, customer reviews, support tickets, speech data, survey responses often in huge quantities. On top of that, there’s a wealth of information buried in the company’s internal documentation like reports, invoices and employee records.</p><p>But how often do we actually use this data effectively and know how to leverage it? That’s where Natural Language Processing (NLP) steps in, providing tools and methods to help machines make sense of messy text data and turn it into valuable insights.</p><h4>So, what is NLP?</h4><p>NLP- Natural language processing is a subset of Artificial Intelligence where machines learn human language to be able to interpret, manipulate and generate it.</p><p>You’ve already seen some NLP in action:</p><ul><li>Email filtering your spam messages</li><li>Auto-suggestion in your search sentences in Google search</li><li>Grammar checks for you writing</li><li>Voice assistant like Siri or Alexa</li><li>Customer reviews summarization on Amazon or Walmart’s review section</li></ul><h4>Let’s explore some key concepts &amp; terms in NLP</h4><p>Now that we have a basic understanding of NLP, let’s take a closer look at some key concepts &amp; terms you will come across when exploring this topic further.</p><p><strong>Text pre-processing</strong></p><p>These are basic cleanup to prepare text for analysis</p><ul><li>Tokenization: Splitting sentences into smaller chunks called ‘tokens’, it can be either words or characters. (e.g. sentence tokens [‘I’, ‘love’, ‘data’] or character tokens [‘I’, ‘l’, ‘o’, ‘v’, ‘e’, ‘d’, ‘a’, ‘t’, ‘a’])</li><li>Removing stop words: Filtering out common words like “and” or “the”</li><li>Lemmatization: Reducing their words to their root word. (e.g. ‘walking’ -&gt; ‘walk’’)</li></ul><p><em>🌟 Why it matters:</em> pre-processing helps machines prepare the data for effective analysis</p><p><strong>Representing Text for Machines</strong></p><p>Finally to do anything useful, machines need numbers, not words:</p><ul><li>Bag of Words(BoW) / TF-IDF — These methods help quantify the texts in a document. (e.g. BoW is the frequency of each word present in the document, a very simplistic approach whereas TF-IDF is the number of documents divided by frequency of the word).</li><li>Word Embeddings (like Word2Vec or BERT) — Smarter numerical representations that capture meaning (e.g. Word2Vec learns static vector representation for each word to generate clusters of similar vectors (words), so “bank” will have same vector value and BERT is a model that generates vectors based on the surrounding words for context, meaning “bank” will be different vector value in context of a river and “bank” will be a different vector value in context of a financial institute)</li></ul><p><em>🌟 Why it matters:</em> The next step in the model is to turn text into numbers so models can learn from it and accurately analyze the data.</p><p><strong>Understanding Meaning</strong></p><p>These methods help machines extract valuable insights from raw text:</p><ul><li>Named Entity Recognition (NER) : Identifies entities like name, people, organizations, locations or dates in a document or a block of text. (e.g. for “Google’s headquarters Mountain View, California” -Identify “Google” as an Organization or “Mountain View” as location (city), California” as location (state)</li><li>Keyword Extraction: Extracts specific keywords present in a single document and is useful to summarize long texts by using the extracted keywords. (e.g. for a document on AI, keyword extraction might pick key terms like “NLP”, “Machine Learning”, “Data Science”)</li><li>Sentiment Analysis: Understands the emotional tone (positive, negative, neutral)</li></ul><p><em>🌟 Why it matters:</em> These help businesses understand what people are talking about and how they feel.</p><p>Some common examples of real-world use cases:</p><ul><li><strong>Retail: </strong>Analyze product reviews to understand customer sentiments. Customer reviews are a goldmine to understand what your customers want and their pain-points.</li><li><strong>Finance:</strong> Analyze communications records to identify suspicious communications to detect fraud.</li><li><strong>Healthcare: </strong>Extracting key symptoms from patient notes for diagnosis or summarizing patient history from their records for easy access.</li></ul><p>If your team produces text data or unstructured data (a topic for another time), NLP can help. Remember when I mentioned earlier that companies collect this data in huge quantities? Yep, NLP helps us automate the analysis, so that we get crucial insights faster!</p><p>Now that you have a basic understanding of NLP and what it can do, how do you plan to apply this in your own work? Whether you want to analyze customer feedback or improve operational efficiency to speed up decision making, the possibilities are endless. I’d love to hear your thoughts on NLP applications in your field, feel free to share it in the comments.</p><p>If you enjoyed this post, I’d love your support — feel free to connect with me or follow me on <a href="https://www.linkedin.com/in/akshatakini/"><strong>LinkedIn</strong></a> and don’t forget to <strong>clap and follow</strong> to stay connected!</p><p>Got thoughts or questions? Drop them in the comments, I’d love to hear your perspective!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa23a10257d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>