<?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 Aashna Kunkolienker on Medium]]></title>
        <description><![CDATA[Stories by Aashna Kunkolienker on Medium]]></description>
        <link>https://medium.com/@aashna-kunk?source=rss-37cb0bd5899b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*IUNR4UtASFReAz6oRFjUhg.png</url>
            <title>Stories by Aashna Kunkolienker on Medium</title>
            <link>https://medium.com/@aashna-kunk?source=rss-37cb0bd5899b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 22:35:25 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aashna-kunk/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[Building a No-Code Multi-Agent Chatbot With RAG, Guardrails, and Widgets in OpenAI’s Agent Builder]]></title>
            <link>https://aashna-kunk.medium.com/using-openais-agent-builder-to-build-a-no-code-multi-agent-chatbot-with-rag-guardrails-and-186f1734d874?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/186f1734d874</guid>
            <category><![CDATA[gpt]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[agent-builder]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 01:07:21 GMT</pubDate>
            <atom:updated>2025-11-02T01:16:28.986Z</atom:updated>
            <content:encoded><![CDATA[<p>I usually build agents through code, but I wanted to see what a no-code approach could accomplish. OpenAI’s new <strong>Agent Builder</strong> promises a way to create multi-agent workflows, connect tools, add Guardrails, and ground data: all without writing a line of code. I was curious how far it could go, so I decided to build something end-to-end inside the Agent Builder and see what the no-code experience feels like.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9P5YEQxHd_xX7dIsVYLhGQ.png" /><figcaption>Workflow overview</figcaption></figure><p>As an experiment, I decided to build a <strong>multi-agent assistant</strong> for a musical instruments company, an internal tool that employees could use to check HR policies, resolve IT issues, explore product details, review customer interactions, and summarize internal data.<br> I also integrated it with Gmail to retrieve past customer conversations and added Guardrails to ensure that any sensitive information remains secure.</p><p>In this walkthrough, I’ll show exactly what I built:</p><h3>1. The Orchestrator</h3><p>The <strong>Orchestrator</strong> is the brain of the entire setup. Its only job is to understand what the user is asking and route the query to the right agent.</p><pre>Start → Orchestrator → If/Else → {HR, IT, PRODUCT, CS, KB} → End</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H5Nv-dEZCoEVdcf6SxujVg.png" /><figcaption>orchestration based on intent</figcaption></figure><p>Here’s how it works:</p><ul><li>The user’s input first goes to the Orchestrator.</li><li>The Orchestrator classifies the intent into one of five categories — <strong>HR</strong>, <strong>IT</strong>, <strong>Product</strong>, <strong>Customer Success</strong>, or <strong>Knowledge Base (KB)</strong>.</li><li>The <strong>If/Else</strong> node then reads that label and passes the query to the corresponding agent.</li></ul><p>In each branch, I used a simple condition such as:</p><pre>input.output_text) == &quot;HR&quot;</pre><p>This setup became the backbone of the whole system.</p><h3>2. Building the Agents</h3><p>Once routing worked, I created the five agents:</p><p>1. The <strong>HR Agent</strong> handles HR policy and leave-related questions.<br>2. The <strong>IT Agent</strong> executes (dummy for now) local tools such as reset_password and create_ticket for support requests.<br>3. The <strong>Product Info Agent</strong> retrieves product details from a file and outputs structured JSON that renders through a widget.<br> 4. The <strong>Customer Success Agent</strong> manages customer interactions and can trigger Gmail MCP for follow-ups. (with user approval) <br> 5. Finally, the <strong>Knowledge Base (KB) Agent</strong> summarizes internal or confidential data, supported by File Search grounding and Guardrails for safety.</p><p>To make it more concrete, here’s the exact instruction prompt I used for the <strong>Product Info Agent</strong>.</p><pre>You are a Product Information Agent.<br><br>You must ALWAYS use the attached File Search tool named &quot;product_data&quot;<br>to look up product information. Do not rely on your own knowledge.<br><br>When a user asks about a product:<br>1. Search the &quot;product_data&quot; file for the product name or category.<br>2. Extract the following fields:<br>   - name<br>   - dimensions<br>   - color<br>   - price<br>3. Output ONLY valid JSON in this format:<br>{<br>  &quot;name&quot;: &quot;string&quot;,<br>  &quot;dimensions&quot;: &quot;string&quot;,<br>  &quot;color&quot;: &quot;string&quot;,<br>  &quot;price&quot;: &quot;string&quot;<br>}<br>If no match is found, use placeholder values (e.g., &quot;Not available&quot;) <br>and do not include any text outside the JSON.</pre><h3>3. Adding Widgets for Structured Output</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*iqOBmPzyYiojDdqa_4z5dQ.png" /><figcaption>Widget outputs</figcaption></figure><p><strong>This part was fun.</strong></p><p>I wanted my Product Info Agent to return structured data instead of paragraphs, so I used the <strong>Widget Builder</strong>.</p><h3>For the widget, I did this:</h3><ul><li><strong>Widget Builder → Gallery -&gt; selected a card I liked</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S7hg9jk0hU04ELaVlw44CQ.png" /><figcaption>Widget Builder UI</figcaption></figure><p>The following fields were auto-populated and customizable:</p><ul><li>A schema with four keys: name, dimensions, color, price</li><li>UI components: Card, Heading, Text, Row, Divider</li></ul><p>Once I was happy with how it looked, I had to just download it as a .widget file.</p><p>Then, in my <strong>Product Info Agent</strong>, I set the <strong>Output Format → Widget → Piano Product Card</strong>.<br> Now whenever the agent outputs valid JSON matching that schema, it renders automatically as a clean UI card.</p><h3>4. Connecting File Search (Grounding / RAG)</h3><p>To make it grounded, I uploaded a dummy product catalog file:</p><pre>name,dimensions,color,price,availability,description<br>Aurora 161GP Grand Piano,&quot;150 × 161 × 101 cm&quot;,Ebony,&quot;$120,999&quot;,In stock,&quot;Concert-grade grand piano.&quot;<br>Celesta 49 Digital Piano,&quot;130 × 45 × 90 cm&quot;,White,&quot;$3,499&quot;,Pre-order,&quot;Compact digital piano.&quot;</pre><p>I attached it to my Product Info Agent through the <strong>File Search tool</strong>. (<strong>Add Tool → File Search, u</strong>ploaded the file as product_data, Updated my prompt to <strong>always use that file</strong> before answering)</p><p>When I tested: “Do we have any acoustic guitars?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gzd0vRrgdXYG-9lkmZHBDg.png" /></figure><p>The agent retreived information about a relevant product from the vector store attached and the widget immediately rendered it into a product card.</p><h3>5. Adding Guardrails and Approvals</h3><p>Next, I wanted to show responsible automation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/858/1*_mePXqu0WKDWY6BFAr8cGQ.png" /><figcaption>Guardrails &amp; User approval</figcaption></figure><p>The builder allows you to add “<strong>Guardrails nodes”. </strong>I added one after my KB Agent which helped check for PII and hallucinations before giving any response.</p><p>For the Customer Success agent, I enabled<strong> user approval</strong> for the Gmail connector call.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/884/1*Qk8NM9IyPlwGTxWwNC3Btw.png" /><figcaption>user approval before accessing emails</figcaption></figure><p>Now, if the agent tries to send an email or reset a password, it pauses and asks for confirmation first.</p><h3>6. Watching the Logs</h3><p>One of my favorite parts of the Builder is that it logs every step.<br> For each run, I can expand the logs (here, at “resp_0b240a0a”) and view:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/1*Bh0v5o9F7L80Acg3B3JnYQ.png" /></figure><ul><li>What the Orchestrator predicted (PRODUCT, KB, etc.)</li><li>What each agent said or which tool it called</li><li>When Guardrails triggered</li><li>How the final widget JSON looked before rendering</li></ul><h3>7. Some issues (and Fixes)</h3><ul><li><strong>Prompt changes sometimes don’t save: </strong>do NOT refresh, you’ll lose your work. It doesn’t auto-save too often. Publish your flow and then perform a hard refresh.</li><li><strong>Widgets crash silently</strong> if the JSON isn’t valid. Use strict quotes and validate in the preview.</li><li><strong>File Search grounding sometimes skips</strong> if your file doesn’t contain the term exactly. Add generic fallback rows (e.g., “Generic Piano”).</li><li><strong>Temperature matters: </strong>keep it between 0.0 and 0.3 for consistent JSON outputs.</li><li><strong>Refresh often: </strong>changes to tools or widgets may only take effect after a hard reload.</li></ul><p>Despite these quirks, it’s really fascinating how much can be built in one interface.</p><h3>8. Deploying or Embedding the Workflow</h3><p>It can be easily integrated into an actual application.<br>OpenAI provides <strong>ChatKit</strong>, a UI component that lets you embed any published workflow directly into your product using a <strong>workflow ID</strong>.<br> You simply copy the ID from the Agent Builder dashboard, paste it into your app, and the entire multi-agent system: including widgets, Guardrails, and tools, runs behind a chat interface that you control.</p><p>This makes it easy to turn a no-code prototype into a usable internal tool or a customer-facing assistant without managing infrastructure or any API calls manually!!</p><h3>9. Who Should Try This</h3><p>This setup is perfect for anyone who wants:</p><ol><li>A low-code chatbot that combines RAG and orchestration.</li></ol><p>2. Quick prototypes without building APIs or UIs.</p><p>3. Flexible logic, file grounding, and simple moderation built in.</p><p>It’s not for complex, deeply integrated systems yet, but it’s excellent for internal copilots, product demos, or rapid experiments.</p><h3>10. Conclusion</h3><p>This project began as an experiment in OpenAI’s new Agent Builder and, within three hours, became a simple orchestrated, widget-enabled assistant, though on dummy data. It routes, grounds, visualizes, and moderates, all without a single line of code. There are still rough edges (saving, refreshing, connector stability), but the flexibility and speed make Agent Builder an impressive step toward true no-code AI development.</p><p><strong>P.S.</strong> I’ll also be experimenting with <strong>AgentKit</strong> next to see how the code-based experience compares with what I’m used to in LangChain and Google’s ADK.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=186f1734d874" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Practical Guide to the NVIDIA AI Infrastructure & Operations Exam [NCA-AIIO]]]></title>
            <link>https://aashna-kunk.medium.com/a-practical-guide-to-the-nvidia-ai-infrastructure-operations-exam-nca-aiio-804f445a528c?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/804f445a528c</guid>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-infrastructure]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[nvidia]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Wed, 29 Oct 2025 22:48:43 GMT</pubDate>
            <atom:updated>2025-10-29T22:48:43.421Z</atom:updated>
            <content:encoded><![CDATA[<p>I recently took NVIDIA’s AI Infrastructure &amp; Ops certification out of curiosity about how large-scale AI systems actually run. Here’s everything I learned, and how you can prepare too!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*fbT4-0sDzGNDw2KFdckRog.png" /><figcaption>NCA-IIO Badge</figcaption></figure><h3>Who should take it</h3><p>Anyone who wants to understand <strong>how AI actually runs under the hood, i.e., </strong>the datacenter, the GPUs, the interconnects, the orchestration.<br> If you’re an <strong>AI/ML engineer</strong>, <strong>MLOps engineer</strong>, <strong>data-center admin</strong>, or just curious about what happens behind large models like ChatGPT and Llama, this cert is for you.</p><h3>Why you should take it</h3><p>AI Infrastructure &amp; Ops is <em>the new gold rush</em>.<br> Every major model: OpenAI GPT, Anthropic Claude, Google Gemini, Meta Llama, depends on NVIDIA systems, InfiniBand networks, and orchestration stacks.</p><p>This exam forces you to learn that ecosystem:</p><ul><li>DGX systems / BasePOD / SuperPOD</li><li>NVLink &amp; InfiniBand fabrics</li><li>Slurm vs Kubernetes scheduling</li><li>CUDA → TensorRT → Triton → NIM software stack</li></ul><p>It’s <strong>200 % worth it</strong> ! You come out understanding how LLM inference, optimization, and scaling really work.</p><h3>How long prep takes</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/764/1*8mCX5W5YQ6yHZKOw7okt0w.png" /></figure><h3>My preparation path:</h3><p>I <strong>bought the official course</strong> from NVIDIA Academy and treated it like a lab.<br> Then, honestly, I just talked to <strong>ChatGPT (Advanced Voice Mode)</strong> and spent hours talking through the entire NVIDIA guide, asking <em>why</em> each piece (InfiniBand, DPUs, NVSwitch) exists and <em>how</em> it fits.</p><p>After that, I:</p><ol><li>Skimmed the “by-heart” details (bandwidths, rack kW limits, DGX models).</li><li>Focused deeply on <strong>networking</strong> (NVLink vs InfiniBand, RDMA vs RoCEv2).</li><li>Mapped the <strong>AI software stack</strong> : CUDA → TensorRT → Triton → NIM.</li><li>Learned NVIDIA’s <strong>product families</strong>: DGX Cloud, BasePOD, SuperPOD, Grace Hopper.</li></ol><p>By exam day I could explain the whole AI Ops pipeline out loud, and that was enough.</p><h3>What to expect</h3><ul><li>~50 multiple-choice questions, 60 minutes.</li><li>Topics: GPU architecture, interconnects, virtualization, storage, orchestration, monitoring.</li><li>No math, pure understanding and system reasoning.</li></ul><p>I finished in ~20 minutes. They’re “you either know it or you don’t” questions.</p><h3>The takeaway</h3><p>This certification helped <em>see the full picture</em>:</p><p><strong><em>How data moves from storage -&gt; network -&gt; GPUs -&gt; Tensor cores -&gt; Triton -&gt; the inference endpoint</em></strong></p><p>If you want to speak the language of AI infrastructure, or work anywhere near LLM training and deployment, this exam gives you the vocabulary and intuition to do it.</p><p><strong>Worth it?</strong><br>Absolutely. 200 %.</p><p>AI Infrastructure is where innovation meets reality, and this certification is your backstage pass to that future! :)</p><p>If you have any questions or just want to have a relevant chat, feel free to reach out: <strong>aashnakunk@gmail.com</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=804f445a528c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From plug-and-play to custom agents: The Complete Path to Building Agents with Google’s AI Stack]]></title>
            <link>https://aashna-kunk.medium.com/from-plug-and-play-to-custom-agents-the-complete-path-to-building-agents-with-googles-ai-stack-b190c81d71fb?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b190c81d71fb</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[google-agentspace]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Thu, 09 Oct 2025 21:25:38 GMT</pubDate>
            <atom:updated>2025-10-09T23:15:55.775Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the past few months, Google has been shaping an entirely new way to build intelligent systems: agents that can reason, act, and connect deeply with enterprise data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/432/1*eycKQ6mhoXY99MxsAx_cNg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/548/1*RH5UFuKMVn258IOb7LBk8w.png" /></figure><p>With <strong>Gemini Enterprise (formerly Agentspace)</strong>, and the <strong>Agent Development Kit (ADK)</strong>, google now offers a complete gradient of control: from quick, no-code agents and connectors to fully customized, autonomous applications.</p><p>Here’s what that journey looks like.</p><h3>1. Native connectors in Gemini Enterprise (formerly called AgentSpace)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/658/1*iLoijBSODNU0SX-HWvbb6w.png" /><figcaption>Native connectors on gemini enterprise</figcaption></figure><p>Here, you can launch an agent in minutes by connecting built-in integrations such as Gmail, Drive, Jira, Confluence, ServiceNow, Zendesk, etc.</p><p>No code, no setup. The connectors handle authentication, data access, and response generation automatically.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hztUIirRzSczqhcexqhUIQ.png" /><figcaption>Jira Native connector — gemini enterprise</figcaption></figure><p>It’s a powerful way to make enterprise data conversational, your agent can fetch tasks, search documents, or analyze tickets out of the box.<br>But this simplicity also means limited flexibility. The agent can only operate within the scope of those prebuilt connectors.</p><h3>2. Connect custom data sources to AgentSpace (Gemini Enterprise)</h3><p>The next step is to bring your own data into the ecosystem.<br> With <strong>Vertex AI Search</strong> (built on Discovery Engine), you can ingest PDFs, manuals, reports, or support content into a custom datastore and connect that datastore to your Agentspace agent. ACLs can also be respected while doing this. Refer to: <a href="https://medium.com/@aashna-kunk/agentspace-searches-on-custom-data-sources-respecting-acls-6b1a03fbc833">https://medium.com/@aashna-kunk/agentspace-searches-on-custom-data-sources-respecting-acls-6b1a03fbc833</a></p><p>This setup allows you to extend the agent’s knowledge far beyond Google’s standard connectors. Instead of relying solely on Gmail or Drive, your agent can now search your own corpus and ground responses in context-specific material.</p><p>It’s the bridge between a general assistant and a domain expert, a tailored RAG (retrieval-augmented generation) system that speaks your organization’s language.</p><h3>3. Build custom agents built in ADK and deployed to Agentspace</h3><p>Once you need more than retrieval, when your agent has to reason, plan, or orchestrate multiple tools: that’s where the <strong>Agent Development Kit</strong> (ADK) comes in.</p><p>ADK allows you to define a complete multi-agent system in code.You can specify its reasoning model, the tools it can call (for instance, Snowflake, custom data, or Vertex Search), and the logic behind its responses.</p><p>When you deploy it to Gemini Enterprise (formerly AgentSpace), your organization gets a managed, observable agent that can handle complex workflows while still integrating into Google’s ecosystem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OItivnev-MQOS1b_CXRp1Q.png" /><figcaption>ADK agent — local testing</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E6ywJG__jKZiuiMoy83bYg.png" /><figcaption>A custom snowflake agent built using ADK, deployed on Gemini Enterprise (AgentSpace)</figcaption></figure><p>This is the sweet spot for most enterprise teams: enough customization to build meaningful logic, but still within a governed, deployable environment.</p><h3>4. Fully custom agents: ADK with your own UI</h3><p>The final level of customization removes the limits altogether.<br> Here, you still use ADK as the reasoning and orchestration engine, but you build your own frontend: perhaps a Streamlit dashboard, a React chat interface, or even an internal web app.</p><p>For instance, I built an <strong>interior design assistant</strong> using Google’s ADK: it combines <strong>BigQuery embeddings</strong> with <strong>data scraped from the store’s website</strong> to generate style-based recommendations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PaVwCXUXcVfyyF0s-af3pA.png" /><figcaption>Fully Custom Interface built on ADK</figcaption></figure><p>In this design, the agent lives entirely on your infrastructure. It can call Vertex AI Search directly, integrate with your existing APIs, manage access control, and surface insights in your own branded experience.<br> This is what powers bespoke copilots: systems that not only talk but act on your behalf, tightly integrated with your data and workflows.</p><p><strong>Google’s Gemini Enterprise and Agent Development stack turns what used to be a maze of APIs into a canvas for real intelligence.<br> It’s where agents stop being demos, and start running your business.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b190c81d71fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Observability in Google ADK: How to Watch Your AI Agent Think, Call Tools, and Debug in Real Time]]></title>
            <link>https://aashna-kunk.medium.com/observability-in-google-adk-how-to-watch-your-ai-agent-think-call-tools-and-debug-in-real-time-2c244cc1fcf8?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c244cc1fcf8</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[google-adk]]></category>
            <category><![CDATA[google-agentspace]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Thu, 09 Oct 2025 19:50:40 GMT</pubDate>
            <atom:updated>2025-10-09T19:50:40.902Z</atom:updated>
            <content:encoded><![CDATA[<p>One of the most underrated parts of working with <strong>Google’s Agent Development Kit (ADK)</strong> is its built-in <strong>observability layer, </strong>the ability to watch your agent <em>think</em>, <em>call tools</em>, and <em>transition between states</em> in real time.</p><p>If you’ve ever built or deployed an AI agent with ADK, you’ve seen this:<br> — Events being logged,<br> — Tool invocations showing up as nodes,<br> — State changes lighting up like a flow diagram.</p><p>That’s <strong>agent introspection, </strong>built into the runtime.</p><h3>When You Run Your ADK Agent</h3><p>When you launch your agent via the CLI or web runtime (for example, with adk web), ADK automatically enables its observability stack.<br> Each tool call, reasoning step, and response is tracked as an <strong>event</strong>.</p><p>If you open the <strong>web view</strong> (as shown in your screenshot), every interaction appears as a <strong>node</strong>, representing a specific event.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GNdv370073tDuKNekFZZvw.png" /><figcaption>The ADK web UI + observability dashboard (left)</figcaption></figure><h3>2. Features of the Observability dashboard:</h3><p>The observability dashboard is accessible directly through the <strong>web UI</strong> that ADK launches: available at the <strong>localhost link</strong> displayed when you run your agent.</p><p>We can view every event as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/734/1*1JbkI-SMlC2bO3320gZEtw.png" /><figcaption>Events in your agent’s flow</figcaption></figure><h4>Tool Transparency in Action</h4><p>One of the most powerful aspects of ADK’s observability is <strong>tool transparency: </strong>you can literally see how your natural language request was translated into structured tool calls.</p><p>For example, when I asked my agent a question about credit scores in my s<strong>nowflake</strong> database, it automatically invoked the RunQueryTool.</p><p>In the observability graph, this appeared as a node under my main agent.<br> When I expanded it, I could see the <strong>exact query</strong> the agent generated and passed as a parameter:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/772/1*DI_qaOii2LEbIgC3cy7pgA.png" /></figure><p>You can expand nodes to view payloads, latency, and outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FiFRPflZFKDwzbUsYvsQcA.png" /><figcaption>Latency view</figcaption></figure><h3>Why It Matters</h3><p>Most frameworks only show you text responses.<br> ADK shows you <strong>how</strong> your agent got there.</p><p>This makes debugging multi-tool or multi-agent systems far easier:</p><ol><li>You can confirm whether the right tool was invoked.</li></ol><p>2. Inspect arguments and outputs without extra instrumentation.</p><p>3. Detect unnecessary loops or slowdowns.</p><p>4. Understand exactly which <strong>state transitions</strong> occurred, from user prompt to tool to final response.</p><h3><strong>Next time, instead of watching logs, try running your ADK project with </strong><strong>adk web, and open the observability dashboard to watch every tool call come alive!</strong></h3><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c244cc1fcf8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Respecting ACLs in Google AgentSpace Custom Data Source Searches]]></title>
            <link>https://aashna-kunk.medium.com/agentspace-searches-on-custom-data-sources-respecting-acls-6b1a03fbc833?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b1a03fbc833</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[google-agentspace]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[access-control]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Mon, 29 Sep 2025 19:48:42 GMT</pubDate>
            <atom:updated>2025-09-29T19:51:29.018Z</atom:updated>
            <content:encoded><![CDATA[<p>When connecting custom data sources (BigQuery/Cloud Storage) to AgentSpace, it’s not just about indexing data, it’s about <strong>respecting ACLs (Access Control Lists)</strong> so that searches return only what the requesting user is authorized to see. In this article, I’ll walk through how to set up a custom data store with ACLs enforced, using Google Cloud Storage as the backing store and JSONL files with ACL metadata.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y5BW8pOvjtOd_4qZP1Xx0w.png" /></figure><h3>1. Why ACLs Matter in AgentSpace</h3><p>AgentSpace builds on <strong>Vertex AI Search</strong>. When you connect a custom source (here, the mock VAIsearch datastore), you need a way to tell AgentSpace:</p><ol><li><strong>Which users or groups can see which document.</strong></li></ol><p><strong>2. How identity information (Google or third-party) maps to those ACLs.</strong></p><p>This is done via an acl_info section in your ingested JSONL or structured data. AgentSpace then matches it against the authenticated principal (google.subject, google.group, etc.) provided via Workforce Identity Federation or Google accounts.</p><h3>2. Creating a Data Store with ACL Support</h3><p>From the AgentSpace App console:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1004/1*U6XuAMwTR0f8ZLgMHTAXaw.png" /></figure><ol><li><strong>Create a new data store.</strong></li><li>Select <strong>Google Cloud Storage</strong> as the source.</li><li><strong>Choose the import type.</strong></li></ol><p>Refer to the <a href="https://cloud.google.com/agentspace/docs/identity">AgentSpace Identity documentation</a> for detailed guidance on how to structure your files in Cloud Storage, depending on the selected import type, so that they’re ingested in the supported format.</p><p><strong>4. Select ingestion mode.</strong></p><p>5. <strong>Select your bucket (file/folder) to import</strong></p><p>On the next (configuration) page,</p><p><strong>6. Enable ACL info.</strong><br> This ensures AgentSpace enforces access checks at query time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/702/1*aOo5HwWFYYqEZ-wGxNToNw.png" /></figure><p>⚠Note: Only the <em>structured</em> data with document IDs and <em>unstructured with metadata and media</em> paths expose the <strong>“Enable ACL info”</strong> option. If it’s greyed out for you, your file type does not support ACL enforcement.</p><h3>4. How Queries Respect ACLs</h3><p>When a user signs in via <strong>Workforce Identity Federation (WIF)</strong> or a Google account, AgentSpace maps their identity into:</p><ul><li>google.subject (commonly the email or unique user ID)</li><li>google.groups (if group claims are emitted)</li></ul><p>At query time, AgentSpace filters results:</p><ul><li>If the document’s acl_info contains the user or any group the user belongs to, result is returned.</li><li>Otherwise, the document is excluded.</li></ul><p>This means the same query can return different results for two users, depending on ACLs.</p><h3>5. My Example Setup</h3><p>In my setup:</p><ul><li><strong>Connector</strong>: GCS bucket with JSONL.</li><li><strong>Data type</strong>: Structured JSONL with document IDs enabled.</li></ul><p>Before ingestion, I uploaded my JSONL files into a Cloud Storage bucket. When creating the data store in AgentSpace, I simply pointed the connector to the path of that bucket (or folder) under <strong>“Import data from Cloud Storage”</strong>. This ensured the files were pulled directly into the data store during ingestion.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/892/1*UmJO6A1Rwh37Im1hwYbD3w.png" /></figure><p>What a sample JSONL object looked like in my datastore:</p><pre><br>{&quot;id&quot;:&quot;KB_ART_4002&quot;,<br> &quot;structData&quot;:{<br>   &quot;title&quot;:&quot;Data Retention Policy&quot;,<br>   &quot;article_id&quot;:&quot;KB002&quot;,<br>   &quot;content&quot;:&quot;Company policy on log and data retention for compliance (GDPR, HIPAA).&quot;,<br>   &quot;source_system&quot;:&quot;KB&quot;,<br>   &quot;object_type&quot;:&quot;article&quot;,<br>   &quot;uri&quot;:&quot;https://kb.example.com/articles/KB002&quot;,<br>   &quot;updated_at&quot;:&quot;2025-09-07T15:45:00Z&quot;<br> },<br> &quot;acl_info&quot;:{<br>   &quot;readers&quot;:[{&quot;principals&quot;:[<br>     {&quot;group_id&quot;:&quot;testgroup@example.com&quot;},<br>     {&quot;user_id&quot;:&quot;monark.unadkat@example.com&quot;},<br>     {&quot;user_id&quot;:&quot;aashna.kunkolienker@example.com&quot;}<br>     {&quot;user_id&quot;:&quot;khushi.patni@example.com&quot;}<br>   ]}]<br> }}</pre><p>Here, ACLs are tied to either <strong>user IDs (emails)</strong> or <strong>group GUIDs</strong> from your provider (Google, Entra ID, Okta, etc.). Mine’s Google.</p><p>Once ingestion finished, I tested searches as different users. Queries only returned documents where my email or groups I was a part of matched acl_info.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_laSZoT7kRIkCocsA2BywA.png" /><figcaption>AgentSpace search with ACLs respected in the response</figcaption></figure><h3><strong>6. Common Errors &amp; Debugging Tips</strong></h3><p>Even if your setup looks correct, a few common issues can trip you up:</p><ol><li><strong>Import failures in the datastore</strong></li></ol><ul><li>Symptom: <em>“Import completed issues”</em> or <em>“X items failed to integrate”</em>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/692/1*O1eV0GkYuOWimAV_bwPALQ.png" /></figure><ul><li>Cause: Usually due to malformed JSONL (missing commas, invalid JSON object structure, or missing required fields like id or structData).</li><li>Fix: Double-check your JSONL formatting and schema. Refer to: <a href="https://cloud.google.com/generative-ai-app-builder/docs/data-source-access-control">Data Source Access Control</a> and make sure your data is in the format AgentSpace expects. Use the <strong>View Logs</strong> option in the datastore to pinpoint the exact file or line that failed.</li></ul><p><strong>2. ACLs not being respected</strong></p><ul><li>Symptom: You’ve listed emails in acl_info, but users can’t see their documents.</li><li>Cause: The identity logged into AgentSpace doesn’t match the principal listed.</li><li>Fix: Check the email under your <strong>profile icon in the AgentSpace login, </strong>this is what AgentSpace interprets as google.subject. Make sure it matches the email in your acl_info.</li></ul><p><strong>3. Workforce pool identity integration errors</strong></p><ul><li>Symptom: Third-party identities (like Entra ID) can’t sign in, or can sign in but can’t access documents or ACLs don’t match.</li><li>Fix: These are usually due to subject claim mismatches (sub vs. preferred_username). See my detailed debugging guides:</li><li><a href="https://medium.com/@aashna-kunk/debugging-third-party-identity-provider-entra-id-and-agentspace-integration-why-sub-breaks-141b4792f0bf">Debugging third-party identity provider (Entra ID) and AgentSpace integration: why “sub” breaks</a> ACLs</li><li><a href="https://medium.com/@aashna-kunk/how-to-decode-and-debug-entra-id-tokens-for-workforce-identity-federation-beddefcc186f">How to decode and debug Entra ID tokens for Workforce Identity Federation</a></li></ul><ol><li><strong>Still stuck?</strong></li></ol><p>Feel free to reach out at <a href="/service/mailto:aashnakunk@gmail.com"><strong>aashnakunk@gmail.com</strong></a><strong> </strong>with a detailed description of the setup and error.<strong> </strong>Happy to help debug through a working session!</p><h3>6. References</h3><ul><li><a href="https://cloud.google.com/generative-ai-app-builder/docs/introduction">Configure access controls for custom data sources</a></li><li><a href="https://cloud.google.com/generative-ai-app-builder/docs">Vertex AI Search Data Ingestion</a></li><li><a href="https://cloud.google.com/generative-ai-app-builder/docs/data-source-access-control">Data Source Access Control</a></li></ul><h3>Conclusion:</h3><p>Respecting ACLs isn’t optional, it’s foundational. By structuring JSONL files with acl_info and mapping them to your identity provider correctly, you ensure AgentSpace searches always stay secure and compliant.</p><p>This approach works across connectors. The principle stays the same: <strong>define ACLs at ingestion, enforce at query.</strong></p><p>❤</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b1a03fbc833" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Debugging Third-party Identity Provider (Entra ID) and AgentSpace Integration: Why “sub” Breaks…]]></title>
            <link>https://aashna-kunk.medium.com/debugging-third-party-identity-provider-entra-id-and-agentspace-integration-why-sub-breaks-141b4792f0bf?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/141b4792f0bf</guid>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[entra-id]]></category>
            <category><![CDATA[google-agentspace]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Sun, 28 Sep 2025 08:21:32 GMT</pubDate>
            <atom:updated>2025-09-28T08:28:17.056Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Debugging T</strong>hird-party Identity Provider<strong> (Entra ID) and AgentSpace Integration: Why “sub” Breaks Your ACLs</strong></h3><p>When integrating <strong>Microsoft Entra ID</strong> into <strong>Google Cloud Workforce Identity Federation (WIF)</strong> for <strong>AgentSpace searches</strong>, one of the issues is getting <strong>user identity mapping</strong> right.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YSXakerxUfG2vfyJ4hmWyQ.png" /></figure><p>I ran into this exact problem while setting up ACL-based search access for custom data sources (<a href="https://cloud.google.com/agentspace/docs/identity">https://cloud.google.com/agentspace/docs/identity</a>), and the fix turned out to be surprisingly simple.</p><h3><strong>1. The Setup</strong></h3><p>I configured a workforce pool provider in GCP that trusted Entra ID as an OIDC source.</p><p>Principal definition:</p><pre>principal:principalSet://iam.googleapis.com/projects/123456789/locations/global/workforcePools/my-pool/subject/aashna.kunkolienker@narachi.fun</pre><p>Attribute mapping (initial attempt):</p><pre>google.subject = assertion.sub</pre><p>At first glance, this looked fine. But…</p><h3><strong>2. The Problem: sub is Opaque</strong></h3><p>When Entra issued the ID token, the sub claim <em>(Read how to find what claims your App emits here: </em><a href="https://medium.com/@aashna-kunk/how-to-decode-and-debug-entra-id-tokens-for-workforce-identity-federation-beddefcc186f"><em>https://medium.com/@aashna-kunk/how-to-decode-and-debug-entra-id-tokens-for-workforce-identity-federation-beddefcc186f</em></a><em>)</em> looked like this:</p><pre>&quot;sub&quot;: &quot;CqFwcpXg8QllsnHlRS2HoXypXymAhepxJT8W3DI0Ue4&quot;</pre><p>That value is:</p><ol><li>Unique per app</li></ol><p>2. Not human-readable</p><p>3. Not an email</p><p>But <strong>AgentSpace ACLs use email addresses</strong> for access control, for example:</p><pre>&quot;acl_info&quot;: {<br>&quot;readers&quot;: [<br>{ &quot;principals&quot;: [{ &quot;user_id&quot;: &quot;aashna.kunkolienker@narachi.fun&quot; }] }<br>]<br>}</pre><p>Result? The ACL check failed silently because the google.subject didn’t match any user_id in my ACL JSON file.</p><h3><strong>3. The Fix: Use preferred_username</strong></h3><p>Instead of sub, I switched the mapping to:</p><pre>google.subject = assertion.preferred_username</pre><p>Now my ID token emitted:</p><pre>&quot;preferred_username&quot;: &quot;aashna.kunkolienker@narachi.fun&quot;</pre><p>Now, AgentSpace now saw the subject as an email, which matched my ACLs exactly.</p><h3><strong>4. Why This Matters in AgentSpace</strong></h3><p><strong>AgentSpace uses google.subject as the user identity</strong> for ACL checks.</p><p>If it’s mapped to sub, you’ll get opaque IDs : ACL mismatch.</p><p>If it’s mapped to preferred_username (or email, depending on your tenant), you’ll get the user’s email : ACLs work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7t0cj3csV7a39lDAh7bgPg.png" /></figure><p>So:</p><ol><li>Define ACLs with <strong>email addresses</strong></li></ol><p>2. Map google.subject to the Entra ID claim that emits the email</p><h3><strong>5. Debugging Tip</strong></h3><p>To see what Entra is actually emitting, read:</p><p><a href="https://aashna-kunk.medium.com/how-to-decode-and-debug-entra-id-tokens-for-workforce-identity-federation-beddefcc186f">How to Decode and Debug Entra ID Tokens for Workforce Identity Federation</a></p><p>❤</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=141b4792f0bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Decode and Debug Entra ID Tokens for Workforce Identity Federation]]></title>
            <link>https://aashna-kunk.medium.com/how-to-decode-and-debug-entra-id-tokens-for-workforce-identity-federation-beddefcc186f?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/beddefcc186f</guid>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[google-agentspace]]></category>
            <category><![CDATA[debugging]]></category>
            <category><![CDATA[workforce-identity]]></category>
            <category><![CDATA[entra-id]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Sun, 28 Sep 2025 08:08:52 GMT</pubDate>
            <atom:updated>2025-09-28T08:30:04.314Z</atom:updated>
            <content:encoded><![CDATA[<p>When setting up <strong>Workforce Identity Federation (WIF)</strong> with Microsoft Entra ID (Azure AD) to secure access to Google Cloud or custom apps like Agentspace, one of the crucial steps is understanding <strong>what claims your ID token actually emits</strong>.</p><p>If your ACLs depend on email, preferred_username, or groups, you need to verify those values. Otherwise, you’ll spend hours chasing why access isn’t working.</p><p>This post walks through how I requested an <strong>ID token from Entra ID</strong>, decoded it, and confirmed what claims were included, helping me debug issues with group-based access in Google Cloud.</p><h3>1. Building the ID Token Request</h3><p>To fetch the ID token, I used my Entra app’s OAuth2 authorization endpoint with the following parameters:</p><pre>https://login.microsoftonline.com/&lt;tenant_id&gt;/oauth2/v2.0/authorize?<br>client_id=&lt;client_id&gt;<br>&amp;response_type=id_token<br>&amp;redirect_uri=&lt;redirect_url&gt;<br>&amp;scope=openid profile email<br>&amp;response_mode=fragment<br>&amp;state=12345<br>&amp;nonce=678910</pre><p>Here’s what’s happening:</p><p>client_id :the App Registration in Entra ID connected to Google Cloud.</p><p>response_type=id_token : tells Entra to return an ID token (not an access token).</p><p>redirect_uri :callback configured in my Entra App connected to Google Clooud.</p><p>scope=openid profile email : ensures OIDC fields like sub, preferred_username, and email are included.</p><p>nonce : used for replay protection, must match validation on callback.</p><p>I also added a <strong>Group claim</strong> under the Entra App Registration’s <strong>Token Configuration</strong> in Entra, so I could test how groups were emitted because my ACLs depended on groups as well.</p><h3>2. Decoding the ID Token</h3><p>Once I authenticated, Entra redirected back to my WIF callback with an ID token in the URL fragment.</p><p>Here’s an example of what I saw:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kxadPKeytb0aexpWbSXPqA.png" /><figcaption>ID_token returned by the endpoint</figcaption></figure><h3>3. Inspecting the Token with <a href="https://jwt.io/">jwt.</a>ms</h3><p>Go to <a href="https://jwt.ms/"><strong>jwt.ms</strong></a> (Microsoft’s official JWT decoder).</p><p>Unlike jwt.io, this tool is designed specifically for Microsoft-issued tokens. It parses the claims with Azure AD context, so you see the claims as Microsoft intends them.</p><p>Paste your ID token in the box.</p><p>You’ll immediately get a <strong>decoded view</strong> with two sections:</p><ol><li><strong>Decoded JSON</strong> :the raw payload as JSON.</li></ol><p><strong>2. Claims Table</strong> : a friendly explanation of each claim (with notes specific to Azure AD).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s3OXo4wXyFvaV25dca5A0w.png" /><figcaption>JWT.MS</figcaption></figure><h3>4. Why This Matters for WIF</h3><p>Google Cloud’s Workforce Identity Federation lets you <strong>map OIDC claims into Google attributes</strong>. For example:</p><pre>attribute_mappings:<br>  google.subject: &quot;assertion.sub&quot;<br>  google.email: &quot;assertion.preferred_username&quot;<br>  google.groups: &quot;assertion.groups&quot;</pre><p>If your Entra token doesn’t include groups or email, those mappings will fail. By manually decoding your token, you can confirm exactly what’s present and adjust your mappings.</p><h3>Mapping Entra ID Claims Correctly for WIF</h3><p>When configuring Workforce Identity Federation (WIF) in GCP, one of the most important steps is deciding <strong>which claim from your identity provider (Entra ID in this case) should map to </strong><strong>google.subject.</strong></p><p>At first, I defined my attribute mapping as <strong>google.subject: “assertion.sub”</strong> and my workforce pool principal like this:</p><pre>principal: principalSet://iam.googleapis.com/projects/123456789/<br>locations/global/workforcePools/my-pool/<br>subject/aashna.kunkolienker@narachi.fun</pre><p>But this didn’t work because Entra ID was emitting the subject (sub) as an opaque identifier like:</p><pre>&quot;sub&quot;: &quot;CqFwcpXg8QllsnHlRS2HoXypXymAhepxJT8W3DI0Ue4&quot;</pre><p>So my WIF principal (which expected email addresses) couldn’t match it.</p><h3>FIX: Map google.subject to the Right Claim</h3><p>Instead of using the raw sub, I updated the attribute mapping in the WIF provider to use the Entra claim preferred_username:</p><pre>google.subject = assertion.preferred_username</pre><p>That way, the subject became the user’s email address:</p><pre>&quot;preferred_username&quot;: &quot;aashna.kunkolienker@narachi.fun&quot;</pre><h3>Why This Matters for AgentSpace</h3><p>AgentSpace uses the google.subject value as the <strong>profile user identity</strong> when checking ACLs.</p><p>If your ACL JSON contains user emails like this:</p><pre>&quot;acl_info&quot;: {  <br> &quot;readers&quot;: [    <br>{ &quot;principals&quot;:<br> [{ &quot;user_id&quot;: &quot;aashna.kunkolienker@narachi.fun&quot; }] }  <br> ] }</pre><p>Then your WIF provider <strong>must map </strong><strong>google.subject to an email claim</strong> (e.g. assertion.preferred_username or assertion.email, depending on what your ID token emits).</p><p>Otherwise, ACL checks will silently fail, because the user ID in your ACLs won’t match the opaque sub.</p><p><a href="https://medium.com/@aashna-kunk/debugging-third-party-identity-provider-entra-id-and-agentspace-integration-why-sub-breaks-141b4792f0bf">https://medium.com/@aashna-kunk/debugging-third-party-identity-provider-entra-id-and-agentspace-integration-why-sub-breaks-141b4792f0bf</a></p><h3>6. Conclusion</h3><p>Whenever you set up Entra ID as an identity provider for Google Cloud WIF, <strong>always decode your ID token first</strong>. This simple debugging step can save hours of troubleshooting!!</p><p>By confirming the subject, email, and group claims up front, you’ll know exactly how to configure your acl_info for custom data sources and attribute mappings in Google Cloud.</p><p>❤</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=beddefcc186f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Turning a MIDI Keyboard into a Game Controller with Python]]></title>
            <link>https://aashna-kunk.medium.com/turning-a-midi-keyboard-into-a-game-controller-with-python-1fe5bb809738?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1fe5bb809738</guid>
            <category><![CDATA[midi]]></category>
            <category><![CDATA[music]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[creative-coding]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Thu, 03 Jul 2025 20:18:44 GMT</pubDate>
            <atom:updated>2025-07-03T20:18:44.233Z</atom:updated>
            <content:encoded><![CDATA[<p>(Disclaimer: I like building things that are silly and surprisingly functional)</p><p>This one stemmed from a simple question:</p><p><strong>Can you play a video game using a MIDI keyboard?</strong></p><h3>The concept:</h3><p>MIDI keyboards send note data when you press keys. Python can listen to those notes using mido and simulate real keyboard presses using pynput.</p><p>press a C note: it triggers the left arrow.<br>play a G chord : it jumps.</p><p>I tested it out on Chrome Dino and Pac-Man.<br>It was surprisingly funnn.</p><h3>The setup</h3><p>All you need is:</p><ul><li>A MIDI keyboard</li><li>Python 3.x</li><li>Install theselibraries:</li></ul><pre>pip install mido python-rtmidi pynput</pre><h3>Main logic :</h3><pre>note_to_key = {<br>    60: Key.down,   # C note maps to the down arrow <br>    62: Key.right,  # D note maps to the right arrow... etc <br>}</pre><pre>jump_chord = frozenset([60, 64, 67])  # C major chord for chord mappings</pre><p>The script listens to incoming notes. If one matches note_to_key, it presses that key. If the current notes form the jump_chord, it holds the spacebar. Once released, the script lifts the key.</p><p>It’s accurate enough for simple games and adaptable for more serious and complex use. Try using it for COD or minecraft wow that’ll be a huge win.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/664/1*DscMfrbzCvioyrlIsacRrQ.png" /><figcaption>Chrome dino game, controlled using the midi keyboard</figcaption></figure><p>The full code is on GitHub: <a href="https://github.com/aashnakunk/midi-to-keyboard">github.com/aashnakunk/</a><a href="https://github.com/aashnakunk/midi-to-keypress">midi-to-keypress</a></p><p>Try it, make a mess.<a href="https://github.com/aashnakunk">github.com/aashnakunk</a> | <a href="https://aashnakunk.com/">aashnakunk.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1fe5bb809738" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bach-man: Pac-Man, but for Music Theory Nerds]]></title>
            <link>https://aashna-kunk.medium.com/bach-man-pac-man-but-for-music-theory-nerds-38565e699d83?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/38565e699d83</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[creative-coding]]></category>
            <category><![CDATA[game-development]]></category>
            <category><![CDATA[lua]]></category>
            <category><![CDATA[music]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Fri, 13 Jun 2025 06:22:23 GMT</pubDate>
            <atom:updated>2025-06-13T06:22:23.207Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Every now and then your free will gets overridden, not by fate, but by a completely unnecessary, wildly specific idea you feel an uncontrollable need to code. One minute you’re thinking, the next your editor is open and you’re three files deep into something no one ever asked for :P</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AwI82_CR21x1Vv4OEF8MOA.png" /><figcaption>Bach-man’s welcome screen!</figcaption></figure><p><strong>What if Pac-Man was made with musical notes instead of pellets?</strong><br> And what if… those notes mattered ? like if you had to gobble only the correct notes in a given musical scale, and eating a note in the “wrong key” would kill you?</p><h3>Game overview:</h3><p>Bach-man is a side-scrolling game where the adorable yellow “pac-man” navigates a musical staff, eating notes in the specified scale. Choose correctly and you score points and go to the next level. Eat a wrong note (like a F# in the key of C) and a red ghost eats you up and ends your life.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ckEFRJQ8XqqjUFSt6Yhtuw.png" /><figcaption>Bach-man’s main game screen</figcaption></figure><h3>How I Coded It:</h3><p>The entire game is written in <strong>Lua</strong>, using the very convenient <a href="https://love2d.org/">LÖVE2D</a> engine which is refreshingly retro and surprisingly forgiving if you, like me, enjoy basic game dev without Unity-level complexity.</p><h3>Musical Logic Integration</h3><p>There’s a dictionary of scales:</p><blockquote>[“C Major”] = {<br> valid = {60, 62, 64, 65, 67, 69, 71, 72},<br> invalid = {61, 63, 66, 68, 70}<br>}</blockquote><p>…and each note eaten is checked against that when it’s ‘eaten’.</p><p>The notes are drawn on a staff using pitch values mapped to y-coordinates. Getting each note to line up correctly on the staff was a bit painful haha.</p><blockquote>local notePositions = {<br> [60] = 360, — C4<br> [62] = 350, — D4<br> …<br>}</blockquote><p>I kept the game pretty simple, but a few things definitely ate into my patience more than I expected:</p><ol><li>Mapping perfect pixel-coordinates to musical pitches was way more annoying than it sounds.</li><li>Making sure it didn’t falsely “game over” on legit note choices because that would make me question all my music theory fundamentals XD</li><li>Making it mobile-compatible was a little annoying. LÖVE2D supports it, but scaling and touch input are finicky.</li><li>I wanted to add support for different modes like Aeolian, Lydian etc, and user accounts and a global scoreboard just to make it competitive. Then I remembered, this is literally a Pac-Man joke I coded only “because I can”</li></ol><h3>You Can Play It!!</h3><p>The current version is deployed on Firebase here:<br><a href="https://bach-man.web.app/"> </a><a href="https://bach-man.web.app/">https://bach-man.web.app/</a></p><p>(It’s still a little buggy and definitely plays best on laptops or iPads thanks to some unresolved scaling chaos. But if this sounds like your kind of weird, give it a go! Also wow, if you’re also the kind of person who’d be into Pac-Man munching musical notes… good to know I’m not alone XD)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YuUz5Yi8KNx9N6tFK0ImpQ.png" /></figure><p>If you have suggestions, or are just curious about how I coded it in Lua (since it’s a pretty niche language), feel free to reach out at <a href="/service/mailto:aashnakunk@gmail.com">aashnakunk@gmail.com</a>. I’m always happy to talk about weird side projects, retro game dev, or even just music in general.</p><p>❤ Aashna</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=38565e699d83" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sniffing an AirDrop connection on Wireshark (iPhone to MacBook)]]></title>
            <link>https://aashna-kunk.medium.com/sniffing-an-airdrop-connection-on-wireshark-iphone-to-macbook-4f682d635d34?source=rss-37cb0bd5899b------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f682d635d34</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[airdrop]]></category>
            <category><![CDATA[wireshark]]></category>
            <category><![CDATA[networking]]></category>
            <category><![CDATA[apple]]></category>
            <dc:creator><![CDATA[Aashna Kunkolienker]]></dc:creator>
            <pubDate>Wed, 26 Feb 2025 17:56:34 GMT</pubDate>
            <atom:updated>2025-02-26T20:23:55.063Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/536/1*ykPghEmpCLHkbB3jt_l_6g.png" /></figure><p>I was casually observing some packets on a wifi network at Cold Spring Coffeehouse, when I noticed an adaptor labeled “awdl0” on wireshark and I figured this might be related to AirDrop connections. Just for observation&#39;s sake, I decided to dive into the detailed process of how the connection takes place on a computer-networking level.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/321/1*8m-8x-3grRt6AqYjP3nTKg.png" /><figcaption>AirDrop from my iPhone to “Gabuchin’s MacBook Air”</figcaption></figure><h3><strong>FINDING EACH OTHER ON THE NETWORK</strong></h3><p><strong><em>iPhone :</em></strong><em> “</em>To everybody in my range: are you ‘Gabuchins MacBook-Air’ or do you know who it is?<em>”</em></p><p><em>[check end of document for more details on the mDNS query]</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*um0dL89WH1Poq7zcpxABWg.png" /><figcaption>MDNS packets</figcaption></figure><p>It starts with an MDNS query (<strong>multicast DNS</strong>) . MDNS is used instead of regular DNS because it has to multicast this on the local network (AirDrop) due to the absence of a central server. This message is sent to ff02::0b, a special multicast address. Btw, the standard query response in mDNS shows airdrop_tcp.local (<em>check the line highlighted in dark blue</em>), means it’s a response to an AirDrop request.</p><p>Gabuchin’s Macbook Air reponds saying:</p><p><strong>MacBook</strong>: “heyy that’s me! You can find me at the MAC address fe80::a854:65ff:fedd….”</p><p>In the TCP handshake:</p><ul><li><strong>iPhone’s MAC addresss:</strong> fe80::709b:d9ff:fe03</li><li><strong>MacBook Air’s MAC Address:</strong> fe80::a854:65ff:fedd</li></ul><h3>THE TLS HANDSHAKE</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3nNuRW_AWqfyHkrl2YEBAA.png" /><figcaption>TLS packets</figcaption></figure><p>Once the discovery phase is complete and the devices know each other, they need to establish a secure channel. This is where <strong>TLS (Transport Layer Security)</strong> comes into play.</p><p>The iPhone initiates the process by sending a <strong>Client Hello</strong> message. This is its way of saying: <em>“</em>Ello, I want to talk securely, I support these ciphers, and here’s my random number for the key”</p><p>The MacBook Air responds with a <strong>Server Hello</strong>, selecting one of the ciphers from the list and sending its own random number. It’s essentially saying:<br>“Ok, let’s use this encryption method. You can have my random number too.”</p><p>On the second last line, you’ll see CHANGE CIPHER SPEC, which means “ From now on, let’s use encryption” and this decision is acknowledged by both my iPhone and MacBook.</p><h3>HANDLING CONNECTION DISRUPTION</h3><p>Line 26 and 27 show MDNS (multicast DNS) packets sent again, i.e., the iPhone doesn’t know and is asked “Hey, who on this network is <em>Gabuchin’s Macbook Air</em>”?</p><p>Why did this happen?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BpgUWCq_pdzUtMN_a3_k4Q.png" /><figcaption>Connection disruption: TCP retransmissions, DUP ACKs and RST</figcaption></figure><p>There was likely a disruption in the connection. We can see a lot of <strong>TCP retransmissions</strong> (highlighted in black) which indicate the iPhone is trying to resend packets that aren’t acknowledged by my MacBook.</p><p>Check line 23: the iPhone does start sending application data but line 26 shows an mDNS and the <strong>entire discovery process happening again</strong>. On line 34, we see a <strong>RST</strong> (RESET) packet meaning the <strong>connection is forcefully shut down.</strong></p><p>After this, we’re all good to go with the TLS handshake happening again, and now the real file transfer can begin!</p><h3>FILE TRANSFER</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_xH1ZvVBsUcRNAbJqv6-ag.png" /><figcaption>File Transfer</figcaption></figure><p>There are 1000+ capture lines only covering the file transfer part for just one image! This might seem quite excessive, but observe the “Info” section carefully: only a few lines have “<strong>Application data</strong>”; the others are mostly SYNs and ACKs and window updates(<strong>TCP’s reliability mechanisms</strong>). Also, if you want to know what all those sequence and acknowledgement numbers mean and how data is sent over a TCP connection, do watch this: <a href="https://www.youtube.com/watch?v=xMtP5ZB3wSk">https://www.youtube.com/watch?v=xMtP5ZB3wSk</a></p><h3>TRANSFER COMPLETE</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OcM3JY4HEmJVnw3Vuot7jA.png" /><figcaption>Connection Termination, file transfer complete</figcaption></figure><p>When the file transfer is complete, we see the <strong>[ FIN, ACK ]</strong> packet at line 1220 which says “I’m done sending packets. Let’s close this conversation properly<em>.</em>&quot;</p><p>However, <strong>we also see RST (Reset) packets</strong> in <strong>lines 1219 and 1221</strong>, which suddenly kill the connection. A <strong>RST</strong> packet is like slamming the door on your friend’s face instead of hugging them goodbye. AirDrop does this for efficiency, maybe.</p><h3><strong>DNS DETAILS</strong></h3><p>For the networking nerds, here’s more detail about the DNS process:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NGdJWmp1e_fa4Dc18VzzRw.png" /><figcaption>MDNS packets</figcaption></figure><p>Let’s open up packet 1, here’s what it looks like:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4G4v3wJu6tURlB_FFQ_0Pg.png" /><figcaption>MDNS packet 1 opened</figcaption></figure><p>The query asks: <em>“Who has an AirDrop service (airdrop_tcp.local) available? </em>This is sent to <strong>ff02::fb</strong> (the mDNS multicast address)</p><p>The DNS Answer section shows: <em>402e0efe49f1._airdrop._tcp.local → type SRV, port 8770, target 2ea34b25–0112–435b-977a-b4f5c8362201.local</em></p><p>Which is the MacBook telling the iPhone that it’s available for AirDrop. Here’ the hostname is linked to the MAC address in the line: <em>2ea34b25–0112–435b-977a-b4f5c8362201.local → fe80::a854:65ff:fedd:d0e5</em></p><p>Also, a fun fact I stumbled upon: The name you see in AirDrop UI (like “Gabuchin’s MacBook Air”) is mapped from the hostname stored in Bonjour (Apple&#39;s Zero-Config Network Service). That is not directly sent over mDNS but is retrieved once the device is found on the network.</p><h3><strong>MORE RANDOM FACTS</strong></h3><p><strong>AirDrop uses ephemeral hostnames</strong><br>Instead of using fixed name like gabuchins-macbook-air.local, the device often responds with a hostname based on UUID like <em>2ea34b25–0112–435b-977a-b4f5c8362201.local. </em>[ prevents persistent tracking ]</p><p><strong>Port 8770 is always used for AirDrop file transfers</strong><br>If you see traffic on port 8770 while monitoring a connection, it’s coming from an AirDrop session.</p><p>Hope you liked reading this article!</p><p>Feel free to send over any fun PCAP observations from your WiFi/AirDrop or any other connections to <strong><em>aashnakunk@gmail.com</em></strong>. I’d love to discuss your findings :))</p><p>Love, <br>Aashna</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f682d635d34" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>