<?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 Prabha Matta on Medium]]></title>
        <description><![CDATA[Stories by Prabha Matta on Medium]]></description>
        <link>https://medium.com/@prabhamatta?source=rss-f607d030fdd9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*joko9bi4HYvlwXGpj_um7w.png</url>
            <title>Stories by Prabha Matta on Medium</title>
            <link>https://medium.com/@prabhamatta?source=rss-f607d030fdd9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 05:20:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@prabhamatta/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[Agentic AI & Agentic Workflow Orchestration - What is it?]]></title>
            <link>https://medium.com/@prabhamatta/ai-agent-vs-agentic-ai-agentic-workflow-orchestration-f3cfa95f4f70?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3cfa95f4f70</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[tech-innovation]]></category>
            <category><![CDATA[ai-automation]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 23:08:24 GMT</pubDate>
            <atom:updated>2025-04-17T16:07:33.595Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Agentic AI &amp; Agentic Workflow Orchestration - What is it?</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SHyDCD3aZM9RStVzZDhZvA.png" /></figure><p>In the rapidly evolving world of artificial intelligence, the distinctions between different approaches can often feel blurry. However, understanding these nuances is key to leveraging AI effectively in our workflows. One fascinating comparison is between <strong>AI Agents</strong> and the emerging concept of <strong>Agentic AI</strong>.</p><h3>What is an AI Agent?</h3><p>An <strong>AI Agent</strong> is designed to handle specific tasks. Think of it as a single-purpose tool — efficient, focused, and task-oriented. These agents:</p><ul><li>Execute predefined tasks autonomously.</li><li>They rely on strict programming or rules to operate.</li><li>And excel in narrow domains, such as scheduling a meeting or processing a form.</li></ul><p>While incredibly useful, AI Agents operate within a limited scope. They don’t dynamically adapt to broader contexts or multiple interdependent tasks.</p><h3><strong>AI Agent vs Agentic AI</strong></h3><p>Here is a high-level overview of the AI workflows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ee5SGLpE51an1_ja9SZOZQ.png" /><figcaption>Adapted from AI Product Management course</figcaption></figure><h4><strong>Assistant workflow (single-step AI Agent)</strong></h4><p>In some cases, the deployment is through a simple assistant workflow. An assistant workflow consists of the model, and additional tools that it might have access to ie., data that it can reference, and then the ability to maintain a form of memory in a thread. But it is typically going to be very single-step.</p><p>For instance, the <a href="https://medium.com/@prabhamatta/ai-powered-multi-modal-document-assistant-fcb0f5a41ebc">document analyzer application</a> I developed, is a single-step AI agent where the input data (could be in multiple formats — pdf, text, website) is analyzed. A question is taken from the user to provide an output after analyzing the data.</p><h4><strong>Agentic workflows (multi-step AI agent)</strong></h4><p>Agentic workflows are ones that can do multiple steps. There is a notion of self-correcting and planning ie., go back, reflect, come back, make a change based on new information, and potentially go back and forth on it.</p><p>The same concepts of data, access to data, access to tools, and access to memory are involved here too, however, all those kinds of pieces work together.</p><p>The reasoning model is not necessarily an agentic workflow, but could be argued that way too. Because the workflow involves going back and forward, it can be considered as thinking and reflecting. For instance, solving a maze puzzle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wzee8D1C9M8c-ATxCZJdUA.png" /></figure><h3><strong>Orchestration</strong></h3><p>When you move into this world of orchestration, you have multi-step things going on, and you want to be able to have the inputs from one become the outputs of the next, or the outputs of one become the inputs of the next. And also being able to orchestrate a few of these things together.</p><p>Now, the question is how do we coordinate these multiple systems, agents, and assistants, along the way for a more complex task.</p><h3>Orchestration Techniques</h3><p>So, how do these Agentic Workflows work?</p><h4><strong>1. Hand Offs</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s862z5DD3DfExYcrdk5pvg.png" /><figcaption>Adapted from AI PM course</figcaption></figure><p>You can think of these as organizational workflows and organizational structures. For some of you who have spent a lot of time in team building and thinking about what the right way to do things is and what the right method of delegation is. Especially those of you who have stood as managers. This is going to look very familiar.</p><p>For instance, the customer comes in, and says, where’s my order? You have a triage agent or assistant who can go through and say, “Okay, well, here are the choices I have to be able to solve this problem. Which one do I go through? It could be issues and repairs. It could be a sales-related thing. It could be an order thing.”</p><p>Since the person has clearly said “order”, it is interpreted as “order thing” against one of the things that I have. And then send it to the “orders agent” to be able to resolve against.</p><p>And then similar to letting the person talk directly back to the user. Here, it is not that the individual is talking, the <strong>triage agent is talking back to the user</strong>. The triage agent l<strong>iterally moves the conversation over to the orders node. And lets the orders node talk back to the user.</strong> It is similar to a call center. Where this is “press one for this, press two for that, press three for this”. And then when you press it, you go ahead and hand it to another person.</p><p>And then when that person can’t solve your problem, they send you back to the first person that potentially sends you to the next person. Or they know directly where to go. So a lot of this is going to feel very “call center” style because that’s kind of just how we as humans have organized ourselves. And not surprisingly, that’s how models can be organized as well.</p><p><strong>2. Nested Calls</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5uGO7lOYrvmSE0Z3j5Oi-A.png" /><figcaption>Adapted from AI PM course</figcaption></figure><p>Nested Call is the next one. In this, it is something along the lines of “I have a triage agent and the triage agent deals with all the communication. Because I have trained the triage agent to communicate in a certain way. And then the internal teams don’t necessarily communicate that way. So <strong>I don’t necessarily have to have multiple front-of-house conversations. I just have the one.”</strong></p><p>And this is probably very akin to what we do on technical teams where there is an issue. <strong>We triage it internally.</strong> We have a conversation about what’s the right way to do things. Then when the problem is solved, there is <strong>a front-facing person</strong>, a product manager, a Customer-success manager, a salesperson, or a support person, whoever they might be, to relay the information back to the user.</p><p>Again, thinking about how you do this, you can trigger off a bunch of subtasks.</p><ul><li>First case, you can potentially make this smarter like being able to know specifically which one or two multiple tasks to be able to go off to and then decide which is the right information to communicate back to the user. You have to figure out who can solve the problem. In this case, point to the specific order agent and ask “Where is my order”?</li><li>Alternatively, you can also do the brute force thing of always sending it to everybody and always getting the information back. You can ask all the sub-agents, “All of you, where is my order?”. Then the order agent says “Oh, that’s me. I’m the one that’s supposed to solve it. Here’s the order.”</li></ul><p>So it depends on how you want to handle it depending on the complexity of your task.</p><p>Then the front-facing main agent says, “Oh, okay. The order is on its way. I’ll let the customer know back exactly what’s going on.”</p><p><strong><em>Pro-tip: This is very valuable when you have multiple different types of subtasks.</em></strong></p><h4>3. Manager Role</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xhJNLihxDMZ8dfRvqaGSvA.png" /><figcaption>Adapted from AI PM course</figcaption></figure><p>The <strong>manager role is an aggregator role. </strong>And what you are doing in this situation is, the manager agent is the one that puts together the plan —</p><ul><li>The plan of how we are going to be able to solve this problem.</li><li>And then runs each individual task.</li><li>The manager will conduct the entire plan, one after the other after the other.</li><li>Then, use the information from the previous steps to be able to formulate what the next step should be or what the final output should be.</li></ul><p>Note: This can take time. So, a few important things have to be taken into consideration: How do you handle when the system takes more time eg: a few minutes or a few hours? What should the interface look like? What would the user do in the meantime? A couple of options here —</p><ul><li>Do I tell them to come back later?</li><li>Do I give them a stream of information as it goes? For instance, Claude AI or ChatGPT shows the stream of information along the way when providing the output. In some applications, the draft does not necessarily reflect the final. Some questions to consider here: do people want to see the draft? Or do people want to see the final?</li></ul><h4>4. Advanced Hand-off ie., Triage Agent initially and then followed by Assistant Node</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zxfDWT2IIyRaNSYvkYZ0mA.png" /><figcaption>Adapted from AI PM course</figcaption></figure><p>Advanced Handoffs are the ones majorly in production today. This is how some organizations have put together <strong>customer support tools</strong>, which have become a very prominent thing that a <strong>lot of enterprises are doing today</strong>.</p><ul><li>First, there is a triage agent. The agent figures out a routing task of which sub-assistant is going to be able to do something.</li><li>And then that assistant takes on the task through its function calling to be able to work through.</li><li>But the entire task is handed off to that assistant to then interact, deal with, and go through multiple steps.</li></ul><p>E.g.: Swarm — This is an orchestration framework that OpenAI has released called <strong>Swarm</strong>.</p><p><strong><em>Swarm basically does the thing of taking these assistant nodes and then piecing them together in a handoff traditional way, or in other ways that we’ve talked about.</em></strong></p><p>How does this mechanism work?</p><p>It is like the Triage agent saying “Okay, well, I don’t have a solution to your problem, but this person does. They will work with you until that problem is solved, and they will work back and forth”.</p><p>And that’s very <strong>akin to customer support scenarios t</strong>oday in real life with humans.</p><p>— — — — -</p><p><em>Note: a Medium article on </em><strong><em>how to use OpenAI’s Swarm </em></strong><em>is WIP currently. I will be posting it soon. Stay tuned!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3cfa95f4f70" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Create an AI-powered Social Media Publishing Agent using Zapier]]></title>
            <link>https://medium.com/@prabhamatta/how-to-create-an-ai-powered-social-media-publishing-agent-using-zapier-f0b196085c75?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0b196085c75</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[social-media-publishing]]></category>
            <category><![CDATA[ai-social-media-manager]]></category>
            <category><![CDATA[zapier]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 04:41:07 GMT</pubDate>
            <atom:updated>2025-01-07T04:41:07.335Z</atom:updated>
            <content:encoded><![CDATA[<p>In this Notion webpage (linked below), we’ll explore how to create an AI-powered Social Media Publishing Agent using ChatGPT and Zapier.</p><p>This agent creates draft posts and requests approval. Once approved, it publishes the content on LinkedIn and further, records the post link and creation date in a Google Sheet.</p><p>Sample Linkedin Post [<a href="https://www.linkedin.com/posts/aipublisher_the-power-of-grit-a-key-ingredient-activity-7282226178113908736-E3yT">link here</a>]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zbrAMvtKulJOKPa0ZsAb2g.png" /></figure><p>Let us dive into the details….</p><p><a href="https://dot-stomach-ade.notion.site/Create-AI-Social-Media-Publishing-Agent-using-Zapier-174e5cf37edb80538b6ec89c999d4a6b">Create AI Social Media Publishing Agent using Zapier | Notion</a></p><p><em>Originally published at </em><a href="https://dot-stomach-ade.notion.site/Create-AI-Social-Media-Publishing-Agent-using-Zapier-174e5cf37edb80538b6ec89c999d4a6b"><em>https://dot-stomach-ade.notion.site</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0b196085c75" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An AI powered Email Communications Coach App]]></title>
            <link>https://medium.com/@prabhamatta/clear-comms-app-an-ai-email-communications-coach-04344f4f6908?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/04344f4f6908</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gen-ai-evaluation]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Thu, 05 Dec 2024 09:03:33 GMT</pubDate>
            <atom:updated>2024-12-14T13:16:52.176Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/334/1*zi6rB53Lc1RU5n5pdh1Jrw.png" /></figure><p><strong>Product</strong>: ClearComms is an AI-powered email communications coach application for product managers, sales professionals and customer success managers, helping them to craft clear, impactful professional emails for the leadership team, prospects and customers.</p><h3>Problem Details</h3><ul><li><strong>Industry</strong>: Professional email communication tool for product management, Sales and Customer Success teams</li><li><strong>Persona</strong>: Product managers who regularly interact with executives and leadership teams. CSMs who regularly interact with customers. Sales team who regularly interact with prospects.</li></ul><h3>User Journey Stages &amp; Activities</h3><ul><li><strong>Email Drafting:</strong> Product managers write an initial email for executives.</li><li><strong>Submission for Feedback:</strong> Users submit their draft to ClearComms for analysis.</li><li><strong>Analysis &amp; Suggestions:</strong> ClearComms evaluates the email and provides structured feedback while giving recommendations for changes</li><li><strong>Revision:</strong> Users revise their email based on detailed suggestions. Alternatively, use the rewritten email provided by ClearComms</li><li><strong>Finalization:</strong> The refined email is sent to the executive or leadership team.</li></ul><h3>Key Pain-points</h3><ul><li>Product Managers struggle to articulate key points succinctly and with clarity due to lack of time.</li><li>Lack of confidence in the tone or strategic alignment of their communication.</li><li>Difficulty balancing sufficient detail with brevity for executive audiences.</li><li>Wasting time iterating on drafts without clear guidance.</li><li>Misaligned communication causes more delays in decision-making or misinterpretation of priorities.</li></ul><h3>Solutions Today</h3><p><strong>How do users address their problems today:</strong> Manual review by peers, mentors, or leadership, which is time-consuming and inconsistent.</p><h3>Improvement Opportunities with LLMs</h3><ul><li>Convert verbose or unclear drafts into concise, clear emails optimized for executive communication.</li><li>Ensure a consistent professional tone and structure across emails.</li><li>Provide actionable feedback on specific areas (e.g., tone, structure, clarity) to guide improvement.</li><li>Enable structured feedback and provide suggestions for rewrites.</li></ul><h3>What is the opportunity?</h3><h4>Users &amp; Customers</h4><ul><li>PMs , CSMs and Sales team.</li></ul><h4>Problems</h4><p>PMs:</p><ul><li>difficulty in communicating efficiently with the leadership team causing delays in decision in decision making.</li><li>Lack of clarity, causing misalignment in priorities.</li><li>Missed opportunities to showcase strategic thinking and rigor in email content.</li></ul><p>CSMs — difficulty in communicating with clarity and maintaining good tone with their upset customers.</p><p>Sales team — difficulty in communicating with confidence and convincingly with their prospects.</p><h4>Assumptions</h4><ul><li>PMs, CSMs and Sales teams will adopt AI-generated feedback if it improves their drafts without increasing their workload.</li><li>The suggested rewrites align with executive communication standards.</li><li>Executives will notice improved communication, validating ClearComms’ utility.</li></ul><h4>Business Problems</h4><ul><li>Reduces time spent iterating on emails, improving productivity.</li><li>Enhances alignment and efficiency in decision-making processes for companies.</li></ul><h3>Solutions</h3><h4>Solution Ideas</h4><ul><li><strong>Solution 1:</strong> Feedback system with specific actionable suggestions (e.g., improving tone, rewriting unclear sections).Tailored examples to show how revisions align with best practices.</li><li><strong>Solution 2:</strong> Provide a “suggested re-write” of the email directly to the user</li><li><strong>Solution 3:</strong> Provide a scoring system for drafts, highlighting areas of improvement. And give some examples of improvements from the email</li><li><strong>Solutions 4</strong> Take in the email draft/content from the user in the form of a voice message, process the message and then provide a “suggested re-write” of the email</li></ul><h4>Solution Prioritization</h4><p><em>Use RICE Framework to prioritize — ClearComms is the prioritized solution</em></p><h3>Success Outcomes/User Value</h3><ul><li>Product managers can confidently send emails that are concise, clear, and aligned with executive standards.</li><li>CSMs can efficiently handle upset customers in highly intense situations</li><li>Sales teams can confidently write replies to their prospects</li></ul><h3>Success Metrics</h3><h4>User Metrics</h4><ul><li>Frequency of email drafts analyzed.</li><li>User satisfaction with feedback provided.</li><li>Time saved per draft revision.</li></ul><h4>Business Metrics</h4><ul><li>Increase in customer retention due to improved communication tools.</li><li>Enhanced productivity and faster decision-making cycles for companies — assume the average time taken by the individual is 15 minutes. Average time taken by the user before changing the chat topic(including abandoning the chat) for a single topic</li></ul><h3>Adoption Strategy</h3><ul><li><strong>Single-segment launch:</strong> Initial launch targeting product management teams.</li><li><strong>Freemium</strong>: Free trials or pilots to demonstrate impact.</li><li>Demos and tutorials/sample cookbooks</li><li>Word of Mouth, Success stories blogs: Testimonials and success stories from leadership teams.</li></ul><h3>_____How to do your Gen AI App Evaluation____</h3><p>See this<strong> </strong><a href="https://medium.com/@prabhamatta/how-to-do-your-gen-ai-application-evaluation-1726152b3255"><strong>medium blogpost</strong></a><strong>.</strong></p><h3>_______How to build ClearComms App________</h3><p><em>Coming soon…</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04344f4f6908" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gen AI Application Evaluation with OpenAI Platform: A Practical Tutorial]]></title>
            <link>https://medium.com/@prabhamatta/how-to-do-your-gen-ai-application-evaluation-1726152b3255?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/1726152b3255</guid>
            <category><![CDATA[gen-ai-evaluation]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Thu, 05 Dec 2024 08:53:22 GMT</pubDate>
            <atom:updated>2024-12-14T13:09:51.548Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8uIHe7ZDZX8BoS8vndPzmg.png" /></figure><p>Before deploying your application to production, it is critical to run evaluations to make sure your application prompt is accurate and reliable. In this post, let us learn how to run automated evaluations on our Gen AI application “ClearComms” using the Open AI API platform. (what is ClearComms? — see <a href="https://medium.com/@prabhamatta/clear-comms-app-an-ai-email-communications-coach-04344f4f6908">here</a>).</p><p>First and foremost, create an account on the OpenAI platform and get an API key.</p><h3>Test Data</h3><p>Here is the <a href="https://docs.google.com/spreadsheets/d/1-fXhmpWeFRJYO1rn97AoaCAP1ACgzVJtMBUyGG00Xgg/edit?usp=sharing">ClearComms Sample Data</a> with input/output pairs in a Google sheet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WNnF8f2ySpYQEEdp" /></figure><h3>Evaluation Steps</h3><h4><strong>Step 1: Provide the system prompt in the “Playground” tab and dry run with some test data</strong></h4><ul><li>Option1: Use Gen AI to create your prompt</li><li>Option2: If you already have a prompt, just use it directly</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CLuhOfGzYXQ3OIp3so1CBQ.png" /></figure><h4><strong>Step 2: Create a new evaluation and import the test data into the platform</strong></h4><p>Under the Dashboard tab, click evaluations in the left side panel to create a new evaluation</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3pV1WO66ni_Kq4KBaPnOFA.png" /></figure><p><strong>Import the test data into the platform</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DOO4mIQF838rfwaL-YW2Tw.png" /></figure><h4><strong>Step 3: Create some test criteria</strong></h4><p>I have created string checks, semantic checks and criteria match checks. Here are the configurations of some of these checks I used for those evaluation criteria —</p><p><strong>Semantic Similarity Check</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hKYJu4z9pZJs98u6WPah9A.png" /></figure><p><strong>Salutation Check</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w1FDRNfY4Cm7TLRYLIPCOg.png" /></figure><p><strong>Signature Check</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o9oOIjk86xJ7w53Vn27HmA.png" /></figure><p><strong>String Check</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AXG6d7hDjEgg729DUAWGWA.png" /></figure><p><strong>Step 4: Run the tests</strong></p><p>Click the Run button below to run the test criteria against the Google sheet test data</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tkKTB7HzOra5TbfQ" /></figure><h4>Evaluation Round 1</h4><p>Evaluation Summary</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cHure9yq1FungidO" /></figure><p><strong>Evaluation detailed results</strong>: Some tests failed!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8P94swIrNc08iRPu" /></figure><h4>Evaluation Round 2</h4><p>I refined the prompt a couple of times to handle the checks and re-ran the tests.</p><p>Evaluation Summary:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dSHFL0dYj7dRveb4" /></figure><p><strong>Evaluation Detailed Results</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GWDqPXd1YSsieIbb" /></figure><p>All tests passed!</p><h3>_________________________________</h3><h3>How to build ClearComms App</h3><p><em>Coming soon…</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1726152b3255" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Every PM will be an AI PM in the future]]></title>
            <link>https://medium.com/@prabhamatta/cracking-the-code-how-product-discovery-for-ai-products-redefines-the-playbook-and-why-its-not-dcd45a1a1cd3?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/dcd45a1a1cd3</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-product-discovery]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[ai-product-management]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Sat, 23 Nov 2024 03:51:42 GMT</pubDate>
            <atom:updated>2024-12-14T20:28:15.800Z</atom:updated>
            <content:encoded><![CDATA[<h3>Every PM will be an AI PM in the future: Product Discovery for AI Products Redefines the PM’s Playbook</h3><p>Everyone talks about the death of Product Management because of AI. Maybe, it seems so in the short term.</p><p>Indeed, we can perceive it the other way around. Product Management is now more essential than ever…we just need to see the Software Development Life Cycle (SDLC) from a different lens! Here is a write-up on various types of AI PMs; and PM’s methodologies in product discovery for AI products, compared to the traditional software discovery approach.</p><p>In the future every Product Manager will be an AI PM and if you are seeking role clarity, here is how to think about AI PM roles:</p><h3>Role 1: Building AI Powered Products</h3><p>These PMs don’t build AI technology but leverage the available AI tech in existing customer journeys to solve customer problems in a better way for their products. Eg: Gen AI powered customer chatbots which give accurate answers on customer queries or Gen AI powered product review summaries seen on Amazon/Flipkart.</p><p>These PMs also work on products which act as a wrapper on top of existing AI tech accessed via APIs. Eg, my pet projects Document Summarizer and Clear Comms for effective email communication. All the image generation, note taking, voice to text transcription products fall in this category.</p><h3>Role 2: Building AI Platform capabilities</h3><p>These PMs build products/features which will be used by AI teams in order to build a AI product. Eg: a PM in MLOps team working on improving data labelling or a PM building observability capabilities for AI workflows.</p><p>Here the product discovery for AI products vary very differently. More details below.</p><h3>Role 3: Building Experience on top of core AI technology</h3><p>These PMs build products where the company’s core technology is centred around AI, such as chatbots or language models, and their job is to commercialise the technology and make it consumable for businesses or consumers. <br>Eg PMs who are working on features for Perplexity, ChatGPT, or Claude.</p><p>Understanding this distinction across roles is key as it will help you understand what adjacent skills you have today and which skills you need to build, which will help you get into a PM role and successfully thrive.</p><p>Now, coming the intial stages of product development, troduct discovery has always been a critical step in building great software. You gather requirements, define features, test with users, and then launch. Easy, right? Well, not anymore — Building AI-powered products is a whole new ball game, and it’s not just about writing code or designing features. Let’s talk about why AI product discovery is so different and how teams are navigating this brave new world.</p><h4>Why AI Product Discovery Feels So Different</h4><p>If you’re stepping into the world of AI products, be ready for a discovery process that’s deeper, more iterative, and centered around data. Building AI products isn’t just a tech challenge — it’s a whole new paradigm for product discovery.</p><p>The rules that worked for traditional software development don’t always cut it here. From the way we gather data to how we test and iterate, success here requires blending technical rigor with user empathy and understanding both the capabilities of AI and the needs of your audience.</p><p>Traditional software development is pretty straightforward — you figure out what users want, build it, and refine it based on feedback. AI, on the other hand, plays by a whole new set of rules:</p><ul><li><strong>Data is King</strong>: Forget static requirements. With AI, your product is only as good as the data feeding it. You’re not just designing features; you’re training models and constantly tweaking them as new data rolls in.</li><li><strong>The Learning Never Stops</strong>: While traditional software might get a version update every few months, AI products need to evolve in real-time. Think of them as dynamic, self-improving systems.</li></ul><h4>How AI Product Discovery Compares to Traditional Software</h4><p>Here’s a quick cheat sheet on how discovery for AI products stacks up against the old-school approach:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y20KCJq0ttVheqpGFGCxnQ.png" /></figure><h4>Challenges (a.k.a. the Stuff That Keeps AI Product Teams Up at Night)</h4><p>AI might sound like magic, but it comes with its fair share of headaches. Here are some big ones:</p><ol><li><strong>Bad Data = Bad AI: </strong>AI depends on good, diverse data. If your dataset is biased or incomplete, your product will reflect that. Ever had a music app recommend the same 5 songs on repeat? That’s what happens when your training data isn’t broad enough.</li><li><strong>Unpredictable Results: </strong>AI doesn’t always play nice. Sometimes, it spits out results no one saw coming. Test, test, and test again. Teams are even creating fake (synthetic) data to simulate tricky scenarios.</li><li><strong>Ethics and Bias: </strong>No one wants an AI product that’s unfair or discriminatory. Model Bias is a also big problem, and it can show up in ways you don’t expect.</li><li><strong>Trust Issues: </strong>Users can be skeptical about AI, especially when they don’t understand how it works. More teams are building in explainability features for AI — basically a “here’s why I made this decision” section or a button. Eg: “What fraud signals show that this is a fradulent transaction”.</li></ol><h4>What’s next for AI-powered products? Here’s where things are headed:</h4><ol><li><strong>More User Control</strong>: Users will be able to fine-tune AI outputs to match their needs. Eg: In Fraud detection, users more-or-less decide which signals are a priority for them or which policy rules should be used to consider something as fraud.</li><li><strong>Real-Time Everything</strong>: AI will adapt on the fly, changing based on what’s happening right now — not what happened last month. Models are constantly trained with constantly new features and new patterns.</li><li><strong>No More Black Boxes</strong>: Explainable AI is going to be a must-have. Users (and teams) will demand transparency.</li></ol><h4>Why This Matters</h4><p>The way we build AI products is changing everything about product discovery. It’s not just about creating smarter software; it’s about creating better, more responsible tools that users trust and love. If you’re diving into AI product discovery, focus on your data, think about ethics, and keep the user front and center. It’s a challenge, but it’s also an opportunity to shape the future of what tech can do.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dcd45a1a1cd3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI-Powered Multi-modal Document Assistant]]></title>
            <link>https://medium.com/@prabhamatta/ai-powered-multi-modal-document-assistant-fcb0f5a41ebc?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcb0f5a41ebc</guid>
            <category><![CDATA[augmented-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[startup-ideas]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Wed, 20 Nov 2024 05:33:46 GMT</pubDate>
            <atom:updated>2024-11-22T22:46:53.330Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI-Powered Multi-modal Doc Insights Assistant</h3><p>This is my attempt to develop an intelligent multi-modal document analysis solution that provides multi-modal, context-aware interactions with documents(manual text, PDFs, or websites) for users like legal professionals, researchers, students, and business consultants. This app aims to simplify workflows, improve efficiency, and enhance the quality of insights derived from complex datasets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/998/1*QPCkQ-UpSJ1e5f2-41B2SA.png" /></figure><h4>Problem Details</h4><ul><li>Industry: Product Manager, Legal, Academic research, content marketing and Business consulting</li><li>Persona: Lawyers, Students, Product Managers/Business consultants</li></ul><h4>User Journey Stages &amp; Activities</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6ap4OSX2jtvwMssNTbxdsA.png" /></figure><p><strong>Step 1: Identify Information Sources -</strong>Information is scattered across multiple formats and platforms and significant time is spent finding the right resources among a big corpus of documents or websites. Eg: Research students manually collect and organize PDFs, books, or notes. Legal researchers search for specific documents (e.g., legal contracts, case law, or compliance regulations). And business consultants and product managers review scattered reports, competitor websites, or industry publications.</p><p><strong>Step 2: Access and Aggregate Content -</strong>Now, the user downloads files from different platforms, copies text manually, or scrapes websites using third-party tools.</p><p><strong>Step 3: Analyze and Process Text -</strong>The user manually reads through documents or relies on basic keyword search tools. Then he identifies sections of interest, such as legal clauses, research summaries, or competitive trends.</p><p><strong>Step 4: Formulate and Answer Questions -</strong>The user thinks about specific questions mentally or through notes, and then searches for answers by rereading text or using online search engines (e.g., Google). Searching for answers within text requires repeated reading or CTRL+F for keywords, which lacks context awareness.</p><p><strong>Step 5: Organize and Utilize Insights -</strong>The user compiles findings into notes, summaries, or reports for personal or professional use. And then shares insights with stakeholders(leadership team, or professors) or apply them to specific tasks (e.g., legal arguments, research papers, business strategies).</p><h4>Key Pain-points</h4><p><strong>Fragmented and Scattered Text Sources: </strong>Pain point in Step 1, Identify Information Sources</p><p><strong>Multiple formats: </strong>Pain point in Step 2, Access and Aggregate Content. Users must switch between multiple tools and formats, increasing inefficiencies.</p><p><strong>Manual and Time-Consuming Text Analysis</strong>: Pain point in Step 3,Analyze and Process Text. Reading, highlighting, and summarizing large texts takes significant time.</p><p><strong>Significant effort for non-tech savvy: </strong>Various formats can be inconsistent and require a significant amount of manual work to synthesize insights. Also, advanced tools for scraping websites or analysis are too complex and require technical skills. Thus a barrier for non-technical users.</p><p><strong>Lack of Context-Aware Search Tools: </strong>Pain point in Step 4, Formulate and Answer Questions. Keyword-based search tools often fail to provide context-sensitive answers leading to inconsistent insights. Also, advanced tools for scraping or analysis require technical skills, creating a barrier for non-technical users.</p><p><strong>Limited Support for Iterative Questioning: </strong>Current tools and workflows don’t allow users to ask follow-up questions or refine their queries to gain a deeper understanding.</p><h3><strong>Solution Ideas</strong></h3><h4><strong>Multi-modal Interactive Document Assistant with the following 2 solution options:</strong></h4><p><strong>Input</strong>: A centralized interface for users to input data via:</p><ul><li>Manual text entry (large text box for quick pasting or typing).</li><li>File upload (supporting PDFs and text files).</li><li>Web scraping (URL input with an optional checkbox to include child pages).</li></ul><p><strong>Interact with the documents -</strong></p><ul><li>Text (question) → Text (answers)</li><li>Text(question) → 1-min audio or 10-min audio(answer) ,</li><li>Audio(question) → text (answer) — may be some kind of formatted answer or diagrams</li><li>Audio (question)→ 1-min audio or 10-min audio(answer)</li></ul><h4><strong>Solution1: Automated Summarization Tool</strong></h4><p>A summarization engine that condenses lengthy documents or scraped content into concise overviews.</p><ul><li><strong>Input</strong>: A centralized interface for users to input data via:</li><li>Manual text entry (large text box for quick pasting or typing).</li><li>File upload (supporting PDFs and text files).</li><li>Web scraping (URL input with an optional checkbox to include child pages).</li></ul><p><strong>Output</strong></p><ul><li>Summary at varying levels of detail (e.g., key points, short summary, full overview).</li><li>Ability to focus on specific sections or topics based on user-defined keywords or queries.</li><li>Exportable summaries in text or PDF format.</li></ul><h4><strong>Solution2: AI-Powered Question Answering</strong></h4><p>A smart question-answering assistant powered by ChatGPT API to process and respond to user queries based on the uploaded/scraped content.</p><p><strong>Key Features</strong>:</p><ul><li>Natural language question input with context-aware responses.</li><li>Highlighting of relevant text segments in the source material.</li><li>Iterative questioning support to refine or follow up on answers.</li><li>Customizable response formats (short answer, detailed explanation, or bullet points).</li></ul><p><strong>Solution4: Semantic Search Engine</strong></p><p>A search tool that allows users to search uploaded or scraped content using semantic understanding instead of basic keyword matching.</p><p><strong>Key Features</strong>:</p><ul><li>Context-sensitive search that understands synonyms and related concepts.</li><li>Highlighting relevant text with explanations of why it matches the query.</li></ul><h4>3 Key Value propositions</h4><p>By integrating LLMs at every stage of the user journey, the application will:</p><ol><li>Provide intelligent, context-aware assistance for faster and more accurate insights.</li><li>Enable seamless integration of diverse text sources for holistic analysis.</li><li>Simplify web scraping and make it accessible to all users.</li></ol><h3>Assumptions &amp; Risks</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n9J1Ue6fwEtsJSuqfSU9OQ.png" /></figure><ul><li>The assumption is that the users will find iterative questioning (follow-ups and refinements) valuable for deeper insights. If users rarely engage with follow-up questions, this app might become underutilized, reducing perceived product differentiation.</li><li>The assumption is that the AI will successfully synthesize from diverse formats (e.g., PDFs, text files, and scraped websites) to provide unified, actionable insights, or else it will lead to user frustration, undermining the app’s core value proposition.</li><li>The assumption is that the users will trust the AI’s ability to provide accurate, context-aware insights and answers. Users may lose confidence and abandon the platform if the AI frequently provides incorrect, irrelevant, or poorly contextualized answers.</li><li>The assumption is that the LLM will be able to process complex or domain-specific text and generate meaningful insights across a wide range of input types, such as legal contracts, academic papers, and web content. If LLM cannot handle complex or domain-specific text, it may fail to meet user expectations.</li><li>The application assumes that it is possible to validate the answers. Therefore, ensuring the accuracy of the output is very important for this application. However, it is very difficult to validate the accuracy without the help of subject matter experts as we do not know what domain-specific text is given to the application for analysis.</li><li>The assumption is that the users will trust the platform with sensitive documents and data because the app provides robust security and privacy safeguards. Therefore, data security is of utmost importance. Any data security risks could deter users from uploading sensitive materials, particularly legal or confidential business documents.</li></ul><h3>What does success look like?</h3><h4>User Value</h4><p><strong>Centralized data management</strong></p><ul><li>Users can upload, scrape, or manually enter text in one unified platform.</li><li>Seamlessly handle diverse input types (manual, files, and web content) without switching between tools.</li></ul><p><strong>Time Savings</strong></p><ul><li>Users save hours previously spent on manual data aggregation, analysis, and insight generation.</li><li>Further, they can instantly query uploaded or scraped content using AI for context-aware answers</li><li>And access quick summaries and insights without reading through lengthy documents.</li></ul><p><strong>Scalable Insights</strong></p><ul><li>Users can process and analyze larger datasets or documents without scaling costs or complexity.</li><li>Can ask follow-up or refined questions to explore data more thoroughly.</li><li>Uncover hidden patterns, relationships, or details that basic keyword searches miss.</li></ul><p><strong>Improved Accuracy</strong></p><ul><li>AI eliminates human errors, providing precise and context-aware insights.</li></ul><p><strong>Increased Productivity</strong></p><ul><li>Users can focus on strategic tasks, leaving repetitive and labor-intensive processes to the app.</li></ul><p><strong>Accessibility for All Users</strong></p><ul><li>Non-technical users gain access to advanced analysis tools, leveling the playing field across industries.</li></ul><p>Success means empowering users to <strong>achieve more with less effort</strong>, transforming their workflows, and <strong>solving problems faster and smarter</strong>. The app will become an indispensable tool for professionals across industries, delivering measurable time savings, enhanced productivity, and improved outcomes.</p><h4>Success Metrics</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i3lM__hymjnytiHT2gPJdw.png" /></figure><p><strong>North Star Metric (NSM) — </strong>“Number of Questions Answered Per Active User Per Week”</p><h4><strong>User Metrics</strong></h4><p><strong>Adoption Metrics</strong></p><ul><li>Number of signups, First-Time Upload Rate, adoption of features</li></ul><p><strong>Engagement Metrics</strong></p><ul><li>DAU, WAU, MAU, Session length, number of queries per session, Multi-source engagement</li></ul><p><strong>Retention Metrics</strong></p><ul><li>Repeat users — Percentage of users who return to the platform after their first session</li><li>Churn Rate — Percentage of users who stop using the platform after a certain time.</li><li>Feature Reuse Rate — Tracks how frequently specific features (e.g., web scraping, question answering) are used over time.</li></ul><h4><strong>Value Metrics</strong></h4><ul><li>Time to Insight — Average time it takes users to receive meaningful results after uploading content or asking a question.</li><li>Question Refinement Rate — Percentage of queries followed by refinements or follow-up questions.</li><li>NPS score</li></ul><h4><strong>Growth Metrics</strong></h4><ul><li>Referral Rate — Percentage of new users who sign up via existing user referrals</li><li>Feature Adoption Over Time — Tracks how quickly new features are adopted after being introduced.</li></ul><h4>Business Metrics</h4><ul><li>Revenue Metrics — MRR, ARR, Average Revenue Per User (ARPU)</li><li>User Acquisition Metrics — Sign-Up Conversion Rate, Customer Acquisition Cost (CAC)</li><li>User Retention Metrics — See “User metrics” section</li><li>Growth Metrics — User Growth Rate, Referral Rate</li><li>Operational Metrics — Cost per Query, Infrastructure Costs</li></ul><p>— — — — — — — — — —</p><p><strong><em>Demo - Coming soon. Stay tuned!! </em></strong>The basic conceptual version of the app is WIP….will share the demo soon.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V8SVzuChF0Y3LU8MA8UFrg.png" /></figure><p>— — — —</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fcb0f5a41ebc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is RAG? How do you use it in your AI application?]]></title>
            <link>https://medium.com/@prabhamatta/what-is-rag-how-do-you-use-it-in-your-ai-application-73bfa5e5400f?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/73bfa5e5400f</guid>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Wed, 13 Nov 2024 00:47:03 GMT</pubDate>
            <atom:updated>2024-11-21T23:14:48.828Z</atom:updated>
            <content:encoded><![CDATA[<p>There are multiple videos and articles and detailed research papers explaining these concepts. Here, I want to give a very basic explanation of RAG and how it is used in <a href="https://medium.com/@prabhamatta/ai-powered-multi-modal-document-assistant-fcb0f5a41ebc">your chat application</a> in the simplest terms.</p><h3><strong>What is RAG?</strong></h3><p>RAG is an acronym for Retrieval Augmented Generation. In simple terms, if you have data or some sources of data at hand, you can use that to answer questions by giving it to the LLM.</p><h3><strong>How does it work? How do you use it in your AI application?</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-tUTKCDv6tDWkHp9Co6o2w.png" /><figcaption>RAG process flow diagram</figcaption></figure><p><strong>Step1: Gather your data</strong></p><p>The first step is to gather all of your data from different data sources. For instance, you can have your data source from text or pdf files, you can have it in websites. Then you add all your sources in one place what we call a RAG pipeline, where you chunk data from all these sources.</p><p><strong>Step2: Chunk your data</strong></p><p>Now, there are multiple techniques for chunking, but the simplest one is to chunk after a specific number of letters or words or phrases. Now you have multiple chunks and every chunk has a meaning behind it or is talking about a specific thing. The trick is to optimize this chunking process to get your app to work well.</p><p><strong>Step3. Embeddings on the chunks</strong></p><p>Then comes the step of Embedding all of those chunks. Embedding is just turning the text into an array of numbers, which are called vectors. Those vectors represent the meaning in those texts and the semantics behind them. Then you store these vectors in a vector database. Now, you have all of your sources vectorized and stored in a vector database.</p><p><strong>Step4. Get the User question</strong></p><p>When a user asks a question, you do the same thing ie., you chunk and embed the question. Now, you have the semantic representation of the question as well.</p><p><strong>Step5. Get the relevant chunks</strong></p><p>And then you go to the vector database and get the most similar chunks to that question. And the most similar chunks to the question will probably be something that talks about the same thing. And it will probably have the answer to that question.</p><p>You might have found multiple chunks that talk about that topic. You can retrieve/choose the top three chunks that talk about this question topic or top three paragraphs.l containing those chunks.</p><p><strong>Step6. Feed into LLM</strong></p><p>Now, you have a question and you have some data that talk about the same topic and the probability that this data has the answer to the question might be very high. So, the last step would be to make a request to an LLM and ask “Given this information, can you answer this question?”. The LLM should be able to answer it if the information is in the question.</p><p>By the way, below is a cool AI-generated flowchart for the same above process:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ewh7UrotJ57x9BQ3nt2zLQ.png" /><figcaption>RAG Process Explained!</figcaption></figure><p>— — —</p><p>Note that your application needs to be tested, and data needs to be validated before deploying it into prod. Perplexity’s CEO articulates this very well; he talks about how Perplexity targets Accuracy, Latency, and Readability…and further on Reliability, the spot-on dimensions.</p><p><a href="https://www.linkedin.com/posts/pmatta_inside-how-perplexity-is-disrupting-online-activity-7261550391459233793-Drvb">https://www.linkedin.com/posts/pmatta_inside-how-perplexity-is-disrupting-online-activity-7261550391459233793-Drvb</a><a href="https://www.linkedin.com/posts/pmatta_inside-how-perplexity-is-disrupting-online-activity-7261550391459233793-Drvb?utm_source=share&amp;utm_medium=member_desktop">p</a></p><p>— — —</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73bfa5e5400f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to create a no-code AI Chatbot]]></title>
            <link>https://medium.com/@prabhamatta/how-to-create-a-no-code-ai-chatbot-bbf6b90a131e?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/bbf6b90a131e</guid>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Wed, 30 Oct 2024 20:37:11 GMT</pubDate>
            <atom:updated>2024-11-01T16:08:28.873Z</atom:updated>
            <content:encoded><![CDATA[<p>This morning when I was brushing my teeth, I thought why don’t I build a simple AI chatbot answering questions regarding my skillset and professional work experiences 🤔</p><p>Well, here we go, this is the <a href="https://app.droxy.ai/guest-agent/67228b96c39918a6b638a69d">link to the AI chatbot</a>. This is really a draft version and might be a little rusty, but wanted to share it anyways :-)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/912/1*vIFC4ZFyxkvvjyOtLBW32Q.png" /><figcaption>Output from my fun little chatbot</figcaption></figure><h3>How can you build it?</h3><p>I used the <a href="https://www.droxy.ai/">droxy.ai</a> free version to make this chatbot. It has a simple UI and a very user-friendly configuration setup. Though I wanted to embed this chatbot in my website, I could not do it since that feature is behind their pay wall 😒</p><p>A picture is worth more than a thousand words so I am posting the screenshots of my chatbot to help you understand various steps. Also, I am too lazy to type today!</p><p><strong>Step1: Upload your data resources</strong></p><p>This is the place where you upload the data eg: Google Drive folder containing docs, your website, your videos or even PDFs. Droxy enables you to build chatbots using these various content formats. This content will serve as your chatbot’s training material.</p><figure><img alt="This is where you upload your data resources" src="https://cdn-images-1.medium.com/max/1024/1*TOz0h3AqlPAj1JnmrM5mNw.png" /><figcaption>This is where you upload your data resources</figcaption></figure><figure><img alt="I used my website as a data resource" src="https://cdn-images-1.medium.com/max/1024/1*TsUm70wF4Jb4spg_yhj7cw.png" /><figcaption>I used my website as a data resource along with some docs</figcaption></figure><p><strong>Step2: Create a new chatbot with the data souces</strong></p><p>Click “Add” on the top right corner to create a new AI agent. The first thing you need to do is select the data sources you want to use for your chatbot. I used my website and uploaded a couple of documents to be used as data resources for my chatbot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iZTw9vc8Qc_NAVmZFYHtbA.png" /><figcaption>Create a Basic agent (FREE version)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vxbSPxaWyQKoqBcMnPJyMg.png" /><figcaption>Select the resource you want to use to train the ChatGPT model</figcaption></figure><p><strong>Step3: Create a personality for your Chatbot</strong></p><p>After selecting your materials, you can tailor the chatbot’s personality, appearance, and tone. The chatbot’s personality is primarily determined by the prompt you provide. The prompt should give details on how you want the chatbot to act. eg: sales manager, service representative, admissions manager, etc. You can also give the details on the tone you want the chatbot to use eg: confidence, compassion, etc</p><figure><img alt="Prompt should give details on how you want the chatbot to act like" src="https://cdn-images-1.medium.com/max/1024/1*CTUfd9oBGlVZLFnShLH42w.png" /><figcaption>Prompt should give details on how you want the chatbot to act like</figcaption></figure><p><strong>Step 4: Customize the look and feel of your chatbot</strong></p><p>This can be done by selecting the Color theme, Welcome message, Title of your chatbot, etc.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ewTQ6HxXVPuIuQEfDp2GvQ.png" /><figcaption>Customize the look and feel of your chatbot</figcaption></figure><p>Save…., and Voila! Your chatbot is ready to go! You will be provided with a link that you can either embed in your website as an iframe (which I am currently lazy to do as my website was launched on Digital Ocean and it is quite a pain to edit my old HTML code) or you can link as a helper bot to the bottom right corner or simply add/share the link in your LinkedIn profile page!</p><p>Please feel free to <a href="https://app.droxy.ai/guest-agent/67228b96c39918a6b638a69d">try my chatbot</a>; also share in the comments on the chatbots you created. I am excited to try them out.</p><p>Cheers!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bbf6b90a131e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pricing Strategy: Price Sensitivity Meter Analysis]]></title>
            <link>https://medium.com/@prabhamatta/pricing-strategy-price-sensitivity-meter-analysis-bbfb9e188ce9?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/bbfb9e188ce9</guid>
            <category><![CDATA[product-analytics]]></category>
            <category><![CDATA[pricing-strategy]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Sat, 09 Jun 2018 19:37:18 GMT</pubDate>
            <atom:updated>2024-02-11T11:48:31.708Z</atom:updated>
            <content:encoded><![CDATA[<p>There are multiple ways to come up with pricing strategy and <strong>Van Westendorp Price Sensitivity Meter model</strong> is one of the quantitative methods to find <strong>“Willingness to Pay”</strong> by a customer. As the name suggests, this analysis is based on customer’s feedback and their expectation regarding the value of the product.</p><h3>Step 1. Collect the data</h3><p>Conduct a survey to get the pulse of the users for these 4 questions.</p><ul><li>At what monthly price is this way too expensive that you would never consider purchasing it?</li><li>At what monthly price is this starting to get expensive, but you’d still consider purchasing it?</li><li>At what monthly price is this a really good deal? (You’d buy it right away.)</li><li>At what monthly price is this way too cheap that you’d question the quality of it?</li></ul><p>Cadence should be taken care in the text while framing the questions. For all the above questions, we need to ask the questions specifically for monthly price, else we might get wrong data due to ambiguity.</p><p>From the survey, data looks something like this -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*whtqms1GO91pUcsJI9ioBw.png" /><figcaption>Survey Data for Price Sensitivity Meter Analysis</figcaption></figure><h3>Step 2. Validate the data</h3><p>Validate the data to make sure we take correct data points from each respondent. In Excel, you can create your formula like this:</p><p><strong>IF(AND(A2&gt;=B2, B2 &gt;= C2, C2 &gt;= D2), “PASS”, “FAIL”)</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gKcD3tx53rEryh_UNZeEjw.png" /><figcaption>Validate the data</figcaption></figure><h3>Step 3. Process the data</h3><p>Create your buckets based on the data. If the data is too nuanced, you can create the bucket size as 5 and if the data is too varied, you can create the bucket size as 50. Use Excel Histogram tool to create frequency distribution of data for each question for our user-defined buckets.</p><p>Now, compute Cumulative Frequency from the above processed data and it should look in this format.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/952/1*ALW6XVsfnkNEP1wLEvwNCQ.png" /><figcaption>Process the data</figcaption></figure><p>For some survey questions we need derived data points. Create inverse Cumulative Frequency for <strong>Expensive </strong>data points as <strong>Not Expensive, </strong>for<strong> Not Cheap(Bargain) </strong>data points<strong> </strong>as<strong> Cheap </strong>and for<strong> Too Cheap.</strong></p><h3>Step 4. Plot the graph</h3><p>Plot the graph for all the data points and label as below</p><ul><li><strong>Point of Marginal Cheapness(PMC)</strong> = lower bound of an acceptable price range = intersection of <strong>Too Cheap </strong>and <strong>Expensive </strong>lines</li><li><strong>Point of Marginal Expensiveness(PME)</strong> = upper bound of an acceptable price range = intersection of <strong>Too Expensive</strong> and <strong>Cheap</strong> lines</li><li><strong>Indifference Price Point (IPP)</strong> = intersection of <strong>Expensive </strong>and <strong>Cheap</strong> lines</li><li><strong>Optimal Price Point (OPP)</strong> = intersection of <strong>Too Cheap</strong> and <strong>Too Expensive</strong> lines</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MWpvLrmp5yFkAZQb" /><figcaption>Price Sensitivity Meter</figcaption></figure><p>From the above analysis, acceptable price range for our product is <strong>$140 - $170 per month.</strong></p><p>Reference: <a href="https://en.wikipedia.org/wiki/Van_Westendorp%27s_Price_Sensitivity_Meter">Wikipedia</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bbfb9e188ce9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Product Idea: Sublet My Place (Part 2)]]></title>
            <link>https://medium.com/@prabhamatta/product-idea-sublet-my-place-part-2-7622856e1974?source=rss-f607d030fdd9------2</link>
            <guid isPermaLink="false">https://medium.com/p/7622856e1974</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[ideas]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Prabha Matta]]></dc:creator>
            <pubDate>Sat, 12 May 2018 19:00:58 GMT</pubDate>
            <atom:updated>2024-06-22T07:41:43.659Z</atom:updated>
            <content:encoded><![CDATA[<h3>Product Roadmap Process: Sublet My Place (Part 2)</h3><p>In <a href="https://medium.com/@prabhamatta/product-idea-sublet-my-place-part1-e5569260b43c">Part 1</a>, we looked at business goals, personas and feature prioritization. Now let’s proceed to user flows.</p><h3>User Flows</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/995/1*GFLjrA_9S6IiybjmxO3yIQ.png" /><figcaption>Sample User Flow Diagrams</figcaption></figure><h3>Product Roadmap: OKRs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*8SOluuMKdACweTcyFAEjpQ.png" /><figcaption>Sample Product Roadmap: OKRs and KPIs</figcaption></figure><h3>Product Roadmap: Main Themes</h3><p>Themes are color coded to identify them in later stages easily.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/625/1*_WPqPjCi1fJIyWrrjNeltA.png" /><figcaption>Product Roadmap: Themes</figcaption></figure><h3>Product Roadmap: Priorities</h3><p>Prioritized features of <a href="https://medium.com/@prabhamatta/product-idea-sublet-my-place-part1-e5569260b43c">Part 1</a> are assigned to individual themes. Features related to each other are color coded based on the <strong>themes</strong> and are planned to be released at the same time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*XsXrLlRD6oXzA5HN9QbP3A.png" /><figcaption>Product Roadmap: Priorities</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*QkNPIubeGV5x0xmbn4Kz1Q.png" /><figcaption>Product Roadmap: Priorities</figcaption></figure><h3>Product Roadmap: Release Schedule</h3><p>Features in black font indicate the ones to be released in Beta version and red font ones will wait for future releases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*kPaLfzzvqEUgwHf2DiVhng.png" /><figcaption>Sample Product Roadmap: Release Schedule</figcaption></figure><h3><strong>Design (Lo-fi Prototypes)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/889/1*RDAq1WyovyBvZpZVRyDp0Q.png" /><figcaption>Main Page</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/644/1*15ulN9HNYii6b279BIO_4A.png" /><figcaption>Sublet Details</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/663/1*6PyRDm7p6WwncaiOeHKc9w.png" /><figcaption>Renter’s view</figcaption></figure><h3><strong>User Stories and Acceptance Criteria</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/701/1*f0cM9HB__Daftd_EP7J15Q.png" /><figcaption>Sample Acceptance Criteria</figcaption></figure><h3><strong>Trello Board for Q1 Releases</strong></h3><p>Here is the link to a <a href="https://trello.com/b/mW5maW7u/sublet-my-place">sample Trello Board</a> to see details of the plan for releases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Trl044JwsH6M983BCn48hw.png" /><figcaption>Sample <a href="https://trello.com/b/mW5maW7u/sublet-my-place">Trello Board</a></figcaption></figure><h3><strong>Yay!! Finally, Release Notes!</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/616/1*Old1q5t8xUfBQpjI9WrsSA.png" /><figcaption>Sample Release Notes</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7622856e1974" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>