<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Tilak Raj Roy</title>
    <description>The latest articles on DEV Community by Tilak Raj Roy (@tilak_rajroy_f81d1723578).</description>
    <link>https://dev.to/tilak_rajroy_f81d1723578</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4078498%2F29c4e4d3-5fc4-4e92-8a4b-d77ceba7a77c.png</url>
      <title>DEV Community: Tilak Raj Roy</title>
      <link>https://dev.to/tilak_rajroy_f81d1723578</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tilak_rajroy_f81d1723578"/>
    <language>en</language>
    <item>
      <title>VaultAudit AI</title>
      <dc:creator>Tilak Raj Roy</dc:creator>
      <pubDate>Wed, 14 Oct 2026 15:35:15 +0000</pubDate>
      <link>https://dev.to/tilak_rajroy_f81d1723578/vaultaudit-ai-1ofi</link>
      <guid>https://dev.to/tilak_rajroy_f81d1723578/vaultaudit-ai-1ofi</guid>
      <description>&lt;p&gt;Here is your completed submission post, written specifically for &lt;strong&gt;VaultAudit AI&lt;/strong&gt; and following the official Hacktoberfest Week 2 template:&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;I built &lt;strong&gt;VaultAudit AI&lt;/strong&gt; 🔒⚡ — an on-device, zero-knowledge financial intake, receipt parser, and tax-deductible expense sorter designed for freelancers, solopreneurs, and independent contractors.&lt;/p&gt;

&lt;h3&gt;
  
  
  The Problem It Solves
&lt;/h3&gt;

&lt;p&gt;Freelancers and boutique agencies handle sensitive financial documents every day: subcontractor invoices, flight &amp;amp; client travel receipts, hosting bills, and payment records. However, preparing these for tax season or client expense reimbursement presents critical hurdles:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;NDA &amp;amp; Privacy Violations:&lt;/strong&gt; Uploading confidential invoices with client names, hourly billing rates, bank wire routing details, and proprietary deliverables to commercial cloud AI vendors (OpenAI, Anthropic, cloud OCR APIs) breaks strict client non-disclosure agreements.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Subscription Fatigue:&lt;/strong&gt; Paying $15–$50/month for cloud expense tracking apps or per-token API costs drains freelance profit margins.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Connectivity Limitations:&lt;/strong&gt; Traveling consultants and digital nomads frequently need to reconcile receipts on flights, trains, or in locations with restricted network access.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  Who Would Pay For It?
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Freelancers &amp;amp; Contractors:&lt;/strong&gt; Willing to pay a one-time license fee ($29–$49) for a lifetime utility that replaces recurring monthly SaaS subscriptions and guarantees client data privacy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CPAs &amp;amp; Boutique Bookkeepers:&lt;/strong&gt; Willing to buy agency packs ($149+) to distribute to privacy-conscious clients who refuse to let their financial data sit on third-party cloud servers.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Interactive Local App:&lt;/strong&gt; Runs directly on &lt;code&gt;localhost:5173&lt;/code&gt; with a modern dark glassmorphism interface.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Key Features Demonstrated:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Instant Multi-Scenario Ingestion:&lt;/strong&gt; Pre-packaged templates for AWS cloud bills, Uber business transit receipts, subcontractor development invoices, and Apple hardware depreciation items.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero-Latency Ingestion:&lt;/strong&gt; Paste raw text or upload receipt files for immediate extraction.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automated Tax Tagging:&lt;/strong&gt; Accurately classifies categories (SaaS, Hardware, Travel, Subcontractors) and IRS Section 179 eligibility accompanied by clear audit justifications.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local Persistence &amp;amp; Export:&lt;/strong&gt; Stores all reviewed entries in browser storage and provides one-click export to accountant-ready CSV (compatible with QuickBooks, TurboTax, and Excel).&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;(Include your demo video, Loom link, or GIF screenshots of the UI processing receipts here)&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;The complete source code is available on GitHub:&lt;/p&gt;

&lt;p&gt;{&lt;a href="https://github.com/tilakraj1108/VaultAudit-AI" rel="noopener noreferrer"&gt;https://github.com/tilakraj1108/VaultAudit-AI&lt;/a&gt;}&lt;/p&gt;

&lt;h3&gt;
  
  
  Stack &amp;amp; Structure
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; Pure HTML5, Vanilla CSS3 (modern glassmorphism design system, typography via Inter &amp;amp; JetBrains Mono), and modern ES6 JavaScript.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local Bridge Backend:&lt;/strong&gt; Python standard library (&lt;code&gt;http.server&lt;/code&gt;, &lt;code&gt;urllib.request&lt;/code&gt;) — requiring &lt;strong&gt;zero external pip dependencies&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference Runtime:&lt;/strong&gt; Local &lt;a href="https://ollama.ai" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt; engine with strict JSON schema decoding.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;VaultAudit AI is architected around &lt;strong&gt;open-weight Llama 3.2 (3B)&lt;/strong&gt; running 100% locally on user hardware.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Open-Source AI Engine
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; Meta's open-weight &lt;code&gt;llama3.2:3b&lt;/code&gt; running via &lt;strong&gt;Ollama&lt;/strong&gt;. With 4-bit quantization, it runs in ~2.0 GB of RAM, making it fast and lightweight on standard modern laptops without requiring cloud GPUs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Prompt Architecture:&lt;/strong&gt; The bridge sends a structured system prompt enforcing a strict JSON schema:

&lt;ul&gt;
&lt;li&gt;Merchant identification &amp;amp; invoice numbers&lt;/li&gt;
&lt;li&gt;Line-item parsing (descriptions, quantities, unit prices)&lt;/li&gt;
&lt;li&gt;Tax calculation &amp;amp; currency resolution&lt;/li&gt;
&lt;li&gt;Deductibility classification with audit justifications&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic Extraction:&lt;/strong&gt; Uses low temperature (&lt;code&gt;0.1&lt;/code&gt;) and Ollama's native &lt;code&gt;format: "json"&lt;/code&gt; mode to eliminate hallucinated JSON formatting errors.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. Zero-Knowledge Local Pipeline
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;The local Python server acts as a clean bridge between the browser frontend and the Ollama REST API (&lt;code&gt;http://127.0.0.1:11434&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No telemetry, no tracking, and no external requests.&lt;/strong&gt; If the network cord is unplugged, the application functions identically.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Open innovation is what makes a privacy-first utility like VaultAudit AI possible:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;True Data Sovereignty:&lt;/strong&gt; Closed cloud APIs invariably require transmitting data packets over the public internet to third-party data centers, where data retention and model training policies can be opaque. Open weights run completely in local memory — client bank numbers and invoices never leave the machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Infinite Inference at Zero Marginal Cost:&lt;/strong&gt; Running 10,000 document audits through proprietary APIs accumulates token fees that make a low-cost, one-time payment desktop tool economically unviable. Open weights reduce inference cost to &lt;strong&gt;$0.00&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Resilience &amp;amp; Independence:&lt;/strong&gt; Closed APIs can deprecate endpoints, change pricing tiers, or enforce unexpected rate limits. Open-source models grant complete autonomy — the software works today, tomorrow, and offline indefinitely.&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;




&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;p&gt;During development, I tested VaultAudit AI with a freelance full-stack engineer who bills multiple overseas clients under strict NDAs. &lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Their Feedback:&lt;/strong&gt; &lt;em&gt;"Normally, sorting through contractor invoices and cloud bills takes me hours at the end of the month, and I couldn't use cloud AI tools because my enterprise client contracts forbid uploading project descriptions and billing information. Running this locally took under a minute for 15 receipts and gave me a clean CSV ready for my accountant."&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Outcome:&lt;/strong&gt; Validated the $29 one-time pricing model as an instant purchase over recurring cloud accounting add-ons.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Primary Track:&lt;/strong&gt; Side Hustle (Hacktoberfest Open-Source AI Challenge Week 2)&lt;/li&gt;
&lt;/ul&gt;




</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>🌿 TrailQuest AI: Using Local AI to Get People Outside</title>
      <dc:creator>Tilak Raj Roy</dc:creator>
      <pubDate>Fri, 09 Oct 2026 07:34:31 +0000</pubDate>
      <link>https://dev.to/tilak_rajroy_f81d1723578/trailquest-ai-using-local-ai-to-get-people-outside-1el7</link>
      <guid>https://dev.to/tilak_rajroy_f81d1723578/trailquest-ai-using-local-ai-to-get-people-outside-1el7</guid>
      <description>&lt;h1&gt;
  
  
  🌿 TrailQuest AI: Using Local AI to Get People Outside
&lt;/h1&gt;

&lt;p&gt;What if AI could help us spend less time on our screens and more time in the real world?&lt;/p&gt;

&lt;p&gt;That question inspired me to build &lt;strong&gt;TrailQuest AI&lt;/strong&gt;, a local-first AI application that transforms free time into personalized outdoor adventures.&lt;/p&gt;

&lt;p&gt;Instead of using AI only for answering questions or generating content, I wanted to build something that encourages people to step outside, observe nature, and explore their surroundings.&lt;/p&gt;

&lt;h2&gt;
  
  
  🌱 What I Built
&lt;/h2&gt;

&lt;p&gt;TrailQuest AI generates personalized nature quests based on a user's location, available time, interests, and preferred pace.&lt;/p&gt;

&lt;p&gt;For example, imagine you have 30 minutes free and want to explore a nearby park. TrailQuest AI can generate a mission centered around observing birds, noticing different plants, listening to outdoor sounds, or taking a mindful nature walk.&lt;/p&gt;

&lt;p&gt;The goal is simple: &lt;strong&gt;let AI plan the adventure, then put the screen away and experience it.&lt;/strong&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  ✨ Key Features
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;🧭 &lt;strong&gt;Personalized Nature Quests:&lt;/strong&gt; Generate activities based on your preferences and available time.&lt;/li&gt;
&lt;li&gt;🐦 &lt;strong&gt;Bird and Wildlife Exploration:&lt;/strong&gt; Discover simple observation activities to try outdoors.&lt;/li&gt;
&lt;li&gt;✅ &lt;strong&gt;Interactive Checklist:&lt;/strong&gt; Track your progress as you complete your mission.&lt;/li&gt;
&lt;li&gt;⏱️ &lt;strong&gt;Activity Timer:&lt;/strong&gt; Keep your outdoor adventure within your planned duration.&lt;/li&gt;
&lt;li&gt;📋 &lt;strong&gt;Take It With Me:&lt;/strong&gt; Copy your mission before leaving your screen behind.&lt;/li&gt;
&lt;li&gt;🔒 &lt;strong&gt;Local AI Inference:&lt;/strong&gt; Generate quests using an AI model running on your own computer.&lt;/li&gt;
&lt;li&gt;🧪 &lt;strong&gt;Preview Mode:&lt;/strong&gt; Explore the app even when the local AI model is unavailable.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  🎬 Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Run it locally:&lt;/strong&gt; Follow the setup instructions in the GitHub repository.&lt;/p&gt;

&lt;p&gt;The demo will show the complete workflow: selecting preferences, generating a nature quest with local AI, using the checklist, and preparing to head outdoors.&lt;/p&gt;

&lt;h2&gt;
  
  
  💻 GitHub Repository
&lt;/h2&gt;

&lt;p&gt;🔗 &lt;strong&gt;Source code:&lt;/strong&gt; &lt;a href="https://github.com/tilakraj1108/trailquest-ai" rel="noopener noreferrer"&gt;https://github.com/tilakraj1108/trailquest-ai&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repository includes the application code, Python backend, local AI integration, setup instructions, and project documentation.&lt;/p&gt;

&lt;h2&gt;
  
  
  🛠️ How I Built It
&lt;/h2&gt;

&lt;p&gt;I built TrailQuest AI using a lightweight web application architecture.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tech stack:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; HTML, CSS, and JavaScript&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; Python&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI runtime:&lt;/strong&gt; Ollama&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI model:&lt;/strong&gt; Llama 3.2 3B, an open-weight language model&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference:&lt;/strong&gt; Local model inference instead of a paid, remote AI API&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  🧠 How the AI Works
&lt;/h3&gt;

&lt;p&gt;The application collects the user's preferences and sends them to the local language model through Ollama.&lt;/p&gt;

&lt;p&gt;The model generates a nature quest containing practical activities and instructions. The application then presents that quest through its interface so the user can track their progress.&lt;/p&gt;

&lt;p&gt;The app also includes a preview mode, which lets people explore the interface without having the model installed. Preview output is separate from actual model-generated output.&lt;/p&gt;

&lt;h2&gt;
  
  
  🔓 Why Open Innovation Matters
&lt;/h2&gt;

&lt;p&gt;For this project, using an open-weight model makes local AI inference possible without depending on a paid hosted AI API.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Privacy and Control
&lt;/h3&gt;

&lt;p&gt;With local inference, the user's quest preferences can be processed on their own computer instead of being sent to a remote AI service for generation. This provides greater control over personal data.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Lower Running Costs
&lt;/h3&gt;

&lt;p&gt;After downloading the model, users can generate quests locally without paying a per-request fee to a hosted AI provider. Running the model still requires suitable computer hardware and electricity.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Freedom to Experiment
&lt;/h3&gt;

&lt;p&gt;Ollama makes it practical to experiment with different compatible models and compare their generated quests. I can explore different prompts and model configurations without rebuilding the entire application.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Learning Through Open Technology
&lt;/h3&gt;

&lt;p&gt;Building with local AI has helped me explore prompt engineering, backend integration, model inference, and the practical limitations of smaller language models.&lt;/p&gt;

&lt;p&gt;A closed AI API could also generate nature quests. However, local inference&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Friend Study Buddy: A Simple AI-Powered Study Companion</title>
      <dc:creator>Tilak Raj Roy</dc:creator>
      <pubDate>Sat, 03 Oct 2026 07:46:16 +0000</pubDate>
      <link>https://dev.to/tilak_rajroy_f81d1723578/friend-study-buddy-a-simple-ai-powered-study-companion-2pp8</link>
      <guid>https://dev.to/tilak_rajroy_f81d1723578/friend-study-buddy-a-simple-ai-powered-study-companion-2pp8</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01"&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Friend Study Buddy&lt;/strong&gt; is a clean, accessible, and lightweight study companion web application designed specifically for students and study groups. &lt;/p&gt;

&lt;p&gt;I built this for my friend and study partner who often gets overwhelmed when trying to plan study sessions for technical subjects like Data Structures, Web Development, and Computer Science fundamentals. During exam crunch times and hackathons, they struggled with:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Deciding how much time to allocate between learning concepts, practicing problems, and doing final revisions.&lt;/li&gt;
&lt;li&gt;Quickly testing fundamental concepts without getting distracted by heavy, complex quiz platforms.&lt;/li&gt;
&lt;li&gt;Finding concise, cheat-sheet-style revision notes for quick refreshers.&lt;/li&gt;
&lt;li&gt;Staying motivated during late-night study marathons.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Friend Study Buddy&lt;/strong&gt; solves this by putting everything in one distraction-free place:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Dynamic Study Planner&lt;/strong&gt;: Calculates tailored time breakdowns (concept learning, examples, hands-on practice, revision, and quick self-testing) based on custom durations and difficulty levels.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Quick CS Quiz&lt;/strong&gt;: An interactive multiple-choice quiz testing core CS fundamentals with instant score calculations and encouraging feedback.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Quick Revision Notes&lt;/strong&gt;: Instant flashcards for HTML, CSS, JavaScript, and DSA fundamentals.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Motivation Generator&lt;/strong&gt;: A random affirmation generator to keep spirits high during tough study sessions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Night-Owl Friendly Dark Mode&lt;/strong&gt;: An accessible, soothing dark theme to reduce eye strain during late-night study sessions.&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live Demo Link&lt;/strong&gt;: &lt;a href="https://tilakraj1108.github.io/friend-study-buddy/" rel="noopener noreferrer"&gt;https://tilakraj1108.github.io/friend-study-buddy/&lt;/a&gt; &lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Preview / Walkthrough&lt;/strong&gt;: A zero-dependency, lightning-fast static web app that runs smoothly on mobile, tablet, and desktop screens with full dark/light theme support.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;The project is built with clean, semantic open web technologies (HTML5, CSS3, and Vanilla JavaScript ES6+) with zero build tools or dependencies required.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GitHub Repository&lt;/strong&gt;: &lt;a href="https://github.com/tilakraj1108/friend-study-buddy" rel="noopener noreferrer"&gt;https://github.com/tilakraj1108/friend-study-buddy&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;friend-study-buddy/
├── index.html   # Semantic HTML5 layout and structured forms
├── style.css    # Responsive design system, CSS variables &amp;amp; dark mode
├── script.js    # Vanilla JS planner logic, quiz engine, &amp;amp; quote generator
└── README.md    # Documentation and local setup guide
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;I designed and developed &lt;strong&gt;Friend Study Buddy&lt;/strong&gt; using open-source agentic workflows and standard modern web primitives:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;AI-Assisted Pair Programming&lt;/strong&gt;: Leveraged open-source AI tooling and agentic coding workflows to rapidly scaffold semantic HTML5 structures, formulate balanced study breakdown algorithms, and create responsive UI layouts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pure Web Stack Architecture&lt;/strong&gt;:

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;HTML5&lt;/strong&gt;: Fully accessible and semantic tags (&lt;code&gt;&amp;lt;header&amp;gt;&lt;/code&gt;, &lt;code&gt;&amp;lt;nav&amp;gt;&lt;/code&gt;, &lt;code&gt;&amp;lt;main&amp;gt;&lt;/code&gt;, &lt;code&gt;&amp;lt;section&amp;gt;&lt;/code&gt;, &lt;code&gt;&amp;lt;form&amp;gt;&lt;/code&gt;, &lt;code&gt;&amp;lt;footer&amp;gt;&lt;/code&gt;, &lt;code&gt;&amp;lt;fieldset&amp;gt;&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CSS3 Design System&lt;/strong&gt;: Custom properties (CSS variables) for seamless Light/Dark theme switching, responsive CSS Grid and Flexbox layouts, and smooth transitions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vanilla JavaScript (ES6+)&lt;/strong&gt;: Clean, dependency-free DOM manipulation, event delegation, and modular state management for the planner calculations and quiz evaluator.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Open innovation and open-source software are the bedrock of modern technology education. When building for a friend, open innovation made it possible to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Zero Barriers to Access&lt;/strong&gt;: Anyone can inspect, modify, fork, or run the project locally without paying subscription fees, managing API keys, or dealing with vendor lock-in.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Longevity and Reliability&lt;/strong&gt;: Because it relies purely on open web standards, this tool will work today, tomorrow, and decades from now without breaking due to closed API deprecations or pricing changes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Collaborative Learning&lt;/strong&gt;: Friends can contribute new quiz questions, add custom subject notes, or customize study algorithms collaboratively via pull requests.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;I created and iterated on this application using an AI pair-programming agent session to draft, refine, and verify the study planner algorithms, accessibility features, and documentation.&lt;/p&gt;




&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/strong&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Building Bazaar Mitra: My 10-Day Journey Building a Voice AI Agent for Local Commerce</title>
      <dc:creator>Tilak Raj Roy</dc:creator>
      <pubDate>Sat, 15 Aug 2026 06:37:43 +0000</pubDate>
      <link>https://dev.to/tilak_rajroy_f81d1723578/building-bazaar-mitra-my-10-day-journey-building-a-voice-ai-agent-for-local-commerce-eo0</link>
      <guid>https://dev.to/tilak_rajroy_f81d1723578/building-bazaar-mitra-my-10-day-journey-building-a-voice-ai-agent-for-local-commerce-eo0</guid>
      <description>&lt;h2&gt;
  
  
  1.INTRODUCTION
&lt;/h2&gt;

&lt;p&gt;Bazaar Mitra is a voice-first AI assistant designed to help local-commerce customers with product enquiries, memory, outbound calls, human escalation, analytics, and specialist handoffs.&lt;/p&gt;

&lt;h2&gt;
  
  
  2.THE PROBLEMS
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Persistent user memory
Making the agent remember a caller's name across different calls.
Solved using SQLite with user_id, name, language preference, and facts.
Also added explicit consent before saving information
2.Product data integration
Initially the product lookup used a manually created local catalogue.
You then moved toward real external catalogue data using DummyJSON API.
The challenge was handling API failures, timeouts, missing products, and preventing the agent from guessing prices.
3.Outbound calling
Browser-based calls were straightforward, but outbound SIP calling required configuring LiveKit SIP and Linphone.
You had to deal with SIP URI, trunk ID, environment variables, and room/agent connection.
4.Frontend/debugging
You encountered frontend/runtime issues such as missing references and component-related errors while customizing the LiveKit starter.
This is a good section because it shows actual debugging rather than pretending everything worked immediately.
5.Human escalation
The agent needed to know when not to solve something itself.
You implemented escalation for situations requiring human assistance, including permission before sharing information.
6.Call analytics
You needed actual call results rather than hardcoded dashboard numbers.
The dashboard tracks:
Total calls
Successful calls
Failed calls
7.Specialist handoff
The main Bazaar Mitra agent shouldn't handle every type of request.
You added a specialist-agent concept so specific requests can be transferred while maintaining conversation context.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  3.FEATURE MAP
&lt;/h2&gt;

&lt;p&gt;Bazaar Mitra is a Local Commerce AI voice assistant that can:&lt;/p&gt;

&lt;p&gt;converse using voice&lt;br&gt;
follow business and safety guardrails&lt;br&gt;
understand Hindi, English, and Hinglish&lt;br&gt;
remember returning callers with consent&lt;br&gt;
look up product information&lt;br&gt;
make outbound calls&lt;br&gt;
escalate payment/order issues to humans&lt;br&gt;
track call outcomes through analytics&lt;br&gt;
hand return/refund conversations to a specialist agent&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Voice + Guardrails - Defined identity, objectives, safety boundaries and concise voice responses.&lt;/li&gt;
&lt;li&gt;Indian / Mixed Language - Hindi, English and Hinglish communication based on the caller.&lt;/li&gt;
&lt;li&gt;Memory - Caller information is stored persistently only after explicit consent.&lt;/li&gt;
&lt;li&gt;Product Lookup - The agent calls a tool/API instead of inventing product information.&lt;/li&gt;
&lt;li&gt;Outbound Calls - The agent can initiate an outbound voice interaction through SIP.&lt;/li&gt;
&lt;li&gt;Human Escalation - Payment/refund or order-dispute issues can become a human-help request.&lt;/li&gt;
&lt;li&gt;Call Analytics - Real call outcomes feed Total, Successful and Failed call counts.&lt;/li&gt;
&lt;li&gt;Specialist Handoff - Returns/refunds conversations move from the general agent to a focused specialist.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  4.ARCHITECTURE
&lt;/h2&gt;

&lt;p&gt;User&lt;br&gt;
  ↓&lt;br&gt;
LiveKit&lt;br&gt;
  ↓&lt;br&gt;
Deepgram STT&lt;br&gt;
  ↓&lt;br&gt;
Gemini LLM&lt;br&gt;
  ↓&lt;br&gt;
Tools / Memory / Handoff&lt;br&gt;
  ↓&lt;br&gt;
Murf Falcon TTS&lt;br&gt;
  ↓&lt;br&gt;
User&lt;/p&gt;

&lt;p&gt;Supporting Components :&lt;br&gt;
SQLite → persistent customer memory&lt;br&gt;
Product API → product lookup&lt;br&gt;
SIP/Linphone → outbound calls&lt;br&gt;
Escalation DB → human-help requests&lt;br&gt;
Analytics DB → call outcomes&lt;br&gt;
Specialist Agent → returns/refunds&lt;/p&gt;

&lt;h2&gt;
  
  
  5.DIFFICULT PARTS
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Memory persistence problem:&lt;/strong&gt;&lt;br&gt;
Initially the agent could save information but couldn't remember it across calls. The issue was traced through the caller identity and database flow. Stable user IDs and SQLite persistence were required.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Tool integration problems:&lt;/strong&gt;&lt;br&gt;
While adding Day 5 tools, tool definitions and API handling caused failures. Debugging logs helped identify whether the tool was being called, whether the API responded, and whether the result reached the voice pipeline.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Outbound SIP setup:&lt;/strong&gt;&lt;br&gt;
Outbound calling required understanding the difference between the LiveKit SIP endpoint, SIP trunk, and destination SIP address. Environment-variable configuration was also important.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. PRACTICAL BUILD GUIDE
&lt;/h2&gt;

&lt;p&gt;Prerequisites: Python environment, LiveKit project, API credentials for the services used, and a public repository.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Basic Stack :&lt;/em&gt;&lt;br&gt;
Speech-to-Text → Deepgram&lt;br&gt;
LLM → Gemini&lt;br&gt;
Text-to-Speech → Murf Falcon&lt;br&gt;
Real-time transport → LiveKit&lt;br&gt;
Database → SQLite&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Setup Sequence :&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Clone the repository&lt;/li&gt;
&lt;li&gt;Install dependencies&lt;/li&gt;
&lt;li&gt;Create .env.local&lt;/li&gt;
&lt;li&gt;Add API keys there&lt;/li&gt;
&lt;li&gt;Never commit .env.local&lt;/li&gt;
&lt;li&gt;Start the LiveKit agent&lt;/li&gt;
&lt;li&gt;Start the frontend&lt;/li&gt;
&lt;li&gt;Open the browser&lt;/li&gt;
&lt;li&gt;Start a voice session&lt;/li&gt;
&lt;li&gt;Test a simple conversation&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  7. Evidence
&lt;/h2&gt;

&lt;p&gt;Bazaar Mitra frontend&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz63jvyfok2ng4693ct1a.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz63jvyfok2ng4693ct1a.png" alt=" " width="800" height="389"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;product tool being called&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdfar9kbf9o0hsmznz06l.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdfar9kbf9o0hsmznz06l.png" alt=" " width="800" height="386"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Call analytics dashboard&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F46zi1qjfs15at51ytzno.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F46zi1qjfs15at51ytzno.png" alt=" " width="800" height="399"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Specialist handoff clip&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fhex13wc7pi1dpcaxs7a1.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fhex13wc7pi1dpcaxs7a1.png" alt=" " width="800" height="384"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  8. LEARNING
&lt;/h2&gt;

&lt;p&gt;Voice agents are not just an LLM plus TTS. Reliability depends on turn detection, STT quality, tool design,&lt;br&gt;
state management, database persistence, permissions, fallback behavior, and clear agent boundaries.&lt;/p&gt;

&lt;p&gt;the next version could include:&lt;/p&gt;

&lt;p&gt;real merchant inventory instead of demo/catalogue data&lt;br&gt;
proper order-management integration&lt;br&gt;
better multilingual voice support&lt;br&gt;
richer analytics&lt;br&gt;
real CRM/help-desk integration&lt;br&gt;
improved specialist routing&lt;br&gt;
authentication and production-grade data protection.&lt;/p&gt;

&lt;h2&gt;
  
  
  9. LINKS
&lt;/h2&gt;

&lt;p&gt;GitHub Repository : &lt;a href="https://github.com/insyncerick/murf-livekit-starter" rel="noopener noreferrer"&gt;https://github.com/insyncerick/murf-livekit-starter&lt;/a&gt;&lt;br&gt;
My Linkedin Profile : &lt;a href="http://www.linkedin.com/in/tilak-raj-roy-83b85440a" rel="noopener noreferrer"&gt;www.linkedin.com/in/tilak-raj-roy-83b85440a&lt;/a&gt;&lt;br&gt;
(there is uploaded my day1 to day 10 full journey posts )&lt;/p&gt;

</description>
      <category>voiceforbharat</category>
      <category>voiceai</category>
      <category>murffalcon</category>
      <category>livekit</category>
    </item>
  </channel>
</rss>
