<?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: Sujay H S</title>
    <description>The latest articles on DEV Community by Sujay H S (@sujay_hs_b482e14368c435a).</description>
    <link>https://dev.to/sujay_hs_b482e14368c435a</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%2F4078654%2F36505fb9-bb13-446d-8395-e67c740feddb.png</url>
      <title>DEV Community: Sujay H S</title>
      <link>https://dev.to/sujay_hs_b482e14368c435a</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/sujay_hs_b482e14368c435a"/>
    <language>en</language>
    <item>
      <title>How We Built Kisan Saathi in 24 Hours: Surviving 3 AM Bugs, Voice AI in Kannada, and Building for Real Farmers 🌾</title>
      <dc:creator>Sujay H S</dc:creator>
      <pubDate>Sun, 15 Nov 2026 16:54:46 +0000</pubDate>
      <link>https://dev.to/sujay_hs_b482e14368c435a/how-we-built-kisan-saathi-in-24-hours-surviving-3-am-bugs-voice-ai-in-kannada-and-building-for-590e</link>
      <guid>https://dev.to/sujay_hs_b482e14368c435a/how-we-built-kisan-saathi-in-24-hours-surviving-3-am-bugs-voice-ai-in-kannada-and-building-for-590e</guid>
      <description>&lt;h1&gt;
  
  
  We Built Kisan Scheme Saathi in 24 Hours: Our Hackathon Journey, 3 AM Bugs, and Lessons from the Judges
&lt;/h1&gt;

&lt;p&gt;In rural Karnataka, accessing government welfare schemes can be difficult for the farmers who need them most.&lt;/p&gt;

&lt;p&gt;Although several government initiatives provide agricultural subsidies, irrigation support, crop insurance, and educational scholarships, many smallholder farmers struggle to access these benefits because of:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Complex government documents&lt;/strong&gt; written in difficult administrative language.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Scattered government portals&lt;/strong&gt;, including Bhoomi, FRUITS, PM-KISAN, and Samrakshane.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dependence on intermediaries&lt;/strong&gt;, sometimes involving unnecessary charges.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Language and digital literacy barriers&lt;/strong&gt; on conventional government websites.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;During a challenging &lt;strong&gt;24-hour hackathon&lt;/strong&gt;, our team set out to address this problem by building &lt;strong&gt;Kisan Scheme Saathi (ಕಿಸಾನ್ ಸ್ಕೀಮ್ ಸಾಥಿ)&lt;/strong&gt; — an open-source, voice-first web application designed to help farmers discover relevant government schemes, understand eligibility requirements, view document checklists, and track important deadlines.&lt;/p&gt;

&lt;p&gt;🌐 &lt;strong&gt;Live Website:&lt;/strong&gt; &lt;a href="https://kisansathi-frontend-tau.vercel.app" rel="noopener noreferrer"&gt;Kisan Scheme Saathi&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;📂 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/Sujayhs07/Kisan_sathi_" rel="noopener noreferrer"&gt;Kisan Scheme Saathi&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;📜 &lt;strong&gt;License:&lt;/strong&gt; MIT Open Source License&lt;/p&gt;

&lt;p&gt;We didn't win the award. But we gave the hackathon everything we had. We built a working product, dealt with unexpected technical challenges, fixed late-night bugs, disagreed over architectural decisions, and faced questions from judges that made us rethink parts of our solution.&lt;/p&gt;

&lt;p&gt;This is the story of those 24 hours — the decisions, challenges, mistakes, and lessons we took away from the experience.&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%2F8s7jeov0rgqtss3gb9tn.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%2F8s7jeov0rgqtss3gb9tn.png" alt=" " width="800" height="430"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The Kisan Scheme Saathi landing page, designed to make the platform accessible through a simple, bilingual interface.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  🎯 Why We Started with Karnataka
&lt;/h2&gt;

&lt;p&gt;When the 24-hour countdown began, our initial ambition was huge:&lt;/p&gt;

&lt;p&gt;&lt;em&gt;"Let's build a unified government scheme matcher for all 28 states of India in 12 languages!"&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;But within the first two hours, we realised that this approach was too ambitious for the time available.&lt;/p&gt;

&lt;p&gt;Agricultural schemes and eligibility requirements vary significantly between states. Karnataka also has its own digital infrastructure and welfare programmes, including:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Bhoomi RTC / Pahani:&lt;/strong&gt; Land record and ownership information.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;FRUITS FID:&lt;/strong&gt; Farmer Registration and Unified Beneficiary Information System.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Samrakshane Portal:&lt;/strong&gt; An important platform for crop insurance-related services.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;State-specific welfare schemes:&lt;/strong&gt; Programmes such as Ganga Kalyana, Raitha Siri, and CM Raitha Vidyanidhi.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We had to make a practical decision: instead of building a broad but shallow application, we would focus on Karnataka and make the experience more meaningful for its farmers.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌─────────────────────────────────────────────────────────────┐
│                    THE HACKATHON REALITY                    │
│                                                             │
│   Build for 28 States without enough depth   ❌             │
│   Focus on Karnataka with greater detail     ✅             │
│                                                             │
│   • Schemes: 8 Central and Karnataka programs               │
│   • Languages: Kannada (ಕನ್ನಡ) and Indian English           │
│   • Architecture: Designed for future expansion             │
└─────────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Because of the strict timeline, we focused on &lt;strong&gt;Karnataka farmers&lt;/strong&gt;, supporting &lt;strong&gt;Kannada (&lt;code&gt;kn-IN&lt;/code&gt;)&lt;/strong&gt; and &lt;strong&gt;Indian English&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;However, we designed the underlying rules engine and JSON data structure with future expansion in mind. The architecture can be extended to other states by adding their respective schemes, eligibility rules, and regional requirements.&lt;/p&gt;




&lt;h2&gt;
  
  
  🥊 The War Room: A Debate That Shaped Our Architecture
&lt;/h2&gt;

&lt;p&gt;Hackathons are not just about writing code. They also test how a team makes decisions under pressure.&lt;/p&gt;

&lt;p&gt;Around the third hour, we had a major disagreement about how the eligibility system should work.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Camp A — The AI-first approach:&lt;/strong&gt; &lt;em&gt;"Why write complex eligibility logic? Let's send the farmer's details to an LLM and ask it to identify the eligible schemes. We can build that quickly!"&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Camp B — The reliability-first approach:&lt;/strong&gt; &lt;em&gt;"What if the AI generates an incorrect eligibility rule? A farmer might spend time and money visiting an office based on that information. We cannot take that risk."&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The discussion helped us arrive at an important architectural decision:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic eligibility evaluation:&lt;/strong&gt; Scheme qualification would be handled through predefined rules and application logic, using details such as landholding, irrigation type, category, and crop information.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI-assisted explanations:&lt;/strong&gt; Google Gemini would support conversational explanations, helping users understand why a scheme might or might not apply to them.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;In short, we decided that AI could help explain the results, but the core eligibility decision should come from explicit rules rather than an LLM-generated answer.&lt;/p&gt;

&lt;p&gt;Once we agreed on this approach, we divided the responsibilities and continued building.&lt;/p&gt;




&lt;h2&gt;
  
  
  🐛 The Battles in the Trenches: 3 AM Bugs and Engineering Challenges
&lt;/h2&gt;

&lt;p&gt;Building a working application in 24 hours meant dealing with problems we had not initially anticipated. Some of our biggest challenges appeared during testing.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. The 3:15 AM Kannada Voice Synthesis Crisis
&lt;/h3&gt;

&lt;p&gt;We wanted Kisan Scheme Saathi to be accessible to farmers who might find reading long instructions on a screen difficult.&lt;/p&gt;

&lt;p&gt;The application therefore needed to support both voice input and audio responses.&lt;/p&gt;

&lt;p&gt;Initially, we tried the browser's native Web Speech API. It worked during our English tests, but Kannada (&lt;code&gt;kn-IN&lt;/code&gt;) presented problems.&lt;/p&gt;

&lt;p&gt;On our testing setup, the browser did not consistently provide a suitable Kannada voice. Some attempts failed, while others produced unnatural pronunciation.&lt;/p&gt;

&lt;p&gt;We needed a more reliable solution.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Integrating Murf AI for Kannada voice output&lt;/span&gt;
&lt;span class="c1"&gt;// Configured Kannada voice options for the application&lt;/span&gt;

&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;getMurfDefaultConfig&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;voiceId&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'Harman'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;       &lt;span class="c1"&gt;// Male Kannada voice&lt;/span&gt;
    &lt;span class="n"&gt;femaleVoiceId&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'Pooja'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;  &lt;span class="c1"&gt;// Female Kannada voice&lt;/span&gt;
    &lt;span class="n"&gt;locale&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'kn-IN'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'Falcon 2.1'&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;We integrated &lt;strong&gt;Murf AI Falcon 2.1&lt;/strong&gt; for Kannada text-to-speech, using the configured voice options to generate audio responses.&lt;/p&gt;

&lt;p&gt;Alongside Murf AI, we integrated &lt;strong&gt;Deepgram Nova-2&lt;/strong&gt; for speech-to-text. This allowed farmers to speak details such as their landholding and crop information through their phone's microphone.&lt;/p&gt;

&lt;p&gt;We also implemented a custom phonetic transliteration fallback, &lt;code&gt;transliterateKannada&lt;/code&gt;, to help handle regional agricultural terms such as &lt;em&gt;Pahani&lt;/em&gt;, &lt;em&gt;Bhoomi&lt;/em&gt;, and &lt;em&gt;Raitha&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;This was an important step towards making the application more accessible to users who prefer speaking rather than typing.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. The Village Connectivity Reality: Building an Offline Caching Database
&lt;/h3&gt;

&lt;p&gt;Many web applications are developed and tested using reliable internet connections. But connectivity in rural areas can be inconsistent.&lt;/p&gt;

&lt;p&gt;We wanted Kisan Scheme Saathi to remain useful even when a farmer experienced slow network speeds or temporary connection failures.&lt;/p&gt;

&lt;p&gt;Imagine a farmer clicking &lt;em&gt;"Check My Eligibility"&lt;/em&gt; and waiting several seconds for a remote database request. If the application appears unresponsive, the farmer may simply leave.&lt;/p&gt;

&lt;p&gt;To address this, we worked on an &lt;strong&gt;offline-first caching database architecture&lt;/strong&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌────────────────────────────────────────────────────────┐
│             VILLAGE CONNECTIVITY REALITY               │
│                                                        │
│   City: 5G / Fiber Optic  ──&amp;gt; Better connectivity      │
│   Village: 2G / Edge      ──&amp;gt; Unstable connectivity    │
└────────────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Our implementation included:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Caching scheme rules, document requirements, and district mappings to reduce repeated server requests.&lt;/li&gt;
&lt;li&gt;A dual-mode persistence mechanism using &lt;code&gt;pgPool.ts&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;A local JSON fallback at &lt;code&gt;data/app-db.json&lt;/code&gt; when the PostgreSQL database was unavailable.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Here is an example of the backend test output during a run where the local JSON fallback was activated:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; kisan-saathi-api-server@2.0.0 &lt;span class="nb"&gt;test&lt;/span&gt;
&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; npx tsx tests/integration.test.ts

&lt;span class="o"&gt;[&lt;/span&gt;TEST SUITE&lt;span class="o"&gt;]&lt;/span&gt; Importing and starting Kisan Scheme Saathi server...
&lt;span class="o"&gt;[&lt;/span&gt;PostgreSQL&lt;span class="o"&gt;]&lt;/span&gt; Database not reachable. Using &lt;span class="kd"&gt;local&lt;/span&gt; JSON fallback &lt;span class="o"&gt;(&lt;/span&gt;data/app-db.json&lt;span class="o"&gt;)&lt;/span&gt;.
Kisan Scheme Saathi backend running on http://localhost:5000

&lt;span class="o"&gt;======================================================&lt;/span&gt;
   RUNNING COMPREHENSIVE BACKEND API TEST HARNESS
&lt;span class="o"&gt;======================================================&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Health Check &lt;span class="o"&gt;(&lt;/span&gt;/api/health&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Schemes Catalog &lt;span class="o"&gt;(&lt;/span&gt;/api/schemes&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Scheme Details &lt;span class="o"&gt;(&lt;/span&gt;/api/schemes/pm-kisan&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Document Checklist &lt;span class="o"&gt;(&lt;/span&gt;/api/schemes/pm-kisan/checklist&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Application Guide &lt;span class="o"&gt;(&lt;/span&gt;/api/schemes/pm-kisan/application-guide&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Eligibility Engine &lt;span class="o"&gt;(&lt;/span&gt;Marginal SC Millet Farmer&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Eligibility Engine Output Distinction &lt;span class="o"&gt;(&lt;/span&gt;Large General Sugarcane Farmer&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Conversational Chatbot Endpoint &lt;span class="o"&gt;(&lt;/span&gt;/api/eligibility-chat&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Chatbot Input Validation: Reject &lt;span class="s2"&gt;"I don't know"&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; Name
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Chatbot Input Validation: Reject Invalid Age
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Chatbot Confirmation: Aadhaar advances to Confirm Step
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Chatbot Confirmation: No/Change routes to Edit Options
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Chatbot Confirmation: Yes Evaluates Schemes
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Farmer Registration &lt;span class="o"&gt;&amp;amp;&lt;/span&gt; OTP &lt;span class="o"&gt;(&lt;/span&gt;/api/auth/register&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Verify OTP &lt;span class="o"&gt;&amp;amp;&lt;/span&gt; Issue JWT &lt;span class="o"&gt;(&lt;/span&gt;/api/auth/verify-otp&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Unregistered Mobile Login Block &lt;span class="o"&gt;(&lt;/span&gt;/api/auth/send-otp LOGIN_VERIFICATION&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Unregistered Mobile Verify OTP Block &lt;span class="o"&gt;(&lt;/span&gt;/api/auth/verify-otp LOGIN_VERIFICATION&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Registered Mobile Login Verification &lt;span class="o"&gt;(&lt;/span&gt;/api/auth/send-otp &lt;span class="o"&gt;&amp;amp;&lt;/span&gt; verify-otp&lt;span class="o"&gt;)&lt;/span&gt;
  ✓ &lt;span class="o"&gt;[&lt;/span&gt;PASS&lt;span class="o"&gt;]&lt;/span&gt; Check User Database Endpoint &lt;span class="o"&gt;(&lt;/span&gt;/api/auth/check-user&lt;span class="o"&gt;)&lt;/span&gt;

&lt;span class="o"&gt;======================================================&lt;/span&gt;
TOTAL SUITE RESULTS: 19/19 PASSED &lt;span class="o"&gt;(&lt;/span&gt;100%&lt;span class="o"&gt;)&lt;/span&gt;
&lt;span class="o"&gt;======================================================&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The test run reported &lt;strong&gt;19 out of 19 checks passing&lt;/strong&gt;, including the database fallback, eligibility engine, chatbot, and authentication flows.&lt;/p&gt;

&lt;p&gt;Although we implemented a local JSON fallback when PostgreSQL was unavailable and worked on caching frequently accessed scheme information, this did not mean that the entire application was fully functional offline. The fallback improved backend persistence resilience, while caching helped reduce repeated requests. Full offline functionality remains an area for future improvement.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. The Big Bug: Accepting Invalid User Input
&lt;/h3&gt;

&lt;p&gt;Around 5:30 AM, while testing different user journeys, we discovered an important validation issue.&lt;/p&gt;

&lt;p&gt;The application was accepting input without sufficiently checking whether the values were valid.&lt;/p&gt;

&lt;p&gt;For example, a user could enter:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Landholding: &lt;code&gt;-50 acres&lt;/code&gt; or random letters.&lt;/li&gt;
&lt;li&gt;Phone number: &lt;code&gt;123&lt;/code&gt; or &lt;code&gt;99999999999999&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Social category: left unselected.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This could lead to incorrect eligibility results. A user entering a negative landholding value, for example, could be processed by the rules engine without the expected validation.&lt;/p&gt;

&lt;p&gt;We paused other development work and spent considerable time strengthening the validation layers.&lt;/p&gt;

&lt;p&gt;The changes included:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Validating phone numbers to accept &lt;strong&gt;10-digit Indian mobile numbers&lt;/strong&gt;, with checks across &lt;code&gt;onKeyDown&lt;/code&gt;, &lt;code&gt;onPaste&lt;/code&gt;, &lt;code&gt;onInput&lt;/code&gt;, and regex validation.&lt;/li&gt;
&lt;li&gt;Restricting landholding inputs to positive numeric values within the application's configured range.&lt;/li&gt;
&lt;li&gt;Adding step-by-step confirmation in the conversational chatbot so that user details are checked before the eligibility evaluation begins.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This experience reminded us that input validation is not just a technical detail. When an application deals with government benefits, incorrect input can lead to misleading results.&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%2Ffjvd80dflf5nzj2wgvbf.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%2Ffjvd80dflf5nzj2wgvbf.png" alt=" " width="799" height="432"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;The matched schemes dashboard showing scheme information, Kannada audio controls, deadline indicators, and relevant application links.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  🔔 Smart Scheme Reminders: Helping Farmers Never Miss a Deadline
&lt;/h2&gt;

&lt;p&gt;Finding an eligible government scheme is only the first step. Farmers also need to know when applications open and when deadlines are approaching.&lt;/p&gt;

&lt;p&gt;Some schemes have specific application windows and seasonal deadlines. Missing these dates could mean losing an opportunity to benefit from a scheme.&lt;/p&gt;

&lt;p&gt;To address this challenge, we designed the &lt;strong&gt;My Reminders&lt;/strong&gt; feature in Kisan Scheme Saathi.&lt;/p&gt;

&lt;p&gt;Farmers can explore eligible and upcoming schemes and use the &lt;strong&gt;Set Reminder&lt;/strong&gt; option to keep track of important application dates.&lt;/p&gt;

&lt;h3&gt;
  
  
  Our Vision for SMS Notifications
&lt;/h3&gt;

&lt;p&gt;As part of our proposed reminder functionality, we plan to introduce SMS notifications that can:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Notify farmers when an upcoming scheme application window opens.&lt;/li&gt;
&lt;li&gt;Remind farmers before an application deadline.&lt;/li&gt;
&lt;li&gt;Provide timely alerts so they can prepare the required documents and apply.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  Why This Feature Matters
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Avoid Missed Deadlines:&lt;/strong&gt; Help farmers keep track of important application dates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stay Informed:&lt;/strong&gt; Make upcoming scheme opportunities easier to follow.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Timely Action:&lt;/strong&gt; Remind farmers to prepare and submit applications before deadlines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Accessibility:&lt;/strong&gt; SMS notifications can reach farmers directly through their mobile phones without requiring them to repeatedly visit the website.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The SMS notification functionality is part of our planned enhancement and was not active in the hackathon implementation.&lt;/p&gt;

&lt;p&gt;Our goal is simple: &lt;strong&gt;Helping farmers discover the right schemes is important, but helping them act at the right time is equally important.&lt;/strong&gt;&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%2Fkgl44qb304hc89p5rila.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%2Fkgl44qb304hc89p5rila.png" alt=" " width="799" height="432"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  &lt;em&gt;The My Reminders interface, designed to help farmers track upcoming scheme application windows and important deadlines.&lt;/em&gt;
&lt;/h2&gt;

&lt;h2&gt;
  
  
  📱 Mobile-First UX: Designed Around How Farmers Access Technology
&lt;/h2&gt;

&lt;p&gt;One important question we considered while designing Kisan Scheme Saathi was: &lt;strong&gt;How will farmers actually access this platform?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;For many farmers, a smartphone is one of the most accessible ways to connect to digital services. Rather than expecting users to visit a computer centre or navigate a complicated desktop website, we wanted them to access scheme information directly from a device they already use in their daily lives.&lt;/p&gt;

&lt;p&gt;That is why we prioritised a &lt;strong&gt;mobile-first, responsive user experience&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Our design focuses on:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Responsive Layout:&lt;/strong&gt; The interface adapts to different smartphone screen sizes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Simple Navigation:&lt;/strong&gt; Important features are easy to find, with fewer unnecessary steps.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Readable Scheme Information:&lt;/strong&gt; Eligibility details, benefits, deadlines, and document requirements are presented clearly.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kannada Voice Interaction:&lt;/strong&gt; Users can provide information through voice input and listen to responses in Kannada.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Touch-Friendly Controls:&lt;/strong&gt; Buttons, forms, and interactive elements are designed for convenient mobile usage.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Easy-to-Understand Scheme Cards:&lt;/strong&gt; Matched schemes are displayed in a structured format, helping users understand their options and next steps.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Our intention was not simply to shrink the desktop website to fit a mobile screen. We wanted to create an experience that feels natural and practical for farmers using smartphones.&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%2F691vdb8d9q52gaxlo80g.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%2F691vdb8d9q52gaxlo80g.png" alt=" " width="800" height="525"&gt;&lt;/a&gt;&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%2Fvsdy8n7doulv2yd4tm2n.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%2Fvsdy8n7doulv2yd4tm2n.png" alt=" " width="800" height="525"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Mobile interface of Kisan Scheme Saathi, demonstrating its responsive layout and farmer-friendly design.&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  🔐 Protecting Farmer Information: OTP-Based Authentication
&lt;/h2&gt;

&lt;p&gt;While building a platform that handles personal information, we understood that accessibility alone was not enough. &lt;strong&gt;The security and privacy of farmers' data also had to be considered.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Farmers may provide personal details while registering and checking their eligibility for government schemes. We therefore implemented an &lt;strong&gt;OTP-based authentication system&lt;/strong&gt; to help protect access to user accounts.&lt;/p&gt;

&lt;p&gt;Our authentication flow includes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Mobile Number Registration:&lt;/strong&gt; Users register using their mobile number.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;OTP Verification:&lt;/strong&gt; A one-time password is sent to the registered mobile number.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Account Verification:&lt;/strong&gt; Users must enter the correct OTP before their account can be verified.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;JWT-Based Authentication:&lt;/strong&gt; After successful verification, the backend issues a JSON Web Token (JWT) for authenticated access.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Login Verification:&lt;/strong&gt; Login verification checks help prevent unregistered mobile numbers from accessing protected account functionality.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;This approach provides an additional layer of account security while keeping the login process familiar and straightforward for mobile users.&lt;/p&gt;

&lt;p&gt;We wanted the authentication experience to remain simple without overlooking the importance of protecting personal information.&lt;/p&gt;

&lt;h2&gt;
  
  
  &lt;strong&gt;Our design principle was simple: make the platform easy to access, but ensure that access to personal accounts is verified.&lt;/strong&gt;
&lt;/h2&gt;

&lt;h2&gt;
  
  
  🏛️ The Three Rounds of Judging: Three Different Experiences
&lt;/h2&gt;

&lt;p&gt;Hackathon judging was one of the most memorable parts of our experience.&lt;/p&gt;

&lt;p&gt;We went through &lt;strong&gt;three evaluation rounds&lt;/strong&gt;, each with a different judge. Every round brought a different experience, and some of the questions challenged us more than we expected.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;┌───────────────┐     ┌───────────────┐     ┌───────────────┐
│    ROUND 1    │     │    ROUND 2    │     │    ROUND 3    │
│   (Judge A)   │     │   (Judge B)   │     │   (Judge C)   │
├───────────────┤     ├───────────────┤     ├───────────────┤
│ The Slang &amp;amp;   │     │ The Silent    │     │ Total         │
│ Paan Question │     │ Stone Wall    │     │ Redemption    │
│  (Went Blank) │     │ (No Feedback) │     │ (Impressed!)  │
└───────────────┘     └───────────────┘     └───────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Round 1: The Slang and Paan Curveball
&lt;/h3&gt;

&lt;p&gt;The first judge arrived at our table. We presented our idea, demonstrated the Kannada voice features, and explained the deterministic scheme-matching system.&lt;/p&gt;

&lt;p&gt;Then came a question we were not prepared for:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"This voice recognition is nice. But what about the regional slangs? The Kannada spoken in Hubballi-Dharwad is totally different from Mandya, Kalaburagi, or Mangaluru Kannada. And what if a farmer is chewing paan (gutka/tobacco) while speaking into the microphone? How will your AI understand muffled speech?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;We went completely blank.&lt;/p&gt;

&lt;p&gt;Regional dialects? Muffled speech? We had integrated Deepgram and Murf, but we had not trained a dedicated speech model for these real-world conditions.&lt;/p&gt;

&lt;p&gt;For a few moments, we struggled to come up with a convincing technical answer.&lt;/p&gt;

&lt;p&gt;After the round, we thought more carefully about the question. We realised that supporting Kannada is not simply about selecting a language option in an API. Real-world speech includes regional accents, local vocabulary, background noise, and differences in pronunciation.&lt;/p&gt;

&lt;p&gt;One possible direction is to conduct &lt;strong&gt;on-ground surveys with college students and local volunteers&lt;/strong&gt; across different taluks. The collected speech samples could help us understand regional variations and explore the development of more suitable speech models.&lt;/p&gt;

&lt;p&gt;It was a difficult question, but it exposed an important limitation in our original approach.&lt;/p&gt;

&lt;h3&gt;
  
  
  Round 2: The Silent Stone Wall
&lt;/h3&gt;

&lt;p&gt;The second judge arrived, and we presented our application again.&lt;/p&gt;

&lt;p&gt;We explained the offline caching approach, demonstrated the four-step eligibility wizard, and showed the document checklist for Grama One centres.&lt;/p&gt;

&lt;p&gt;She listened quietly, nodded a few times, wrote notes on her evaluation sheet, and left without asking any questions.&lt;/p&gt;

&lt;p&gt;We were confused.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Did she like the project? Was something missing? Did we fail to explain an important feature?&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;We still don't know what she thought about the presentation.&lt;/p&gt;

&lt;p&gt;In a hackathon, direct feedback gives you something to work with. Silence can be much harder to interpret.&lt;/p&gt;

&lt;p&gt;We felt that our difficulty answering the first judge's question, combined with the lack of feedback in the second round, had affected our confidence going into the final evaluation.&lt;/p&gt;

&lt;h3&gt;
  
  
  Round 3: Total Redemption
&lt;/h3&gt;

&lt;p&gt;By the third round, we had decided to approach the presentation differently.&lt;/p&gt;

&lt;p&gt;Instead of waiting for questions, we began by explaining the practical challenges a platform like ours would face in rural areas.&lt;/p&gt;

&lt;p&gt;We addressed:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Regional Kannada dialect variations.&lt;/li&gt;
&lt;li&gt;Background noise and unclear speech.&lt;/li&gt;
&lt;li&gt;Unstable internet connectivity.&lt;/li&gt;
&lt;li&gt;Incomplete documentation at service centres.&lt;/li&gt;
&lt;li&gt;The importance of reliable eligibility calculations.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We explained how our caching approach addressed connectivity challenges, why the eligibility engine relied on deterministic rules, and how future field surveys could help us improve regional speech support.&lt;/p&gt;

&lt;p&gt;Then we demonstrated the working application:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Spoke into the microphone in Kannada.&lt;/li&gt;
&lt;li&gt;Generated matching schemes, including PMKSY micro-irrigation, Ganga Kalyana, and Raitha Vidyanidhi.&lt;/li&gt;
&lt;li&gt;Played the Kannada audio responses.&lt;/li&gt;
&lt;li&gt;Demonstrated deadline countdowns and document checklists.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This time, the presentation went much better.&lt;/p&gt;

&lt;p&gt;The judge engaged with our architecture, asked questions, and appreciated the effort we had put into understanding the practical challenges faced by farmers.&lt;/p&gt;

&lt;p&gt;We walked out of the third round feeling much more confident.&lt;/p&gt;




&lt;h2&gt;
  
  
  🏆 We Didn't Win the Award, But We Gave It Our Best
&lt;/h2&gt;

&lt;p&gt;When the final awards were announced, our name wasn't called.&lt;/p&gt;

&lt;p&gt;Of course, we were disappointed.&lt;/p&gt;

&lt;p&gt;We had spent 24 hours working with very little sleep, dealt with unexpected bugs, made difficult technical decisions, and put everything we had into the project. Naturally, we wanted to win.&lt;/p&gt;

&lt;p&gt;But as we packed our bags and looked at the working Kisan Saathi application running on our phones, we realised that the experience had given us something valuable.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;We had built an application focused on a real problem faced by farming families.&lt;/li&gt;
&lt;li&gt;We had learned to think about low-connectivity environments and caching.&lt;/li&gt;
&lt;li&gt;We had understood why deterministic logic matters in applications involving eligibility decisions.&lt;/li&gt;
&lt;li&gt;We had experienced the importance of input validation, testing, and clear communication.&lt;/li&gt;
&lt;li&gt;We had received questions that helped us identify areas where our solution needed improvement.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;We didn't take home the trophy, but we left with practical experience, technical lessons, and a clearer understanding of what it takes to build something useful.&lt;/p&gt;

&lt;h2&gt;
  
  
  🗺️ What's Next for Kisan Scheme Saathi?
&lt;/h2&gt;

&lt;p&gt;The hackathon was only the beginning. We have several ideas for improving and expanding the platform.&lt;/p&gt;

&lt;p&gt;Our roadmap includes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Expanding Beyond Karnataka:&lt;/strong&gt; Extending the application to states such as Maharashtra, Andhra Pradesh, Telangana, Punjab, and Uttar Pradesh, with support for their regional schemes and languages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;On-Ground Dialect Surveys:&lt;/strong&gt; Working with student volunteers and local communities to collect regional speech samples and better understand rural language variations.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;WhatsApp and IVR Voice Channels:&lt;/strong&gt; Exploring ways to make scheme information accessible through WhatsApp voice notes and phone-based services, including for people using basic mobile devices.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Our long-term goal is to make government scheme information easier to understand and access, regardless of a user's language or internet connectivity.&lt;/p&gt;




&lt;h2&gt;
  
  
  🔗 Project Links and Source Code
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;🌐 &lt;strong&gt;Live Website:&lt;/strong&gt; &lt;a href="https://kisansathi-frontend-tau.vercel.app" rel="noopener noreferrer"&gt;Kisan Scheme Saathi&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;📂 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/Sujayhs07/Kisan_sathi_" rel="noopener noreferrer"&gt;Kisan Scheme Saathi&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;📜 &lt;strong&gt;License:&lt;/strong&gt; MIT Open Source License&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  💻 Tech Stack
&lt;/h2&gt;

&lt;p&gt;Here are the technologies and services we used to build the application:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; React 18, Vite 5, Tailwind CSS v3, Lucide Icons&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Voice and AI:&lt;/strong&gt; Deepgram Nova-2 (Speech-to-Text), Murf AI Falcon 2.1 (Kannada Text-to-Speech), Google Gemini (Conversational Assistance)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; Node.js, Express, TypeScript, with a Spring Boot 3.4 Java module&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Storage and Caching:&lt;/strong&gt; PostgreSQL 14+ with a local JSON caching fallback&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  🌾 Final Thought
&lt;/h2&gt;

&lt;p&gt;If you're participating in a hackathon, don't focus only on building something impressive for the judges.&lt;/p&gt;

&lt;p&gt;Try to build something meaningful for people whose problems are often overlooked by technology.&lt;/p&gt;

&lt;p&gt;Our experience with Kisan Scheme Saathi taught us that a project does not have to win an award to be worth building. Sometimes, the most valuable outcome is discovering a real problem, attempting to solve it, learning from your mistakes, and understanding what needs to improve next.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Technology is at its greatest when it has mud on its boots.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Kisan Scheme Saathi is an open-source project licensed under the MIT License.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;If you enjoyed reading about our journey, feel free to leave a ❤️, 🦄, or comment below.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Have you ever faced an unexpected question during a hackathon presentation? We'd love to hear your experience!&lt;/strong&gt;&lt;/p&gt;

</description>
      <category>hackathon</category>
      <category>ai</category>
      <category>webdev</category>
      <category>opensource</category>
    </item>
    <item>
      <title># FINORA AI: Building a Voice-First Financial Assistant for Bharat with Murf Falcon 🎙️🇮🇳</title>
      <dc:creator>Sujay H S</dc:creator>
      <pubDate>Sat, 15 Aug 2026 09:53:52 +0000</pubDate>
      <link>https://dev.to/sujay_hs_b482e14368c435a/-finora-ai-building-a-voice-first-financial-assistant-for-bharat-with-murf-falcon-1oij</link>
      <guid>https://dev.to/sujay_hs_b482e14368c435a/-finora-ai-building-a-voice-first-financial-assistant-for-bharat-with-murf-falcon-1oij</guid>
      <description>&lt;h1&gt;
  
  
  FINORA AI: Building a Voice-First Financial Assistant for Bharat with Murf Falcon 🎙️🇮🇳
&lt;/h1&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;What if getting help with financial services was as simple as having a conversation?&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Introduction
&lt;/h2&gt;

&lt;p&gt;When I started the &lt;strong&gt;10 Days of Voice Agents — VoiceForBharat Edition&lt;/strong&gt;, I wanted to build something more than a basic voice chatbot.&lt;/p&gt;

&lt;p&gt;I wanted to build an assistant that could actually help users with financial questions, remember useful context, communicate naturally, know when it needed human support, and hand specialized questions to another agent.&lt;/p&gt;

&lt;p&gt;That idea became &lt;strong&gt;FINORA AI&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;FINORA AI is a voice-first financial services assistant designed for Bharat. The main voice agent, &lt;strong&gt;Dia&lt;/strong&gt;, interacts with users through voice and supports English and Hindi.&lt;/p&gt;

&lt;p&gt;As the challenge progressed, FINORA AI evolved from a simple voice agent into a complete voice-AI system with memory, voice workflows, outbound calling, human escalation, call analytics, and a dedicated &lt;strong&gt;Government Scheme Specialist named Krishna&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The goal was simple:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Make financial assistance less complicated and more conversational — just a voice away.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;In this blog, I'll walk through what I built, how the system works, the challenges I faced, and what I learned during the ten-day journey.&lt;/p&gt;




&lt;h1&gt;
  
  
  🌱 The Problem
&lt;/h1&gt;

&lt;p&gt;Financial services can often feel complicated.&lt;/p&gt;

&lt;p&gt;Users may have to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Search through different websites&lt;/li&gt;
&lt;li&gt;Understand financial terminology&lt;/li&gt;
&lt;li&gt;Navigate multiple pages&lt;/li&gt;
&lt;li&gt;Find the right government scheme&lt;/li&gt;
&lt;li&gt;Understand eligibility requirements&lt;/li&gt;
&lt;li&gt;Figure out where to get help&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The information may already exist, but &lt;strong&gt;finding and understanding the right information can be the difficult part&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;I wanted to make that experience more natural.&lt;/p&gt;

&lt;p&gt;Instead of:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Search → Navigate → Read → Understand
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I wanted:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Ask → Talk → Understand → Get Help
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's where voice becomes useful.&lt;/p&gt;




&lt;h1&gt;
  
  
  🎙️ Meet FINORA AI
&lt;/h1&gt;

&lt;p&gt;FINORA AI is a voice-first financial services assistant.&lt;/p&gt;

&lt;p&gt;The main agent is &lt;strong&gt;Dia&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Dia handles general financial conversations and acts as the primary entry point for the user.&lt;/p&gt;

&lt;p&gt;FINORA AI can:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Handle voice conversations&lt;/li&gt;
&lt;li&gt;Support English and Hindi&lt;/li&gt;
&lt;li&gt;Remember relevant conversational context&lt;/li&gt;
&lt;li&gt;Handle financial-service related questions&lt;/li&gt;
&lt;li&gt;Make outbound calls&lt;/li&gt;
&lt;li&gt;Escalate sensitive situations to humans&lt;/li&gt;
&lt;li&gt;Hand government-scheme questions to Krishna&lt;/li&gt;
&lt;li&gt;Track call outcomes&lt;/li&gt;
&lt;li&gt;Track agent handoffs&lt;/li&gt;
&lt;li&gt;Display analytics through a dashboard&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The idea wasn't to make Dia sound like a robot reading information.&lt;/p&gt;

&lt;p&gt;I wanted the interaction to feel more like talking to an assistant:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Sure, let me help you with that."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;rather than:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"According to the available information..."&lt;/p&gt;
&lt;/blockquote&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%2Fgxi99xwexxzhb8wseh1b.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%2Fgxi99xwexxzhb8wseh1b.png" alt=" " width="800" height="687"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Caption: FINORA AI's minimalistic interface, designed to keep voice interaction simple and easy to understand.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;The website follows the same philosophy as the voice experience:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Minimalistic. Clean. Easy to understand.&lt;/strong&gt;&lt;/p&gt;


&lt;h1&gt;
  
  
  🎯 Why Voice?
&lt;/h1&gt;

&lt;p&gt;Voice removes a layer of complexity.&lt;/p&gt;

&lt;p&gt;Users don't have to figure out exactly what to type or where to click.&lt;/p&gt;

&lt;p&gt;They can simply say:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"What documents do I need?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;or:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Can you explain this government scheme?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;They can also speak naturally in Hindi.&lt;/p&gt;

&lt;p&gt;For FINORA AI, voice isn't just another feature.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Voice is the primary interface.&lt;/strong&gt;&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%2F2236u7t2mqsxl2lo03ub.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%2F2236u7t2mqsxl2lo03ub.png" alt=" " width="800" height="414"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: A live voice conversation with Dia through the FINORA AI interface.&lt;/em&gt;&lt;/p&gt;


&lt;h1&gt;
  
  
  🏗️ How FINORA AI Works
&lt;/h1&gt;

&lt;p&gt;The overall architecture looks like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;flowchart TD
    U[User] --&amp;gt; L[LiveKit]

    L --&amp;gt; STT[Speech to Text]

    STT --&amp;gt; D[Dia - Main Agent]

    D --&amp;gt; LLM[LLM Reasoning]
    LLM --&amp;gt; M[Memory / Relevant Context]

    D --&amp;gt; K[Krishna - Government Scheme Specialist]

    D --&amp;gt; H[Human Escalation]

    K --&amp;gt; GS[Government Scheme Guidance]

    D --&amp;gt; TTS[Murf Falcon]
    K --&amp;gt; TTS

    TTS --&amp;gt; L
    L --&amp;gt; U

    L --&amp;gt; A[Call Analytics]
    A --&amp;gt; DB[(Analytics Database)]
    DB --&amp;gt; DASH[FINORA Dashboard]
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;At a high level, the voice flow is:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User Speech
     ↓
LiveKit
     ↓
Speech-to-Text
     ↓
Dia
     ↓
LLM Reasoning
     ↓
Response / Specialist / Human Escalation
     ↓
Murf Falcon
     ↓
User
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;LiveKit handles the real-time communication layer.&lt;/p&gt;

&lt;p&gt;The agents handle the conversation and decision-making.&lt;/p&gt;

&lt;p&gt;Murf Falcon provides the voice output.&lt;/p&gt;

&lt;p&gt;The analytics system records call-level information so the dashboard can show how the system is performing.&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%2F6ew9a74cwm5usxvft1wb.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%2F6ew9a74cwm5usxvft1wb.png" alt=" " width="800" height="533"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: High-level architecture of FINORA AI, showing the voice pipeline, Dia, Krishna, human escalation, Murf Falcon, and call analytics.&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  🔊 Murf Falcon — Giving FINORA AI a Voice
&lt;/h1&gt;

&lt;p&gt;A voice assistant isn't only about what it says.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How it sounds matters too.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;For FINORA AI, I used &lt;strong&gt;Murf Falcon&lt;/strong&gt;, the fastest TTS API highlighted in the challenge.&lt;/p&gt;

&lt;p&gt;The goal was to make the voice experience:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Natural&lt;/li&gt;
&lt;li&gt;Clear&lt;/li&gt;
&lt;li&gt;Fast&lt;/li&gt;
&lt;li&gt;Friendly&lt;/li&gt;
&lt;li&gt;Conversational&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Instead of feeling like a traditional IVR system, I wanted the interaction to feel like a real conversation.&lt;/p&gt;

&lt;p&gt;This was one of the most interesting parts of the project because voice changes how you design the entire agent.&lt;/p&gt;

&lt;p&gt;A text chatbot can give a long answer.&lt;/p&gt;

&lt;p&gt;A voice assistant needs to think about:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;What should I say, and how should I say it?&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h1&gt;
  
  
  🇮🇳 English + Hindi Voice Support
&lt;/h1&gt;

&lt;p&gt;FINORA AI supports &lt;strong&gt;English and Hindi across both agents — Dia and Krishna&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The language doesn't stop when the conversation is handed from one agent to another.&lt;/p&gt;

&lt;p&gt;If a user starts speaking in English, Dia and Krishna can continue in English.&lt;/p&gt;

&lt;p&gt;If the user starts speaking in Hindi, the conversation can continue in Hindi even after the handoff.&lt;/p&gt;

&lt;h2&gt;
  
  
  English
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"What documents do I need for PM-KISAN?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;Dia:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"This is a government-scheme question, so I'll connect you with Krishna, our government-scheme specialist."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;Krishna:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Hi, I understand you're asking about the documents required for PM-KISAN. Let me help you with that."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Hindi
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"PM-KISAN के लिए कौन-कौन से documents चाहिए?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;Dia:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"यह सरकारी योजना से जुड़ा सवाल है, इसलिए मैं आपको कृष्ण, हमारे सरकारी योजना विशेषज्ञ से जोड़ता हूँ।"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;strong&gt;Krishna:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"नमस्ते, मुझे समझ आया कि आप PM-KISAN के लिए जरूरी दस्तावेजों के बारे में पूछ रहे हैं। मैं आपकी मदद करता हूँ।"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This was important to me because &lt;strong&gt;building for Bharat isn't just about adding Indian languages — the entire conversation should continue naturally in the user's preferred language, even when another specialist agent takes over.&lt;/strong&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  🧠 Memory and Context
&lt;/h1&gt;

&lt;p&gt;Another important part of FINORA AI is memory.&lt;/p&gt;

&lt;p&gt;A useful assistant shouldn't force returning users to repeat relevant information unnecessarily.&lt;/p&gt;

&lt;p&gt;FINORA AI can retain useful conversational context where appropriate, helping the interaction feel more continuous.&lt;/p&gt;

&lt;p&gt;However, because FINORA is related to financial services, I designed the system with clear privacy boundaries.&lt;/p&gt;

&lt;p&gt;Sensitive information such as:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;OTPs&lt;/li&gt;
&lt;li&gt;PINs&lt;/li&gt;
&lt;li&gt;Passwords&lt;/li&gt;
&lt;li&gt;CVVs&lt;/li&gt;
&lt;li&gt;Authentication secrets&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;should not be stored as ordinary conversational memory.&lt;/p&gt;

&lt;p&gt;The goal is to remember &lt;strong&gt;useful context&lt;/strong&gt;, not sensitive credentials.&lt;/p&gt;




&lt;h1&gt;
  
  
  🛠️ Agent Capabilities and Workflows
&lt;/h1&gt;

&lt;p&gt;Rather than trying to make one agent responsible for everything, FINORA AI separates responsibilities.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Dia&lt;/strong&gt; is the main financial assistant.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Krishna&lt;/strong&gt; focuses specifically on government schemes.&lt;/p&gt;

&lt;p&gt;And when something requires a human, the system can use the existing &lt;strong&gt;human escalation workflow&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;This creates a simple separation:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;General Financial Assistance
            ↓
           DIA

Government Scheme Questions
            ↓
         KRISHNA

Sensitive / Human-Only Issues
            ↓
      HUMAN SUPPORT
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This separation became one of the most important architectural decisions in the project.&lt;/p&gt;




&lt;h1&gt;
  
  
  📞 Outbound Calling
&lt;/h1&gt;

&lt;p&gt;FINORA AI also supports outbound calling workflows.&lt;/p&gt;

&lt;p&gt;This changes the interaction from:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Ask me something."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;to:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Let me help you with something you need."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Building outbound calls made me think more carefully about how an agent starts a conversation.&lt;/p&gt;

&lt;p&gt;An outbound voice agent needs to:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Introduce itself&lt;/li&gt;
&lt;li&gt;Explain why it is calling&lt;/li&gt;
&lt;li&gt;Establish context quickly&lt;/li&gt;
&lt;li&gt;Avoid sounding overly scripted&lt;/li&gt;
&lt;li&gt;Give the user a clear next step&lt;/li&gt;
&lt;/ul&gt;




&lt;h1&gt;
  
  
  🧑‍💼 When AI Should Ask for Human Help
&lt;/h1&gt;

&lt;p&gt;One of the most important things I learned during the challenge was that &lt;strong&gt;AI shouldn't try to solve everything&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;For financial services, this becomes even more important.&lt;/p&gt;

&lt;p&gt;For example, imagine a user says:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"I see a transaction that I didn't make."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Dia shouldn't pretend it can investigate or resolve the issue itself.&lt;/p&gt;

&lt;p&gt;Instead, the system recognizes that human assistance may be required.&lt;/p&gt;

&lt;p&gt;The flow becomes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User reports possible fraud
          ↓
Dia identifies the situation
          ↓
Dia explains why human help is needed
          ↓
Ask for user permission
          ↓
User agrees
          ↓
Create escalation
          ↓
Generate reference ID
          ↓
Send request to human support
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The user receives a reference ID and a clear explanation of what happens next.&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%2Fto901jitzro4khmx4y8f.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%2Fto901jitzro4khmx4y8f.png" alt=" " width="800" height="576"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: FINORA AI creating a human-support request after receiving the user's permission.&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  🔐 Privacy by Design
&lt;/h1&gt;

&lt;p&gt;Since FINORA AI is related to financial services, privacy cannot be an afterthought.&lt;/p&gt;

&lt;p&gt;The escalation workflow is designed to avoid sharing unnecessary sensitive information.&lt;/p&gt;

&lt;p&gt;It should never send:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Passwords&lt;/li&gt;
&lt;li&gt;OTPs&lt;/li&gt;
&lt;li&gt;PINs&lt;/li&gt;
&lt;li&gt;CVVs&lt;/li&gt;
&lt;li&gt;Full account numbers&lt;/li&gt;
&lt;li&gt;Authentication secrets&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Instead, the human receives a short summary containing only useful information.&lt;/p&gt;

&lt;p&gt;For example:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Reference: FIN-2026-0001

Reason:
Possible Fraud

Summary:
Caller reports an unauthorized transaction.

Urgency:
High

Language:
Hindi

Preferred Follow-up:
Phone
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The principle is simple:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Give the human enough information to help without giving them information they don't need.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h1&gt;
  
  
  🤝 Meet Krishna — Government Scheme Specialist
&lt;/h1&gt;

&lt;p&gt;As FINORA AI grew, I realized that Dia shouldn't have to be an expert in everything.&lt;/p&gt;

&lt;p&gt;So I created a separate specialist agent:&lt;/p&gt;

&lt;h2&gt;
  
  
  Krishna
&lt;/h2&gt;

&lt;p&gt;Krishna is FINORA AI's &lt;strong&gt;Government Scheme Specialist&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;His job is focused specifically on government financial and welfare schemes.&lt;/p&gt;

&lt;p&gt;For example:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;PM-KISAN&lt;/li&gt;
&lt;li&gt;PM Jan Dhan Yojana&lt;/li&gt;
&lt;li&gt;PM SVANidhi&lt;/li&gt;
&lt;li&gt;Mudra-related schemes&lt;/li&gt;
&lt;li&gt;Pension schemes&lt;/li&gt;
&lt;li&gt;Insurance schemes&lt;/li&gt;
&lt;li&gt;Scholarship schemes&lt;/li&gt;
&lt;li&gt;Other supported government schemes&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The key idea is that Krishna has a &lt;strong&gt;smaller and more focused responsibility&lt;/strong&gt; than Dia.&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%2Fuo62ciz33d1txrsxb5dn.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%2Fuo62ciz33d1txrsxb5dn.png" alt=" " width="799" height="409"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Caption: Krishna, FINORA AI's dedicated Government Scheme Specialist.&lt;/em&gt;&lt;/p&gt;


&lt;h1&gt;
  
  
  🔄 Real Agent-to-Agent Handoff
&lt;/h1&gt;

&lt;p&gt;This is one of the features I am most excited about.&lt;/p&gt;

&lt;p&gt;The system doesn't simply ask Dia to generate a response about government schemes.&lt;/p&gt;

&lt;p&gt;Instead, Dia can hand the conversation over to &lt;strong&gt;Krishna, a separate specialist agent&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;For example:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"What documents are required for PM-KISAN?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Dia responds:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"This is a government-scheme question, so I'll connect you with Krishna, our government-scheme specialist. You won't need to repeat your question."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Then Krishna takes over:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Hi, I'm Krishna, the government-scheme specialist. I understand you're asking about the documents required for PM-KISAN. Let me help you with that."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The important part is that the user &lt;strong&gt;doesn't have to explain the entire problem again&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The relevant context is passed during the handoff.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;flowchart LR
    U[User] --&amp;gt; D[Dia]

    D --&amp;gt;|Government Scheme Question| K[Krishna]

    K --&amp;gt;|Continues Same Conversation| U

    D --&amp;gt;|Sensitive Issue| H[Human Support]
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This makes FINORA AI a &lt;strong&gt;multi-agent voice system&lt;/strong&gt;, rather than just a single chatbot with many prompts.&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%2Fjzo5pq7k6fouqfnvag9z.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%2Fjzo5pq7k6fouqfnvag9z.png" alt=" " width="799" height="414"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: Dia handing the conversation to Krishna without requiring the user to repeat their question.&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  📊 Building the Call Analytics Dashboard
&lt;/h1&gt;

&lt;p&gt;After adding multiple capabilities, I wanted to answer a simple question:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Is the agent actually working?&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That's why I built a call analytics dashboard.&lt;/p&gt;

&lt;p&gt;The dashboard tracks real calls and displays:&lt;/p&gt;

&lt;h3&gt;
  
  
  Total Calls
&lt;/h3&gt;

&lt;p&gt;The total number of calls handled.&lt;/p&gt;

&lt;h3&gt;
  
  
  Successful Calls
&lt;/h3&gt;

&lt;p&gt;Calls that reached the defined success condition.&lt;/p&gt;

&lt;h3&gt;
  
  
  Failed Calls
&lt;/h3&gt;

&lt;p&gt;Calls that ended without reaching the intended outcome.&lt;/p&gt;

&lt;p&gt;I also added agent handoff information.&lt;/p&gt;

&lt;p&gt;For example:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Total Calls: 25

Successful Calls: 21

Failed Calls: 4

Agent Handoffs: 7

Handoffs to Krishna: 7
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Replace these example numbers with your actual dashboard values before publishing.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;These numbers come from actual call activity rather than being hardcoded.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;[INSERT IMAGE: FINORA AI CALL ANALYTICS DASHBOARD]&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: FINORA AI's call analytics dashboard showing real call outcomes and specialist handoffs.&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  📈 Handoff Analytics
&lt;/h1&gt;

&lt;p&gt;An important design decision was that a specialist handoff should &lt;strong&gt;not create a second call&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;For example:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;CALL-001
   │
   ├── Dia
   │
   └── Krishna
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This is still:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;ONE CALL&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;with the handoff recorded as part of that call.&lt;/p&gt;

&lt;p&gt;This keeps the analytics accurate.&lt;/p&gt;

&lt;p&gt;The dashboard can therefore show the complete journey of a call without artificially inflating the number of calls.&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%2Fbkbuj7mtfcn336bay2m0.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%2Fbkbuj7mtfcn336bay2m0.png" alt=" " width="800" height="727"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: Call analytics recording a Dia-to-Krishna handoff as part of the same conversation and call.&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  🎨 Designing a Simple User Experience
&lt;/h1&gt;

&lt;p&gt;While building all these features, I didn't want the website to become complicated.&lt;/p&gt;

&lt;p&gt;The FINORA AI interface is intentionally:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Minimalistic&lt;/li&gt;
&lt;li&gt;Clean&lt;/li&gt;
&lt;li&gt;Easy to understand&lt;/li&gt;
&lt;li&gt;Voice-focused&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The main navigation includes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Voice Assistant&lt;/li&gt;
&lt;li&gt;Finance Dashboard&lt;/li&gt;
&lt;li&gt;Call Analytics&lt;/li&gt;
&lt;li&gt;Settings&lt;/li&gt;
&lt;li&gt;Help&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The user should be able to open the website and immediately understand what to do.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Talk to Dia. Get help.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The technology stays in the background.&lt;/p&gt;




&lt;h1&gt;
  
  
  🧩 The Challenges I Faced
&lt;/h1&gt;

&lt;p&gt;Building FINORA AI over ten days wasn't a simple process where everything worked on the first attempt.&lt;/p&gt;

&lt;p&gt;There were several challenges.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Building a Natural Voice Experience
&lt;/h2&gt;

&lt;p&gt;Connecting real-time voice communication, the agent, and TTS required understanding how all the components interact.&lt;/p&gt;

&lt;p&gt;A small configuration problem could mean the agent connected successfully but didn't respond correctly.&lt;/p&gt;

&lt;p&gt;I had to debug the voice pipeline step by step instead of assuming the issue was always with the LLM.&lt;/p&gt;




&lt;h2&gt;
  
  
  2. Deciding When AI Should Stop
&lt;/h2&gt;

&lt;p&gt;One of the biggest design questions was:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;When should AI answer, and when should AI stop?&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Too many escalations make the experience frustrating.&lt;/p&gt;

&lt;p&gt;Too few can make the system unsafe.&lt;/p&gt;

&lt;p&gt;So I created clear conditions for human escalation and specialist handoff.&lt;/p&gt;




&lt;h2&gt;
  
  
  3. Keeping Analytics Real
&lt;/h2&gt;

&lt;p&gt;Displaying numbers on a dashboard is easy.&lt;/p&gt;

&lt;p&gt;Making those numbers represent &lt;strong&gt;actual calls&lt;/strong&gt; is much harder.&lt;/p&gt;

&lt;p&gt;The system needs to record the call outcome and handoff information from the actual voice session.&lt;/p&gt;

&lt;p&gt;This helped me understand the difference between building a UI and building a system that produces meaningful data.&lt;/p&gt;




&lt;h2&gt;
  
  
  4. Multi-Agent Handoff
&lt;/h2&gt;

&lt;p&gt;The biggest challenge with Krishna was making the handoff feel like one continuous conversation.&lt;/p&gt;

&lt;p&gt;Krishna needs to understand:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;What the user asked&lt;/li&gt;
&lt;li&gt;Why Dia handed the conversation over&lt;/li&gt;
&lt;li&gt;What language the user is using&lt;/li&gt;
&lt;li&gt;What context is relevant&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The user shouldn't have to start again with:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Hello, I was asking..."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The handoff should feel natural.&lt;/p&gt;




&lt;h1&gt;
  
  
  🚀 How You Can Build Your Own Voice Agent
&lt;/h1&gt;

&lt;p&gt;If you want to build a voice agent yourself, there are four major pieces you need.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Speech-to-Text
&lt;/h2&gt;

&lt;p&gt;Converts the user's speech into text.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User Voice
   ↓
Speech-to-Text
   ↓
Text
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  2. LLM / Agent
&lt;/h2&gt;

&lt;p&gt;The agent understands the request and decides what to do.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Text
 ↓
LLM
 ↓
Decision
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  3. Text-to-Speech
&lt;/h2&gt;

&lt;p&gt;The response is converted back into voice.&lt;/p&gt;

&lt;p&gt;For FINORA AI, I use &lt;strong&gt;Murf Falcon&lt;/strong&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Agent Response
      ↓
Murf Falcon
      ↓
Natural Voice
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  4. Real-Time Transport
&lt;/h2&gt;

&lt;p&gt;A real-time communication layer connects the user and the voice agent.&lt;/p&gt;

&lt;p&gt;FINORA AI uses &lt;strong&gt;LiveKit&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The simplified pipeline is:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User
 ↓
Microphone
 ↓
LiveKit
 ↓
Speech-to-Text
 ↓
Dia / Krishna
 ↓
LLM
 ↓
Murf Falcon
 ↓
LiveKit
 ↓
User
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h1&gt;
  
  
  ⚙️ Running the Project
&lt;/h1&gt;

&lt;p&gt;Start by cloning the project:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/Sujayhs07/murf-livekit-starter/tree/DAY10
&lt;span class="nb"&gt;cd &lt;/span&gt;murf-livekit-starter
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Create a &lt;code&gt;.env&lt;/code&gt; file for your API keys and configuration.&lt;/p&gt;

&lt;p&gt;For example:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;MURF_API_KEY=your_murf_api_key

LIVEKIT_URL=your_livekit_url
LIVEKIT_API_KEY=your_livekit_api_key
LIVEKIT_API_SECRET=your_livekit_api_secret
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Add any additional credentials required by your implementation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Never commit your &lt;code&gt;.env&lt;/code&gt; file to GitHub.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Add it to &lt;code&gt;.gitignore&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;.env
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then install the dependencies and start the LiveKit agent according to the project's setup instructions.&lt;/p&gt;

&lt;p&gt;Once the browser client connects, you can start a conversation with Dia.&lt;/p&gt;




&lt;h1&gt;
  
  
  🧪 Testing FINORA AI
&lt;/h1&gt;

&lt;p&gt;I tested different conversation paths.&lt;/p&gt;

&lt;h2&gt;
  
  
  Test 1 — Normal Question
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"What is compound interest?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Expected:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User → Dia
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No specialist handoff.&lt;/p&gt;




&lt;h2&gt;
  
  
  Test 2 — Government Scheme
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"What documents do I need for PM-KISAN?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Expected:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User
 ↓
Dia
 ↓
Krishna
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The user should not have to repeat the question.&lt;/p&gt;




&lt;h2&gt;
  
  
  Test 3 — Hindi Government Scheme
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"PM-KISAN के लिए कौन-कौन से documents चाहिए?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Expected:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User
 ↓
Dia — Hindi
 ↓
Krishna — Hindi
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The language should remain consistent across the handoff.&lt;/p&gt;




&lt;h2&gt;
  
  
  Test 4 — Possible Fraud
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;User:&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Someone made a transaction from my account."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Expected:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User
 ↓
Dia
 ↓
Permission
 ↓
Human Escalation
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It should &lt;strong&gt;not&lt;/strong&gt; be routed to Krishna.&lt;/p&gt;




&lt;h2&gt;
  
  
  Test 5 — Analytics
&lt;/h2&gt;

&lt;p&gt;After the call ends, the dashboard should reflect the real result:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Total Calls ↑

Successful Calls ↑

Failed Calls ↑

Handoffs to Krishna ↑
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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%2Frx6ousezakiaq4efwa3i.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%2Frx6ousezakiaq4efwa3i.png" alt=" " width="799" height="549"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Caption: A snapshot of FINORA AI's end-to-end voice-agent workflow.&lt;/em&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  📚 What I Learned
&lt;/h1&gt;

&lt;p&gt;The biggest lesson I learned from this challenge is that building a voice agent isn't really about making an AI talk.&lt;/p&gt;

&lt;p&gt;The interesting engineering begins after that.&lt;/p&gt;

&lt;p&gt;You have to think about:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;What should the agent know?&lt;/li&gt;
&lt;li&gt;What should it remember?&lt;/li&gt;
&lt;li&gt;When should it use a tool?&lt;/li&gt;
&lt;li&gt;When should it stop?&lt;/li&gt;
&lt;li&gt;When should a human take over?&lt;/li&gt;
&lt;li&gt;When should another agent take over?&lt;/li&gt;
&lt;li&gt;How should context move between agents?&lt;/li&gt;
&lt;li&gt;How do we measure success?&lt;/li&gt;
&lt;li&gt;How do we protect user information?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These questions completely changed how I think about voice AI.&lt;/p&gt;




&lt;h1&gt;
  
  
  🔮 What I'd Build Next
&lt;/h1&gt;

&lt;p&gt;FINORA AI still has a lot of room to grow.&lt;/p&gt;

&lt;p&gt;Some things I'd like to work on next are:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;More Indian languages&lt;/li&gt;
&lt;li&gt;More specialist agents&lt;/li&gt;
&lt;li&gt;Better multilingual conversations&lt;/li&gt;
&lt;li&gt;Improved personalization&lt;/li&gt;
&lt;li&gt;More advanced analytics&lt;/li&gt;
&lt;li&gt;More government-scheme coverage&lt;/li&gt;
&lt;li&gt;Better accessibility&lt;/li&gt;
&lt;li&gt;Smarter follow-up workflows&lt;/li&gt;
&lt;li&gt;A richer human-support dashboard&lt;/li&gt;
&lt;li&gt;More robust real-world integrations&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I'd especially like to expand the multi-agent architecture so FINORA AI can route users to different specialists depending on what they need.&lt;/p&gt;




&lt;h1&gt;
  
  
  ❤️ What 10 Days Taught Me
&lt;/h1&gt;

&lt;p&gt;When I started the challenge, I thought the hardest part would be getting an AI agent to speak.&lt;/p&gt;

&lt;p&gt;By Day 10, I realized:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Voice is only the beginning.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The real challenge is building an agent that knows how to behave.&lt;/p&gt;

&lt;p&gt;An agent that can say:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"I can help with that."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;But also knows when to say:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Let me connect you with Krishna, our specialist."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;And sometimes:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"This needs human assistance."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That's what I wanted FINORA AI to become.&lt;/p&gt;

&lt;p&gt;Not just an AI that talks.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;An AI that knows what to do next.&lt;/strong&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  🔗 Project Links
&lt;/h1&gt;

&lt;h3&gt;
  
  
  GitHub Repository
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://github.com/Sujayhs07/murf-livekit-starter/tree/DAY10" rel="noopener noreferrer"&gt;https://github.com/Sujayhs07/murf-livekit-starter/tree/DAY10&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  Challenge
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/murf-ai/voice-for-bharat-challenge-2026/tree/main/challenges" rel="noopener noreferrer"&gt;10 Days of Voice Agents — VoiceForBharat Edition&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;




&lt;h1&gt;
  
  
  🙌 Final Thoughts
&lt;/h1&gt;

&lt;p&gt;Building FINORA AI during the &lt;strong&gt;10 Days of Voice Agents — VoiceForBharat Edition&lt;/strong&gt; has been one of the most hands-on AI projects I've worked on.&lt;/p&gt;

&lt;p&gt;Over these ten days, I explored real-time voice communication, TTS, agent design, memory, outbound calls, human escalation, call analytics, and multi-agent handoffs.&lt;/p&gt;

&lt;p&gt;More importantly, I learned that building a useful AI system isn't only about adding more features.&lt;/p&gt;

&lt;p&gt;It's about knowing &lt;strong&gt;when each feature should be used&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Dia doesn't need to know everything.&lt;/p&gt;

&lt;p&gt;Krishna doesn't need to handle everything.&lt;/p&gt;

&lt;p&gt;And AI doesn't need to handle everything either.&lt;/p&gt;

&lt;p&gt;Sometimes the best decision an AI can make is to &lt;strong&gt;ask another specialist or a human for help&lt;/strong&gt;.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;A good voice agent isn't the one that tries to do everything. It's the one that knows what it can do, knows what it can't, and knows what to do next.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Thanks to &lt;strong&gt;Murf AI&lt;/strong&gt; for organizing the &lt;strong&gt;10 Days of Voice Agents — VoiceForBharat Edition&lt;/strong&gt; and giving me the opportunity to build, experiment, break things, fix them, and learn along the way.&lt;/p&gt;

&lt;h2&gt;
  
  
  🛠️ Built With
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;🎙️ &lt;strong&gt;Murf Falcon&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;⚡ &lt;strong&gt;LiveKit&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🧠 &lt;strong&gt;LLM-powered Agents&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🗣️ &lt;strong&gt;English + Hindi&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🤝 &lt;strong&gt;Human Escalation&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🔄 &lt;strong&gt;Multi-Agent Handoff&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;📊 &lt;strong&gt;Call Analytics&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;🎨 &lt;strong&gt;Minimalistic Web Interface&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;#VoiceForBharat #10DaysOfVoiceAgents #MurfAI #VoiceAI #GenerativeAI #AI #FinTech&lt;/strong&gt;@MURFAI&lt;/p&gt;

</description>
      <category>ai</category>
      <category>python</category>
      <category>beginners</category>
      <category>api</category>
    </item>
  </channel>
</rss>
