🎙️ Munshi (मुंशी)

Hire a voice receptionist by explaining the job. Not by writing prompts.

Build Arena Hackathon · Problem Statement 2: No-Code Voice AI Agent Builder (ElevenLabs)
Team: DOT DOT COM (Ansh Varshney)
Live Production: https://munshi-hackathon-prototype.vercel.app

---

🌟 The Vision

Out of every 100 incoming calls to small clinics, salons, and repair shops, 46 go unanswered. Dr. Mira Sharma runs a two-chair dental clinic in Indirapuram. When her hands are busy treating patients, the phone rings off the hook.

She knows her front-desk operations inside and out, but she cannot code, and current voice agent builders demand complex system prompts, JSON schemas, API keys, and webhook pipelines.

Munshi solves this entirely. Instead of writing prompts or configuring nodes, the business owner explains the job out loud—exactly like briefing a new human employee. Munshi listens, compiles a structured rule card, runs the receptionist through a rigorous 24-run probabilistic Trial Day, allows her to coach mistakes in one plain sentence, and safeguards all tool calls through ArmorIQ intent-aware runtime governance before going live.

       🗣️ EXPLAIN                      📋 CHECK                     🧪 TEST
┌──────────────────────┐        ┌──────────────────────┐     ┌──────────────────────┐
│ Voice Interview /    │ ─────▶ │ Structured Job Card  │ ──▶ │ Trial Day: 8 Tests   │
│ Plain-Text Starter   │        │ (Zod Validated)      │     │ × 3 Probabilistic    │
└──────────────────────┘        └──────────────────────┘     └──────────┬───────────┘
                                                                        │
                                                                        ▼
       🚀 GO LIVE                      🛡️ SECURE                    🔧 COACH
┌──────────────────────┐        ┌──────────────────────┐     ┌──────────────────────┐
│ Verified Merge       │ ◀───── │ ArmorIQ Runtime      │ ◀── │ Spoken One-Sentence  │
│ 3/3 Pass Gate Only   │        │ Sentry & Guardrails  │     │ Fix + Auto-Test Gen  │
└──────────────────────┘        └──────────────────────┘     └──────────────────────┘

---

📸 Product Walkthrough & Visual Proof

1. Bilingual Landing Page & 60-Second Explainer Video

A warm, human interface designed with a tailored palette (warm paper #FBF8F3, deep teal #0D7377, saffron #E8833A). Includes a real-time interactive phone simulator and a 60-second bilingual explainer video with synchronized English and Hindi narration tracks that preserve timestamp on switch.

English Landing Experience	Native Hindi Experience (633 Translated Strings)

60-Second Video Player (Dual Audio Stream)	Mobile-First Responsive Experience

---

2. The 5-Step Agent Lifecycle

Step 1: Voice Interview — "Explain the Job"

The owner taps the microphone and speaks naturally. Munshi's conversational interviewer asks 8 everyday questions (business name, hours, services, prices, booking rules, emergency hand-offs, what to never do, and what to say when unsure).

Step 2: The Job Card — Human-Readable Rules

Gemini distills the voice transcript into a clear, structured Job Card. The owner can inspect approved prices, edit hours, or select a voice from the ElevenLabs library with one-click audio previews. No technical jargon, no raw prompts.

Step 3: Trial Day — 8 Tricky Calls, 3 Runs Each

Before a single real customer calls, Munshi subjects the receptionist to Trial Day: 8 edge-case calls evaluated 3 times each (24 total runs) using the ElevenLabs Agent Testing API. In version 1, a caller asks for teeth whitening; since it wasn't on the price list, the receptionist hallucinates a price of ₹3,000. Trial Day catches it immediately.

Step 4: Spoken Coaching — "Fix It in One Sentence"

The owner doesn't edit a 2,000-word prompt. She simply speaks or types one corrective sentence:

"If a price is not on the list, never guess. Say you will confirm and take a callback number."

Gemini analyzes the correction, updates the Job Card, and automatically synthesizes a permanent regression test. Munshi branches the ElevenLabs agent into Version 2.

Step 5: Version 2 Verification & Gated Go-Live

Trial Day re-runs all tests against Version 2. All 8 situations pass 3 out of 3 (24/24 passed). The Go-Live gate unlocks, merging the branch to production and generating:

• A public call link for customers (/talk/[clinicId])
• An embeddable <elevenlabs-convai> web widget snippet
• Verified telephone routing via SIP/Twilio

Version 2: 24/24 Tests Passed	Gated Go-Live Dashboard

---

🛠️ Complete Technology Stack & Tools Used

1. 🎙️ ElevenLabs Conversational AI Suite

Munshi leverages the full depth of the ElevenLabs platform:

• Agents API (@elevenlabs/react, @elevenlabs/elevenlabs-js): Orchestrates conversational state machines with low-latency WebSockets.
• Client Tools: Mid-interview extraction where the interviewer calls client tools to update the UI live as the user speaks.
• Server Webhook Tools: The receptionist autonomously executes real-time backend actions mid-call (check-slots, book-slot, log-callback).
• Agent Testing API: Uses POST /v1/convai/agent-testing/create and POST /v1/convai/agents/{id}/run-tests with repeat_count: 3 to evaluate next-reply accuracy, tool calls, and simulation mocks.
• Branching & Versioning: Isolated branch creation (/branches) and atomic merges (/merge) guarantee that untested prompt changes never reach patients.
• Scribe v2 Speech-to-Text: High-accuracy transcription for vocal interview answers and spoken coaching.
• Flash v2.5 Multilingual TTS: Ultra-low latency voice synthesis in English and Hindi (featuring warm Indian voices like "PP").
• Post-Call Webhook: HMAC-SHA256 signature verification for call transcripts, duration, audio recordings, and sentiment.

---

2. 🛡️ ArmorIQ — Runtime Agent Security & Intent Governance

Integrating ArmorIQ provides enterprise-grade safety and intent verification for our autonomous voice receptionist:

                  ┌──────────────────────────────────────────────┐
                  │          ElevenLabs Voice Receptionist       │
                  └──────────────────────┬───────────────────────┘
                                         │ Tool Call: book_slot
                                         ▼
                  ┌──────────────────────────────────────────────┐
                  │       🛡️ ArmorIQ Runtime Sentry Layer        │
                  ├──────────────────────────────────────────────┤
                  │ 1. Intent Verification: Authorized action?   │
                  │ 2. Parameter Sanitization: Valid datetime?   │
                  │ 3. Guardrails: No hallucinated discounts     │
                  │ 4. Prompt Injection & Jailbreak Defense      │
                  └──────────────────────┬───────────────────────┘
                                         │ Evaluated: ALLOW
                                         ▼
                  ┌──────────────────────────────────────────────┐
                  │      MongoDB Atlas Verified Slot Booking     │
                  └──────────────────────────────────────────────┘

• Intent-Aware Policy Enforcement: Before any tool call (book-slot, check-slots, log-callback) is executed on our database, ArmorIQ validates the agent's intent against the clinic's declared policy rules.
• Sentry Interceptor: Protects against prompt injection attacks (e.g. callers tricking the AI into booking appointments for free or altering clinic operating parameters).
• Price Hallucination Barrier: Strict boundary policies prevent the voice agent from quoting rates or discounts outside the verified Job Card list.
• Cryptographic Audit Logs: Every autonomous action taken by the voice agent is recorded with cryptographic provenance for healthcare compliance and dispute resolution.

---

3. 🧠 Google Gemini 2.5

• Conversational Slot Filling: Extracts 8 structured fields from raw voice interviews into strict Zod schemas.
• Rule Compiler: Compiles human-readable Job Cards into prompt guidelines without exposing raw engineering tokens.
• Coach & Test Synthesizer: Converts one-sentence human feedback into operational rules and generates negative test cases for the ElevenLabs test pack.
• Conversation Intelligence: Performs post-call sentiment analysis, caller intent classification, and resolution scoring.

---

4. 🗄️ MongoDB Atlas (Verified Transactional Calendar)

• Zero Fake Bookings: A booking is only valid if book_slot returns a verified bk_ ID generated by MongoDB with atomic slot reservation.
• Double-Booking Prevention: Strict compound unique index (clinicId + start) stops race conditions across simultaneous callers.
• Audit Verification: The post-call processor cross-checks any booking ID spoken by the agent against MongoDB records. If the agent uttered an unverified booking ID, it is immediately flagged as a hallucination in the dashboard.

---

5. ⚡ Next.js 15, React 19 & Three.js

• Next.js 15.5.28 App Router: Route handlers, streaming UI, and security headers.
• Zero Browser API Key Exposure: All client voice sessions run via short-lived signed URLs (GET /v1/convai/conversation/get-signed-url).
• 3D Audio-Reactive Orb: Custom @react-three/fiber visualizer with real-time frequency analysis during active voice calls.
• Complete Bilingual i18n: Bespoke client-side localization system (src/lib/i18n.tsx) supporting 633 English and Hindi terms with zero external runtime overhead.

---

🏛️ System Architecture

                    ┌───────────────────────────────────────────┐
                    │     Client Browser / Mobile / Widget      │
                    │   (Next.js 15, React 19, Three.js Orb)    │
                    └─────────────────────┬─────────────────────┘
                                          │ 1. Ephemeral Signed URL Request
                                          ▼
┌────────────────────────────────────────────────────────────────────────────────────────┐
│                                 Munshi Core Engine                                      │
├────────────────────────────────────────────────────────────────────────────────────────┤
│  • Interview Coordinator          • Rule & Prompt Compiler (compile.ts)               │
│  • Zod Card Validator (jobcard.ts) • Test Pack Orchestrator (testpack.ts)               │
│  • Speech-to-Text Pipeline (STT)  • Versioning & Branch Manager (workflow.ts)          │
└───────────────┬─────────────────────────┬───────────────────────────────┬──────────────┘
                │                         │                               │
                ▼                         ▼                               ▼
    ┌──────────────────────┐  ┌──────────────────────┐  ┌────────────────────────────────┐
    │  ElevenLabs Platform │  │   Google Gemini 2.5  │  │   🛡️ ArmorIQ Security Sentry   │
    ├──────────────────────┤  ├──────────────────────┤  ├────────────────────────────────┤
    │ • Conversational AI  │  │ • Job Card Analyzer  │  │ • Intent-Aware Guardrails      │
    │ • Agent Testing Suite│  │ • Rule Synthesis     │  │ • Tool Call Interception       │
    │ • Flash v2.5 TTS     │  │ • Test Generation    │  │ • Prompt Injection Defense     │
    │ • Scribe v2 STT      │  │ • Sentiment Scoring  │  │ • Audit Log & Provenance       │
    └───────────┬──────────┘  └──────────────────────┘  └───────────────┬────────────────┘
                │                                                       │
                └─────────────────────────┬─────────────────────────────┘
                                          ▼ Verified Action
                              ┌──────────────────────┐
                              │  MongoDB Atlas       │
                              │  • Atomic Calendars  │
                              │  • Call Logs         │
                              │  • Booking Integrity │
                              └──────────────────────┘

---

🛡️ Responsible AI & Safety Guarantees

1. Mandatory Identity Disclosure: Every receptionist greeting begins with explicit confirmation that the caller is speaking with an AI assistant.
2. Medical Advice Prohibition: The agent strictly declines medical inquiries, symptom diagnosis, and prescription questions, automatically initiating an urgent callback for a licensed doctor.
3. No Untested Version in Production: The Go-Live engine cryptographically verifies that the current branch has passed 100% of required blocker tests 3/3 before permitting deployment.
4. Instant Rollback: Every version exists on an immutable ElevenLabs agent branch; rolling back to a previous build takes under 2 seconds.

---

🚀 Getting Started Locally

Prerequisites

• Node.js: v20 or higher
• Package Manager: npm or pnpm
• MongoDB Atlas cluster URI
• ElevenLabs API Key
• Google Gemini API Key

Installation

1. Clone the repository:

   git clone [email protected]:anshvarshney1502/prototype.git munshi
   cd munshi
   

2. Install dependencies:

   npm install
   

3. Configure Environment Variables: Create a .env.local file in the root directory:

   # Database
   MONGODB_URI=mongodb+srv://<user>:<password>@cluster.mongodb.net/munshi?retryWrites=true&w=majority
   
   # AI Providers
   ELEVENLABS_API_KEY=your_elevenlabs_api_key
   ELEVENLABS_WEBHOOK_SECRET=your_webhook_signing_secret
   GEMINI_API_KEY=your_gemini_api_key
   
   # Tool Security
   TOOL_SECRET=your_internal_shared_secret_min_32_chars
   
   # Public Base URL (for webhooks & tools)
   PUBLIC_BASE_URL=https://your-deployment-domain.com
   

4. Run the Development Server:

   npm run dev
   

   Open http://localhost:3000 in your browser.

5. Run Production Build & Verification:

   npm run build
   npm run start
   

---

📁 Repository Structure

munshi/
├── docs/
│   └── images/              # High-resolution screenshots of the entire product flow
├── public/
│   ├── explainer/           # 60s English & Hindi MP4 explainer videos + poster
│   └── favicon.svg          # Munshi brand icon
├── scripts/
│   ├── live-booking.ts      # Live end-to-end booking verification test
│   ├── smoke-elevenlabs.ts  # ElevenLabs Agent & Testing API connectivity test
│   ├── test-access.sh       # Multi-tenant isolation and security test suite
│   ├── test-calendar.ts     # MongoDB double-booking prevention test
│   └── test-rollback.ts     # Instant branch rollback verification test
├── src/
│   ├── app/
│   │   ├── api/             # Next.js route handlers (clinics, versions, tools, runs, voices)
│   │   ├── c/[id]/          # Main workspace: Explain, Card, Practice, Fix, Go Live, Calls
│   │   ├── talk/[id]/       # Public customer-facing voice receptionist interface
│   │   ├── layout.tsx       # Root layout with font optimization & i18n provider
│   │   └── page.tsx         # Redesigned bilingual landing page
│   ├── components/
│   │   ├── landing/         # Landing hero, 60s dual-audio explainer player, practice demo
│   │   ├── workspace/       # Interview, JobCardStep, TrialDay, Coach, GoLive, Calls
│   │   ├── Builder.tsx      # Start-from-scratch or starter selector
│   │   ├── Icons.tsx        # Tailored SVG icon system
│   │   ├── Logo.tsx         # Brand mark and wordmark
│   │   └── Orb.tsx          # Three.js 3D audio-reactive orb visualizer
│   └── lib/
│       ├── calendar.ts      # Slot calculation & atomic booking engine
│       ├── compile.ts       # Job card to prompt compiler
│       ├── db.ts            # MongoDB Atlas connection manager & schemas
│       ├── elevenlabs.ts    # ElevenLabs SDK wrapper (Agents, Tests, Branches, Scribe)
│       ├── gemini.ts        # Google Gemini extraction & synthesis pipeline
│       ├── i18n.tsx         # Zero-overhead English & Hindi localization engine
│       └── jobcard.ts       # Strict Zod schema for receptionist rule cards
└── vercel.json              # Vercel deployment configuration

---

🏆 Hackathon Submission Details

• Event: Build Arena Hackathon (2026)
• Problem Statement: #2 — No-Code Voice AI Agent Builder (ElevenLabs)
• Team Name: DOT DOT COM
• Developer: Ansh Varshney
• Live Deployment: https://munshi-hackathon-prototype.vercel.app
• GitHub Repository: https://github.com/anshvarshney1502/prototype
• Explainer Videos: Included directly in the web player with English / Hindi live audio toggle (/explainer/munshi-en.mp4, /explainer/munshi-hi.mp4)

---

Made with ❤️ for small business owners who deserve a receptionist that never sleeps.