The Defense Panel

One-Line Pitch

A multi-persona AI panel that watches you over WebRTC, reads your project document, and challenges you in real time, so you walk into your pitch, viva, or interview having already faced the hardest questions.

---

Summary

Founders, researchers, and job-seekers routinely fail high-stakes presentations not because they lack knowledge, but because they have never practiced under real adversarial pressure.

The Defense Panel puts you in front of two hyper-realistic Beyond Presence avatars — one interrogating your logic, one tracking your delivery — that have read your actual document and will interrupt you when you waffle or contradict yourself.

Sessions are automatically graded and stored in Supabase so you can track improvement over time. The product runs entirely in the browser via WebRTC with no plugins required.

---

Submission Details

Field	Details
Track	Best Use of Beyond Presence
Beyond Presence Integration	Two live WebRTC avatar agents run in parallel per session, created dynamically via the Beyond Presence REST API with GPT-4o-generated system prompts derived from the user's uploaded document
Demo URL	https://defense-panel-two.vercel.app
Tech Stack	React 18 · Vite 6 · Vercel Functions · OpenAI GPT-4o · Beyond Presence · LiveKit · Supabase · Dodo Payments

---

Problem Statement

The Problem

Founders preparing investor pitches, postgraduate researchers approaching their viva voce, and candidates facing technical interviews share a common blind spot: they rehearse their material but never rehearse under adversarial, real-time pressure.

When a VC says:

"Your CAC payback period doesn't add up"

or an examiner says:

"You contradict your own methodology on page 47"

most people freeze — not because they lack the answer, but because they have never been asked that way before.

If left unsolved:

• Founders lose funding rounds
• PhD candidates fail vivas
• Qualified candidates fail interviews they were capable of passing

Why It Matters

• Startup pitch success rates for first-time founders hover around 1–3%
• Viva failure and major correction rates exceed 15%
• Technical interview pass rates at top-tier companies are below 20%

Existing practice methods are:

• Expensive
• Difficult to schedule
• Rarely adversarial enough

Root Cause

The issue is not a lack of preparation tools.

There are:

• Flashcard apps
• Slide coaches
• Interview question banks

But none provide:

• Persistent contextual memory
• Real-time adversarial interruption
• Awareness of the uploaded document
• Dynamic challenge generation

---

Proposed Solution

What the Product Does

Users upload:

• Pitch decks
• Theses
• CVs

The backend:

1. Extracts document text
2. Sends it to GPT-4o
3. Generates a structured Context Matrix

Two Beyond Presence avatars are then created:

• Interrogator → challenges logic
• Ah-Counter → tracks delivery quality

Users enter a live WebRTC arena where both avatars:

• See and hear them
• Interrupt weak answers
• React contextually

At the end:

• GPT-4o grades the transcript
• Results are stored in Supabase

---

Key Features

Feature	User Need Addressed
Document-aware AI personas	Catch contradictions specific to the user's content
Live WebRTC barge-in interruption	Simulates real pressure
Dual avatar panel	Separates logic critique from delivery feedback
Three simulation modes	Startup Pitch / Viva / Interview
Difficulty levels	Friendly / Standard / Brutal
Auto-grading & transcript persistence	Tracks improvement over time
Inline document preview	Mirrors real presentation setup
Vision mode toggle	Optional body-language commentary

---

Scope

In Scope

• Three simulation modes
• Dual Beyond Presence agents
• LiveKit real-time media bridge
• GPT-4o grading pipeline
• Supabase authentication and session history
• PDF / DOCX / PPTX extraction
• Dodo Payments integration
• Public deployment

Out of Scope

• Native mobile apps
• Enterprise SSO
• Custom avatar creation
• Voice cloning
• Real-time transcript display

---

Functional Requirements

Must Have

• Upload PDF/DOCX/PPTX
• GPT-generated Context Matrix within 10 seconds
• Live avatar session
• Context-aware interruptions
• Simulation mode selection
• Auto-generated grading
• Session persistence
• Email/password and Google OAuth authentication
• Session bookmarking
• Difficulty and session-length configuration

Should Have

• Second avatar observer
• Practice goal selection
• Inline document preview
• Vision mode

Could Have

• Native mobile apps
• Real-time transcript display
• Collaborative sessions
• Voice/avatar customization
• Email reports

---

Non-Functional Requirements

Performance

• Document processing < 10 seconds
• Agent creation < 30 seconds
• Grading < 60 seconds

Reliability

• Structured error handling
• Frontend toast notifications
• Graceful extraction fallbacks
• Automatic agent cleanup

Usability

• Single-page application flow
• Minimal controls during session
• Chromium browser recommendation

Scalability

• Stateless API sessions
• Supabase RLS multi-tenancy
• Serverless auto-scaling

---

Technical Architecture

Stack

Layer	Technology
Frontend	React 18 + Vite 6 + Tailwind CSS
Realtime Media	Beyond Presence + LiveKit
Backend	Vercel Functions
Parsing	pdf-parse, mammoth, jszip
AI	OpenAI GPT-4o
Database/Auth	Supabase
Payments	Dodo Payments
Hosting	Vercel

---

Core Data Flow

1. User uploads document
2. Backend extracts text
3. GPT-4o generates Context Matrix
4. User configures session
5. Backend creates Beyond Presence agents
6. Browser launches WebRTC arena
7. Session runs with live interruptions
8. User ends session
9. GPT grades transcript
10. Results saved to Supabase

---

AI Integration

GPT-4o performs:

• Context extraction
• Persona generation
• Transcript grading

Outputs:

• System prompt
• Greeting
• Role objectives
• Conversation structure
• Starting script
• Document summary

Grading returns:

• Overall score
• Filler word count
• Critical feedback

---

Known Technical Limitations

• WebRTC interruption latency may vary
• PPTX extraction weak for image-heavy decks
• Dodo sidecar is single-process
• Transcript retrieval may occasionally fail

---

Security

Authentication

• Supabase Auth
• JWT verification
• Row Level Security

Data Handling

• Documents processed in-memory only
• No payment data stored

Secret Management

Secrets stored only server-side:

• OPENAI_API_KEY
• BEYOND_PRESENCE_API_KEY
• SUPABASE_SERVICE_ROLE_KEY
• DODO_PAYMENTS_API_KEY

Validation

• MIME validation
• 10MB upload limit
• JSON validation for AI outputs

---

User Stories

Startup Founder

I want to upload my pitch deck and be challenged by an AI investor panel.

PhD Researcher

I want to simulate my viva with AI examiners.

Job Candidate

I want to practice technical interviews with AI recruiters.

Returning User

I want to track my progress across sessions.

---

Startup Pitch Use Case

1. User uploads deck
2. AI processes document
3. User configures session
4. Arena launches
5. AI interrupts weak reasoning
6. Session continues dynamically
7. Grading begins
8. Feedback dashboard appears
9. Session saved to history

---

Edge Cases

• Image-heavy PPTX uploads
• Blocked microphone permissions
• Very short sessions
• User leaves session early

---

Target Users

Primary Users

• Startup founders
• PhD researchers
• Technical interview candidates

Market Opportunity

Targets:

• Viva preparation
• Startup fundraising
• Technical interview coaching

---

Competitive Landscape

Competitor	Weakness
Yoodli / Poised	Delivery analysis only
ChatGPT / Claude	No real-time presence
Human mock panels	Expensive and hard to schedule

---

Business Model

Freemium SaaS

Free Tier

• 3 sessions/month
• Standard difficulty
• Single avatar

Pro Tier

• Unlimited sessions
• Dual avatars
• Brutal difficulty
• Full history

Pricing

• £12–£18/month
• Annual discount
• Institutional plans

---

Go-To-Market Strategy

• University outreach
• Startup accelerator partnerships
• Short-form social content
• Sponsor collaborations

---

Roadmap

Built

• Dual avatars
• GPT grading
• Session history
• Payments integration

0–3 Months

• Real-time transcript display
• Better PPTX extraction
• Progress analytics

1 Year

• Voice cloning
• Institutional dashboards
• Native mobile apps

---

Why This Project Leads the Track

Technical Edge

• Dynamic document-generated prompts
• Dual Beyond Presence agents
• Real-time contextual interruptions
• Full AI orchestration lifecycle

Problem-Solution Fit

Every feature directly supports adversarial practice under pressure.

Execution Quality

• Public deployment
• Working production stack
• End-to-end integration
• Payments + persistence included

Real-World Potential

Applicable to:

• Startups
• Academia
• Technical hiring

The product has:

• Clear monetization
• Viral demo moments
• Institutional expansion potential

---

Team & Roles

Member	Responsibility
Kaveesha Ginodh	Team leadership, coordination, debugging
Dulaksha Dulan	UI/UX and frontend
Thesaru Praneeth	Backend and AI integration
Prabuddha Paranawithana	Payments and database architecture

---

Why This Team

The team covers:

• Frontend
• Backend
• AI
• Infrastructure
• Payments
• Deployment

All integrated into a working production-grade system within 24 hours.

Setup

See REQUIREMENTS.md for full clone-and-run instructions (Node 22.13+, all env vars, Supabase SQL, dual-terminal API setup, and troubleshooting).

Quick start:

npm install
cp .env.example .env

Fill in .env from .env.example (Supabase, OpenAI, Beyond Presence). Run supabase/pitch_sessions.sql in Supabase (and optionally supabase/pitch_sessions_extended.sql for dashboard/debrief enrichment fields). Enable Email auth. For briefing deploy and live agents, also run npx vercel dev --listen 3000 in a second terminal (see REQUIREMENTS.md).

Run

npm run dev

Flow

1. Auth — Login (email/password) or Register (username, email, password)
2. Lobby — Select Startup Pitch, Academic Viva, or Technical Interview
3. Briefing — Upload mode-specific PDF, DOCX, or PPTX context
4. Arena — Startup Pitch: multi-format presentation preview + presenter PIP + panelists; other modes: full webcam + panelists
5. Debrief — Performance analytics (mock data until API is wired)

Build

npm run build

Deploy to Vercel

This repo is configured for one-click deployment as a Vite SPA with serverless functions in /api.

1. Push to GitHub / GitLab / Bitbucket

git push

2. Import the project in Vercel

• Vercel will auto-detect the Vite framework (also pinned in vercel.json).
• Build command: npm run build · Output directory: dist · Install: npm install.
• Node runtime is pinned to >=20 via package.json#engines.

3. Set environment variables (Project Settings -> Environment Variables)

Required for all environments (Production / Preview / Development):

Variable	Used by
VITE_SUPABASE_URL	Frontend Supabase client
VITE_SUPABASE_ANON_KEY	Frontend Supabase client
SUPABASE_URL	/api/* serverless
SUPABASE_ANON_KEY	/api/end-session, /api/sessions (JWT verification)
SUPABASE_SERVICE_ROLE_KEY	/api/end-session, /api/sessions (admin inserts)
OPENAI_API_KEY	/api/process-document, /api/end-session
BEYOND_PRESENCE_API_KEY	/api/start-session, /api/start-call, /api/end-session
BEY_AVATAR_ID	/api/start-session (Beyond Presence avatar persona)
DODO_PAYMENTS_API_KEY	/api/create-checkout (Dodo Payments secret key)
DODO_API_URL	/api/create-checkout (e.g. https://test.dodopayments.com or https://live.dodopayments.com)
PRODUCT_ID	/api/create-checkout (default Dodo product)

Optional:

Variable	Default	Purpose
BEY_CHAT_EMBED_ORIGIN	https://bey.chat	Override the iframe-fallback embed origin
BEY_AVATAR_ID2	unset	Second Beyond avatar: silent panelist agent + LiveKit slot 2 in the arena
VITE_USE_API_SESSIONS	unset	Force the frontend to use /api/sessions instead of direct Supabase reads. Production already uses /api/* automatically.
FRONTEND_URL	request origin	Fallback return_url host for /api/create-checkout after payment
VITE_DODO_API_URL	unset	Point Pricing checkout at the standalone dodo/ Express gateway instead of /api/create-checkout

Never expose OPENAI_API_KEY, BEYOND_PRESENCE_API_KEY, or SUPABASE_SERVICE_ROLE_KEY to the frontend — only the VITE_-prefixed pair is bundled into the client.

4. Run Supabase setup

In the Supabase SQL editor, run:

1. supabase/pitch_sessions.sql
2. supabase/pitch_sessions_extended.sql (debrief scores)
3. supabase/pitch_sessions_transcript.sql (Beyond transcript columns)

Enable Email auth in Supabase, and add your Vercel domain to the Site URL / redirect allowlist.

5. Function limits

vercel.json bumps memory and maxDuration for the AI endpoints (process-document, end-session) so OpenAI + Beyond Presence calls have time to complete. On the Hobby plan max duration is capped at 60s — the values in vercel.json stay within that limit.