Startup Copilot

Your AI co-founding team: from raw idea to investor-ready pitch in minutes. Five specialized agents orchestrate market research, assets, and live voice pitching—powered by Amazon Nova.

Comment

Inspiration

The project was inspired by the "Founder’s Wall" — the overwhelming cognitive load early-stage entrepreneurs face when trying to be researchers, designers, and spokespeople all at once.

I wanted to move beyond simple chatbots and build a Multi-Agent Ecosystem that acts as a true co-founder, providing institutional-grade strategy and safe environments for high-stakes practice.

What it does

A founder uploads a document about their startup. From that moment, five coordinated AI agents go to work.

Market Intelligence

Produces a full market report including:

  • TAM/SAM analysis
  • Competitor mapping
  • Interactive charts

Asset Forge

Generates:

  • A professional pitch deck
  • AI-designed slide images
  • A cinematic pitch reel video with narrated voiceover

VC Scout

Finds investors whose thesis, stage, and sector match the venture.

Code Lab

Creates a technical blueprint and starter code scaffold.

Virtual Tank

A real-time voice pitch simulation where three AI sharks with distinct personalities challenge the founder live using bidirectional speech-to-speech streaming.

There is no scripted Q&A — just real conversational pushback.

How I built it

Frontend

  • Next.js
  • Tailwind CSS
  • Recharts for data visualization

Backend

  • FastAPI (Python) orchestrating all agents through a shared VentureDNA workspace

The workspace is a structured representation of the founder's:

  • Problem
  • Solution
  • Market

These are extracted from the uploaded document.

Amazon Nova models used

Model Purpose
Nova Pro Document ingestion (Venture DNA extraction), pitch deck narrative, code blueprints
Nova 2 Lite Market analysis, VC matching, Virtual Tank logic
Nova 2 Sonic Bidirectional speech-to-speech for Virtual Tank (real-time voice)
Nova Canvas Slide images, architecture diagrams
Nova Reel Cinematic pitch reel video generation
Nova Multimodal Embeddings Venture vector representation for semantic search

Supporting AWS services

  • Amazon Polly – per-shark TTS voices
  • S3 – pitch reel storage
  • Bedrock – unified API for all Nova calls

Architecture Pattern

Agentic AI

Each orb is an autonomous agent that:

  1. Receives venture context
  2. Calls the optimal Nova model for its task
  3. Writes results back to the shared workspace

Agents can run:

  • Independently
  • In sequence

The workspace acts as shared memory.

Challenges I ran into

Nova Sonic Bidirectional Streaming

The Virtual Tank required true real-time conversation, not request-response.

Key challenges:

  • Wiring Nova Sonic’s bidirectional stream through a FastAPI WebSocket
  • Handling barge-in (founder interrupts a shark mid-sentence)
  • Routing audio to three different Polly voices
  • Maintaining low conversational latency

Achieving real-time responsiveness required careful async orchestration.

Nova Reel Compositing Pipeline

Nova Reel generates silent video asynchronously to S3.

We built a full post-processing pipeline:

  1. Polly synthesizes narration
  2. FFmpeg composites video + audio + burned-in subtitles
  3. Final MP4 uploaded with a presigned URL

Coordinating:

  • async job polling
  • audio timing
  • subtitle synchronization

across three separate services was non-trivial.

Keeping Six Models Coherent

When five agents call different models, maintaining consistent venture context is critical.

A pitch deck that contradicts the market report destroys credibility.

The solution:

Shared VentureDNA Workspace Pattern

  • Extract once with Nova Pro
  • Share context everywhere

Web Grounding Restrictions

Our hackathon account's organization SCP denied bedrock:InvokeTool, which blocked Nova's built-in web grounding.

We built a custom web_search tool using DuckDuckGo that Nova invokes as a tool call.

This allowed:

  • Market Intelligence
  • VC Scout

to access live web data without native grounding permissions.

Accomplishments that we're proud of

Six Nova Models Working as One Team

I integrated:

  • Nova Pro
  • Nova 2 Lite
  • Nova 2 Sonic
  • Nova Canvas
  • Nova Reel
  • Nova Multimodal Embeddings

into a single coherent application, each model doing what it does best while sharing the same venture context.

Real-time Voice Pitch Practice

The Virtual Tank uses Nova 2 Sonic bidirectional streaming to create a genuinely interactive pitch simulation.

Three AI sharks with distinct personalities:

  • interrupt
  • challenge
  • push back

in real time.

It feels like a panel, not a chatbot.

Achieving sub-second conversational latency through: WebSocket → Sonic → Polly

was the hardest technical win.

One Document In, Everything Out

A founder uploads one PDF and receives:

  • Market report
  • Pitch deck with AI-generated images
  • Cinematic pitch reel video with narrated voiceover
  • Curated investor shortlist
  • Technical blueprint
  • Finance critique
  • Unlimited pitch practice sessions

The journey from "I have an idea" → "I'm investor-ready" happens in one sitting.

What I learned

Orchestration Over Scale

The most effective agentic systems don’t use the largest model for everything.

Matching tasks to the right model (e.g., low-latency voice vs deep reasoning) creates a far better user experience.

Shared Memory is Key

Agents must share context to be effective.

By building a unified Mission Graph, discoveries made by the Market Intelligence agent become immediately known by the AI Sharks in the Virtual Tank.

What's next for Startup Copilot

Virtual Tank Scoring and Coaching

After each pitch session, generate a structured scorecard:

  • Clarity
  • Confidence
  • Handling tough questions
  • Pacing

Track improvement across sessions so founders can quantitatively measure progress.

RAG over the Venture Workspace

Using Nova Multimodal Embeddings, enable semantic search across all generated outputs.

Example queries:

  • "Find the slide that addresses our competitive advantage."
  • "What did the sharks push back on hardest?"

Adversarial Memory for the Virtual Tank

Future AI Sharks will use Venture DNA to remember weaknesses found during Market Intelligence.

Example:

  • If research identifies a competitor gap, sharks will press the founder on that exact weakness during the pitch.

Stateful Agentic Synergy

The goal is full agent collaboration.

For example:

When the Market Intelligence agent finds new data, the Asset Forge automatically updates the pitch deck's market slide without requiring manual resync.

Built With

Share this project:

Updates