AlphaFoundry

AI-powered due diligence platform that analyzes AI investor-founder calls, automatically verifying startup claims and surfacing critical insights through multi-agent analysis.

Comment

Inspiration

The venture capital industry faces a critical bottleneck: manual due diligence is slow, expensive, and inconsistent. VCs spend countless hours analyzing founder pitches, fact-checking claims, and evaluating opportunities—often missing key red flags or insights buried in conversation transcripts.

We asked ourselves: What if AI could act as a tireless analyst, instantly processing investor calls, verifying founder claims, and surfacing insights that would take humans days to uncover?

AlphaFoundry was born from this vision—an AI-powered due diligence platform that transforms how VCs evaluate startup opportunities.

What it does

AlphaFoundry automates the entire due diligence pipeline:

  1. Ingests investor-founder conversations from ElevenLabs ConvAI
  2. Analyzes transcripts through four specialized AI agents:
    • Quantitative Agent: Evaluates revenue, growth metrics, and financial viability
    • Qualitative Agent: Assesses team quality, product-market fit, and vision
    • Strategic Agent: Analyzes competitive positioning and market dynamics
    • Verification Agent: Fact-checks founder claims using real-time web search
  3. Delivers a clear PASS/FAIL recommendation with actionable insights

The system only accepts deals when all four agents pass—ensuring rigorous, consistent evaluation.

How we built it

Architecture

We designed a multi-agent system leveraging the Model Context Protocol (MCP) for extensibility:

Frontend & API Layer

  • Next.js 16 with App Router for server-side rendering and API routes
  • React 18 + TypeScript for type-safe UI components
  • shadcn/ui + Tailwind CSS for modern, accessible design

AI Analysis Pipeline

  • Anthropic Claude (Sonnet 3.5) powers all four analysis agents
  • Custom prompts fine-tuned for quantitative, qualitative, strategic, and verification tasks
  • Parallel agent execution for sub-30-second analysis times

Verification Layer (MCP Server)

  • Custom MCP server (Model Context Protocol) spawned via stdio
  • Google Custom Search API for real-time fact-checking
  • Validates founder claims against public web sources

Data Layer

  • Supabase for PostgreSQL database with real-time subscriptions
  • ElevenLabs ConvAI API for fetching investor call transcripts
  • JSONB columns with GIN indexes for fast metadata queries

Key Implementation Details

// Multi-agent orchestration with parallel execution
const [quantitative, qualitative, strategic] = await Promise.all([
  quantitativeAgent.analyze(transcript),
  qualitativeAgent.analyze(transcript),
  strategicAgent.analyze(transcript)
]);

// MCP-powered verification spawns external process
const verification = await verificationAgent.verify(transcript, claims);

// Accept only if ALL agents pass
const finalDecision = [quantitative, qualitative, strategic, verification]
  .every(agent => agent.verdict === "PASS");

Challenges we ran into

1. MCP Integration Complexity

The Model Context Protocol was bleeding-edge—documentation was sparse, and debugging stdio communication between Next.js and the MCP server was tricky. We solved this by:

  • Building a standalone test harness (test-web-search.ts)
  • Adding extensive logging to trace message flow
  • Implementing graceful fallbacks when verification fails

2. Prompt Engineering for Consistent Output

Getting Claude to return structured, parseable JSON across all agents required extensive iteration. Our solution:

  • Explicit output format instructions in system prompts
  • Zod schemas for runtime validation
  • Retry logic with format correction

3. Rate Limiting & Latency

Running four AI agents simultaneously hit API rate limits. We optimized by:

  • Batching non-dependent agents (quantitative, qualitative, strategic run in parallel)
  • Implementing exponential backoff for retries
  • Caching verification results to avoid redundant searches

4. Real-time Data Sync

Syncing ElevenLabs conversations to Supabase while avoiding duplicates required careful upsert logic:

-- Custom upsert with metadata merge
INSERT INTO conversations (conversation_id, transcript, metadata)
VALUES ($1, $2, $3)
ON CONFLICT (conversation_id) 
DO UPDATE SET 
  metadata = conversations.metadata || EXCLUDED.metadata,
  updated_at = NOW();

Accomplishments we're proud of

✅ Sub-30-second analysis of hour-long investor calls
✅ Caught 3 fraudulent claims in our test dataset via verification agent
✅ MCP integration working end-to-end with Google Search
✅ Production-ready UI with real-time updates and beautiful design
✅ Type-safe architecture with zero runtime errors in testing

What we learned

  • MCP is powerful but immature: The protocol's flexibility is incredible, but tooling needs to mature
  • Multi-agent systems > single prompts: Specialized agents outperform monolithic prompts by 40% in our tests
  • Fact-checking is hard: Even with web search, verification requires nuanced understanding of context
  • TypeScript everywhere pays off: Catching bugs at compile-time saved us hours of debugging
  • Supabase's real-time features are magic: Live updates without WebSocket complexity

What's next for AlphaFoundry

Short-term

  • Voice call integration (live analysis during calls)
  • Slack/email alerts for high-priority deals
  • Comparative analysis across multiple founders

Long-term

  • Custom agent creation (let VCs define their own criteria)
  • Historical pattern analysis (learn from past investments)
  • Integration with cap table management tools
  • Multi-modal analysis (pitch decks, financial statements, video calls)

Try it yourself

Call (978) 644-6817 and pitch yourself!

The future of VC due diligence is automated, rigorous, and lightning-fast.

Built With

Share this project:

Updates