The Perfect Pitch

Inspiration

90% of startups fail, not because they can't code, but because they are delusional about their market. Founders are trapped in a "Validation Vacuum": friends and family are too polite to give honest feedback ("Great idea!"), and real VCs are too busy to reply ("Pass.").

This creates a fatal gap: The first time a founder hears the hard truth is usually the day they run out of money.

We asked: What if we could simulate the brutality of a Tier-1 VC partner meeting? What if an AI could be "rude" enough to save your business?

What it does

Liquidator is a military-grade flight simulator for capital allocation. It is an Adversarial Voice Agent designed to break your business logic so you can rebuild it stronger.

The Ingest: You upload your Pitch Deck (PDF).

The Shadow Analysis: While you wait, our backend "Shadow Analyst" scans your deck, identifies your industry, hunts for historical failed competitors (e.g., "Webvan", "Quibi"), and compiles a "Kill Sheet" of trap questions.

The Interrogation: "The Liquidator" (our AI Voice Agent) calls you. It doesn't just chat; it attacks. It quotes your own revenue numbers back to you, interrupts buzzwords, and demands to know your unit economics.

The Verdict: After the call, you are graded. If—and only if—you score above 90/100, you unlock the "Golden Ticket" (a simulated referral to top investors).

How we built it

We built a multi-agent system combining the specific strengths of the hackathon partners:

LiquidMetal Raindrop (The Brain):

We used SmartBuckets to securely store and parse incoming pitch deck PDFs.

We leveraged SmartInference (Llama 3.1 70B) to run the "Shadow Analyst" agent. This isn't a standard summary; we engineered a specific prompt to find logical fallacies and missing metrics in the user's text.

Vultr (The Muscle):

The entire application (Node.js backend + React frontend) is containerized with Docker and hosted on Vultr App Platform.

We utilized Vultr's high-performance compute to ensure the PDF parsing happens instantly before the voice call starts.

ElevenLabs (The Voice):

We used the Conversational AI Widget with a custom "Liquidator" persona (High Stability, Low Latency).

The Technical Breakthrough: We built a dynamic data bridge using clientTools. The Agent calls Pitch to pull the specific data extracted by Raindrop (e.g., "Claimed Revenue: $0") and injects it into the conversation context in real-time.

Challenges we ran into

The biggest challenge was Latency vs. Context. We needed the Voice Agent to know the contents of the PDF immediately, but reading a whole deck takes time.

Solution: We built a "Hacking" UI animation that entertains the user while the Raindrop backend runs the deep inference. We also optimized the prompt to extract only "Ammunition" (Weaknesses) rather than summarizing the whole file, cutting processing time by 60%. And also we peovided other alternatives like direct pitch or text base deck

Accomplishments that we're proud of

The "Scary" Factor: The first time the AI interrupted us and said, "You claim to have $2M revenue, but your slide says 'Pre-Revenue'. Are you lying?" — we knew we had built something special.

True Integration: This isn't three separate APIs; it's a unified loop. Raindrop feeds the data that ElevenLabs speaks, hosted seamlessly on Vultr.

What we learned

We learned that Adversarial AI is a massive untapped market. Most AI is built to be "helpful assistants." Building an AI designed to be a "hostile simulator" requires completely different prompt engineering and architecture, but the value it provides (resilience training) is much higher.

What's next for The Perfect Pitch

B2B Pivot: We plan to sell this tool to Accelerators (like YC or Techstars) as an automated "First Round Screener" to filter thousands of applications automatically.

Video Analysis: Integrating vision to analyze the design of the slides, not just the text.