Problem

Most founders walk into a VC room having only ever practiced on people who wanted them to succeed. That's rehearsal with a safety net — not real preparation. The YC standard demands 20–30 sessions of hostile, deliberate pressure before a real pitch. That caliber of preparation is inaccessible to 99% of founders who don't have warm introductions to seasoned investors willing to tear their deck apart.

Solution

Built PitchBuddy — a full-stack AI pitch coaching platform that simulates a real investor conversation with no safety net. The core engine runs on Gemini 2.5 Flash in a dual-role loop: first responding as a hostile VC who challenges assumptions, demands evidence, and attacks weak logic — then immediately switching to structured coach mode, explaining what flopped and how to fix it before the next exchange.

Founders select from 5 investor personas (YC Partner, Angel, Series A VC, Corporate VC, Impact Investor), each with distinct behavioral prompts and pressure styles, and a funding stage (Idea → Pre-seed → Seed → Series A+) that scales the AI's expectations and rigor accordingly.

LangChain manages full multi-turn conversation state via thread IDs. Firebase Firestore persists session history across conversations so the AI retains context from previous pitch attempts — simulating a real investor who remembers what you said last time.

Deployed on Google Cloud Run with Secret Manager for credential management and Cloud Build + Container Registry for a full CI/CD pipeline.

Impact

• Built and deployed solo in under 9 hours at GDG Nuremberg hackathon
• Placed 2nd of 16 teams by audience vote
• Accurate dual-role AI coaching after every exchange with no post-processing delay
• Persistent cross-session memory enabling progressive pitch improvement over multiple sessions
• Full GCP production deployment with zero hardcoded credentials

Tech Stack

Streamlit, Gemini 2.5 Flash, LangChain, Firebase Firestore, Google Cloud Run, Google Secret Manager, Google Cloud Build, Google Container Registry.