PitchBuddy

AI-powered startup pitch simulator. Practice your pitch against a skeptical VC, get coached in real time, and sharpen your story before the real meeting.

---

Overview

PitchBuddy combines two AI roles in every response:

• [VC] Asks one sharp, targeted question aimed at your weakest assumption
• [COACH] Tells you what landed, what didn't, and one concrete improvement

Select your investor type and funding stage before you start — the AI adjusts its behavior, scrutiny level, and focus areas accordingly. Conversations are saved to Firestore and can be resumed within the same browser session.

---

Features

• 5 investor personas — Angel, Seed VC, Series A VC, YC Partner, Corporate VC, each with distinct behavior and priorities
• 4 founder stages — Idea through Series A+, which controls how aggressive the AI's questioning gets
• Persistent chat history — Conversations saved to Firestore and restored with the original persona and stage (scoped to the browser session — closing the tab starts fresh)
• Suggestion cards — One-click prompts to drill into TAM, business model, defensibility, traction, funding ask
• Zero friction — No account creation; each browser session is automatically isolated

---

Tech Stack

Layer	Technology	Role
UI	Streamlit	App framework, chat interface, custom CSS
LLM	Gemini 2.5 Flash	AI responses via LangChain
Database	Firebase Firestore	Conversation persistence, user-scoped
Session auth	Anonymous UUID	Per-browser-tab isolation, no login required
Retry logic	Tenacity	Handles Gemini 429 / 502 / 503 errors
Containerization	Docker	Reproducible builds
Deployment	Google Cloud Run	Serverless, scales to zero
CI/CD	Google Cloud Build	Auto-deploy on push
Secrets	Google Secret Manager	API keys and credentials at runtime

---

Local Setup

Prerequisites: Python 3.11+, a Google AI Studio API key, Firebase project with Firestore enabled.

# Clone and enter
git clone https://github.com/gmMustafa/PitchBuddy.git
cd pitchbuddy

# Create virtualenv
python -m venv .venv

# macOS / Linux
source .venv/bin/activate
# Windows
.venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Set up environment
cp .env.example .env
# Open .env and paste your GOOGLE_API_KEY

# Add Firebase service account
# Firebase Console → Project Settings → Service Accounts → Generate new private key
# Save the downloaded file as: firebase-credentials.json

# Run
streamlit run app.py

App runs at http://localhost:8501.

---

Project Structure

pitchbuddy/
├── app.py                    # Entry point — layout, routing, session handling
├── src/
│   ├── config.py             # System prompt, persona prompts, stages, UI copy
│   ├── ai/
│   │   ├── agent.py          # LangChain chain, dynamic system prompt assembly
│   │   └── clients.py        # Gemini LLM client (Streamlit-cached)
│   ├── db/
│   │   └── firestore.py      # Firestore read/write, user-scoped collections
│   ├── ui/
│   │   ├── auth.py           # Anonymous UUID session isolation
│   │   ├── chat.py           # Welcome screen, chat history rendering
│   │   ├── input_bar.py      # Message input and submission
│   │   ├── session.py        # Message processing, conversation loading
│   │   ├── sidebar.py        # Persona/stage selectors, conversation list
│   │   └── styles.py         # Full light theme CSS
│   └── utils/
│       └── retry.py          # Tenacity retry decorator for API calls
├── Dockerfile
├── cloudbuild.yaml
├── requirements.txt
└── .env.example

---

Deployment

Full step-by-step guide: reproducibility.md

Covers Firebase setup, GCP project config, Secret Manager, Cloud Build trigger, and Cloud Run deploy flags.

---

Design Decisions

No login required. Each Streamlit browser session gets a UUID on first load. All Firestore writes are scoped to users/{uuid}/conversations/. Sessions are private by construction — there is no shared state between tabs or users.

Persona injected per call, not stored in history. The selected investor persona and founder stage are appended to the system prompt on every API call. This means switching personas mid-conversation takes effect immediately, and the full behavioral instructions are never diluted by growing chat history.

thinking_budget=0. Gemini 2.5 Flash supports extended thinking which can add 30–120s of latency. Disabled here — the base model handles pitch coaching well, and fast back-and-forth is more valuable than extra reasoning depth.

Two-phase rerun for conversation switching. Loading a saved conversation needs to restore persona and stage into session state before the sidebar selectboxes render. A direct call would raise a StreamlitAPIException. Instead, clicking a conversation sets a _load_conv_id key and reruns; on the next run, the app processes it at the top before any widgets are instantiated.