**⚔️ Click the arena above to try PitchFight AI.**
**⚔️ Click the arena above to try PitchFight AI.**
**Live Space:** [https://huggingface.co/spaces/build-small-hackathon/PITCHFIGHT_AI](https://huggingface.co/spaces/build-small-hackathon/PITCHFIGHT_AI)
**Build write-up:** [https://huggingface.co/blog/prakhar811/pitchfight-ai-practice-the-pitch-before-the-real-r](https://huggingface.co/blog/prakhar811/pitchfight-ai-practice-the-pitch-before-the-real-r)
**GitHub Repository:** [https://github.com/prakhar811/PitchFight/](https://github.com/prakhar811/PitchFight/)
---
## How it is built
PitchFight AI runs as a Hugging Face Gradio Space, but the user experience is built as a custom frontend rather than a default Gradio interface.
The frontend talks to backend routes in `app.py`. The backend handles pitch structuring, judge personas, battle state, voice mode, deal rounds, and scorecard generation. NVIDIA Nemotron is used as the core reasoning model for the AI judge.
```mermaid
flowchart LR
founder[Student Founder] --> ui[Custom PitchFight Frontend]
ui --> api[Gradio Backend / app.py]
api --> router[Model Router]
router --> nemotron[NVIDIA Nemotron API]
nemotron --> judge[Judge Questions + Feedback]
judge --> ui
api --> scorecard[Scorecard + Coaching]
scorecard --> ui
```
---
## System flow
1. The founder submits or records a startup pitch.
2. The backend structures the pitch into a founder briefing.
3. The founder chooses an opponent and difficulty mode.
4. The AI judge asks multi-round follow-up questions.
5. The founder answers under pressure.
6. The system evaluates clarity, confidence, defensibility, and traction.
7. The founder enters a deal-style round.
8. The final scorecard highlights strengths, weak points, and retry suggestions.
---
## Tech stack
| Layer | Technology |
| -------------------- | ----------------------------------- |
| Hosting | Hugging Face Spaces |
| Runtime | Gradio |
| Backend | Python, `app.py`, custom API routes |
| Frontend | Custom HTML, CSS, JavaScript |
| AI Judge | NVIDIA Nemotron |
| Voice Support | Nemotron Omni / voice pipeline |
| Audio Support | ffmpeg |
| Optional Persistence | MongoDB |
| Deployment Secrets | Hugging Face Space Secrets |
---
## NVIDIA usage
PitchFight AI uses NVIDIA Nemotron as the core judge and reasoning model.
The backend calls the NVIDIA API through:
```text
https://integrate.api.nvidia.com/v1
```
Primary model:
```text
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
```
The model is used for:
- founder brief structuring
- judge persona reasoning
- pitch battle follow-up questions
- deal round pressure
- final scorecard feedback
`NVIDIA_API_KEY` is stored only as a backend secret and is never exposed to frontend JavaScript.
---
## Built for the Build Small Hackathon
PitchFight AI was built for the Hugging Face **Build Small Hackathon**.
The goal was to create a focused AI product that feels genuinely useful through structured reasoning, judge personas, voice interaction, and a polished custom interface.
| Prize / Badge | Status | Why PitchFight AI qualifies |
| ------------------------- | --------- | --------------------------------------------------------------------------------- |
| **Backyard AI** | Submitted | Helps student founders practice for real-world pitch situations. |
| **NVIDIA Nemotron Quest** | Submitted | Uses NVIDIA Nemotron as the core reasoning model. |
| **Off Brand** | Submitted | Fully custom cinematic frontend, not default Gradio UI. |
| **Best Demo** | Submitted | Live Space, demo video, and blog write-up are linked. |
| **Field Notes** | Submitted | Hugging Face Blog and `FIELD_NOTES.md` explain the idea, build, and tech. |
| **Community Choice** | Eligible | Public Space is available for users to try and upvote. |
| **Judges' Wildcard** | Eligible | PitchFight AI combines founder tooling, coaching, voice, and pressure simulation. |
### Not claiming
- **OpenBMB / MiniCPM:** PitchFight AI uses NVIDIA Nemotron, not MiniCPM.
- **Modal:** Not claimed unless Modal is used in development or runtime.
- **Tiny Titan:** Not claimed because the primary model is not ≤4B.
- **Off the Grid:** Not claimed because the app uses the NVIDIA hosted API.
- **Fine-tuning:** Not claimed. Inference uses the hosted Nemotron API only.
---
## Submission links
| Item | Link |
| ----------------- | ---------------------------------- |
| Live Space | https://huggingface.co/spaces/build-small-hackathon/PITCHFIGHT_AI |
| Demo Video | https://www.youtube.com/watch?v=s4_BzIBhqxc |
| Hugging Face Blog | https://huggingface.co/blog/prakhar811/pitchfight-ai-practice-the-pitch-before-the-real-r |
| GitHub Repository | https://github.com/prakhar811/PitchFight/ |
| Field Notes | [`FIELD_NOTES.md`](FIELD_NOTES.md) |
---
## Hugging Face deployment
Required Space secret:
```bash
NVIDIA_API_KEY=
```
Optional variables:
```bash
APP_ENV=production
MAX_ROUNDS=6
DEFAULT_MODEL_MODE=premium_nvidia
NVIDIA_BASE_URL=https://integrate.api.nvidia.com/v1
NVIDIA_OMNI_MODEL=nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
ENABLE_DECK_CRITIQUE=true
ENABLE_DEAL_BATTLE=true
ENABLE_VOICE_MODE=true
MONGODB_URI=
```
Never put API keys in frontend code or commit real secrets to GitHub.
---
## Run locally
```bash
python -m venv venv
.\venv\Scripts\Activate.ps1
pip install -r requirements.txt
cp .env.example .env
python app.py
```
Open:
```bash
http://127.0.0.1:7860
```
Install `ffmpeg` locally if using voice mode.
---
## Project structure
```text
.
├── app.py Gradio app entrypoint and backend routes
├── requirements.txt Python dependencies
├── packages.txt System packages such as ffmpeg
├── .env.example Local environment variable template
├── frontend/ Custom HTML, CSS, JS, and UI assets
├── core/ Backend handlers, model router, clients
├── FIELD_NOTES.md Build notes for the Field Notes badge
└── README.md Space card and hackathon submission write-up
```
---
## Safety note
PitchFight AI is a practice and coaching tool. It does not replace real mentors, investors, or legal/financial advisors. It is designed to help founders sharpen their pitch before entering the real room.
---