Skip to content

AI Pitch Studio

AI Pitch Studio is a dual-function pitch preparation and video editing platform designed specifically for Problem Statement Two: More Dimensions.

Code Video Cerebral Valley

Core trick

Argon writes FFmpeg edit scripts, and a separate critic instance scores each rendered video and feeds visual notes back for rewrites.

Tags

voiceagents

Description

AI Pitch Studio is a dual-function pitch preparation and video editing platform designed specifically for Problem Statement Two: More Dimensions. The platform consists of two main engines: 1. Live On-Device Coaching: A local rehearsal environment that uses Gemma to provide real-time nudges on pace, filler words, tone, and articulation. It pairs with a browser-based eye-contact model powered by MediaPipe FaceLandmarker. 2. Self-Correcting Video Editor: An autonomous editing loop powered by Gemini 4 Argon. Argon ingests raw footage and user goals, identifies key moments, and writes an FFmpeg script to execute the edit. A separate critic instance then watches the rendered video, scores it, and passes visual feedback back to Argon, which iteratively rewrites its code until the final render aligns perfectly with the goal.
GerardCapRivers/Argon_4_Hackathon★ 00 forksTypeScriptCreated Oct 25, 2026 · last push Oct 25, 2026

README excerpt

# AI Pitch Studio

Live investor coaching and self-correcting pitch video creation, built for the Gemini 4 Argon hackathon
(Problem Statement Two: *More Dimensions*).

1. **Live coaching on the device (Gemma).** Rehearse on camera and get real-time nudges on pace, filler words,
   articulation, tone and eye contact. Audio is transcribed locally (faster-whisper) and Gemma runs in Ollama.
   Eye contact is detected in the browser by a per-user model trained during a 6-second calibration.
   At the end, Gemini 4 Argon watches the recording together with the live feedback log and returns a score
   plus action items.
2. **A self-correcting editor (Gemini 4 Argon).** You give a goal, audience, style, format and footage.
   Argon finds key moments and you pick which to keep. Argon then **writes an ffmpeg script**, which is
   rendered. A separate **critic** instance watches the render and critiques it against the goal, Argon
   **revises its code**, and the critic **scores the new version**. The result is v1 → v2 with a code diff
   and before/after scores. Your own feedback triggers further revisions.

```
footage + goal ─▶ Argon: key moments ─▶ you pick ─▶ Argon writes edit code ─▶ render (sandbox)
                                                         ▲                        │
                                                         │                        ▼
                                          Argon revises code ◀── critic watches & critiques
```

## Layout

```
backen

Team