DEMO_VIDEO_PITCH.md

# DiffSense Demo Video Pitch

## 20-Second Selling Pitch

DiffSense is a private, local-first pull request reviewer for teams that cannot send proprietary code to cloud review bots. Paste a diff or public GitHub PR URL, click **Review diff**, and get inline severity-tagged findings plus structured JSON. The deterministic reviewer works immediately, while optional bridges for Mellum, Nemotron, Tiny Titan, MiniCPM-V, and Modal are ready for hosted providers or local checkpoints under `/data`.

## 60-Second Demo Script

Hi, this is DiffSense, our Build Small hackathon project.

The problem is that AI code review is useful, but most review bots require sending private source code to a hosted SaaS. That does not work for security-sensitive teams, regulated teams, or unreleased products.

DiffSense is a local-first alternative. On the left, I configure optional model passes: Mellum for summaries, Nemotron for routing, Tiny Titan for a lightweight checker, MiniCPM-V for screenshots and diagrams, and a Modal bridge for hosted inference.

In the center, I paste a public GitHub PR URL or a unified diff. I can also attach PR screenshots or diagrams for the vision path. Now I click **Review diff**.

The app fetches or parses the diff, runs a deterministic review engine, and returns a summary. The model runtime panel shows that `/data` is mounted and writable, with persistent checkpoint slots ready for local model weights.

On the right, DiffSense renders the detailed review as an inline diff: file headers, hunk headers, line numbers, severity badges, comments, and suggested fixes. Under that, it exposes structured JSON so the output can be copied into PR automation.

The key design choice is reliability. The deterministic review path always works, and model bridges enhance it when OAuth, provider routes, Modal, or local checkpoints are available.

So the product is useful now, private by default, and ready for small-model local inference.

## User Flow Checklist

1. Show the title and tagline: private, offline-first PR review.
2. Point to the sidebar model toggles.
3. Paste or keep the public PR URL in the center input.
4. Upload an image if you want to show the MiniCPM-V path.
5. Click **Review diff**.
6. Read the summary and model runtime status.
7. Move to the right pane and show the inline review.
8. Scroll to the structured JSON.
9. Close with the privacy and reliability point.

## One-Line Close

DiffSense turns a diff into a review artifact, not a chat transcript: private by default, useful without a GPU, and ready for local small-model checkpoints.