*** Begin Reindexed PRD*

Kaizen — Product Requirements Document (PRD)

Version: 0.1 — Hackathon MVP Date: 2025-10-15

Table of contents

1. Objective & Scope
2. Problem Statement
3. Vision & Value Proposition
4. Target Users & Personas
5. System Overview
6. MVP Features (user stories + acceptance criteria)
7. Architecture & Key Components
8. Tech Stack & Non-functional Requirements
9. Roadmap & Milestones
10. Success Metrics & Risks
11. Resources

---

1. Objective & Scope

Objective: build a privacy-first Chrome extension that turns Chrome into a locally-run AI co-pilot for browsing (Kaizen). The scope for the hackathon MVP (20 days) focuses on local summarization, basic tab automation, voice/text commands, and a minimal knowledge graph.

In scope for MVP:

• Voice and text command interface (popup)
• Tab grouping and management
• Local summarization for active tab / grouped tabs
• Lightweight RAG knowledge graph visualization (side-panel)
• Doomscroll detection with behavioral nudge
• Privacy & consent summarizer

Out of scope for MVP:

• Cloud sync / account integration
• Full-scale offline LLM training
• Advanced multi-device sync

---

2. Problem Statement

People hesitate to use AI browser features due to privacy, lack of control, and latency from cloud inference. Kaizen solves this by providing an on-device, privacy-focused assistant using Gemini Nano and Chrome-built-in AI APIs.

---

3. Vision & Value Proposition

Kaizen transforms Chrome into a context-aware assistant that helps research, automates repetitive browsing tasks, and protects users from unsafe content — all without sending their browsing data to remote servers.

Key value props:

• Privacy-first local AI
• Context-aware automation (tabs, groups, navigation)
• Lightweight knowledge graph to surface relationships

---

4. Target Users & Personas

Primary users:

• Researchers and knowledge workers who open many tabs and want to synthesize findings quickly
• Power users who automate workflows (tab management, autofill)

Secondary users:

• Privacy-conscious users who want AI assistance without cloud data sharing

---

5. System Overview

High-level flow:

• User issues a command (voice or text) in popup → AI Command Layer parses intent → Background/AI service performs actions (tabs, summarize, build graph) → UI updates in popup/side-panel.

Core runtime boundaries:

• UI (popup/side-panel) — React + Vite dev server for development
• Background service worker — message routing and tabs API
• Content scripts — page-level extraction and behavioral monitoring
• Local AI core — Gemini Nano via Chrome Built-in AI APIs (production)

---

6. MVP Features (User stories + Acceptance Criteria)

6.1 Voice & Text Command Interface

• Story: As a user, I want to tell Kaizen to "summarize this tab" so I get a short summary without leaving the page.
• Acceptance: Popup accepts text/voice input, returns a 2–4 sentence summary for the active tab.

6.2 Tab Grouping & Management

• Story: As a researcher, I want Kaizen to group tabs related to a topic.
• Acceptance: Kaizen can create a tab group and move selected tabs into it via a command.

6.3 Local Summarization & RAG

• Story: As a user, I want summaries of a tab group and a lightweight knowledge graph of extracted concepts.
• Acceptance: Kaizen produces a group summary and a small graph (nodes/edges) for the session.

6.4 Doomscroll Detection & Nudge

• Story: As a user, I want Kaizen to detect extended scrolling and offer a summary or a break.
• Acceptance: After N minutes of continuous scrolling on the same domain, Kaizen displays a nudge in the popup.

6.5 Privacy & Consent Summarizer

• Story: As a user, I want Kaizen to detect consent dialogs or policy text and provide a short risk summary.
• Acceptance: On page load, Kaizen flags major consent text and provides a 1–2 line summary.

---

7. Architecture & Key Components

7.1 Frontend

• Popup (voice/text input, actions)
• Side-panel (knowledge graph visualization)

7.2 Background (API Gateway)

• Routes messages, manages state, executes tab actions

7.3 Content Scripts

• Extract page text, detect doomscrolling and consent banners

7.4 Local AI (Gemini Nano)

• Responsible for summarization, rewriting, and graph embedding generation

7.5 Storage

• Chrome storage / IndexedDB for session graphs and local preferences

---

8. Tech Stack & Non-functional Requirements

Tech choices:

• Frontend: React + Vite
• Graph: Cytoscape.js or D3
• Storage: IndexedDB / chrome.storage.local
• Dev: pnpm workspaces, turbo (optional)

Non-functional:

• Privacy: no network calls for content by default
• Performance: local inference should not block UI; use web workers where appropriate
• Compatibility: Chrome stable supporting Chrome Built-in AI APIs (note feature flags may be required)

---

9. Roadmap & Milestones (Hackathon)

Day 1–3: scaffold, popup UI, background routing Day 4–8: local summarization + content extraction Day 9–13: tab grouping + side-panel graph UI Day 14–18: doomscroll detection + privacy summarizer Day 19–20: polish, testing, demo prep

---

10. Success Metrics & Risks

Success Metrics (for MVP):

• Completion of 6 core features above
• Demoable local summarization latency < 2s (for short pages)
• User flow time-to-summarize < 10s end-to-end

Top risks:

• Gemini Nano availability and compatibility across Chrome versions
• Local performance and memory usage
• Permission and manifest restrictions for side panel APIs

Mitigations:

• Fall back to lighter summarization methods; use web workers; document required Chrome flags

---

11. Resources

• https://developer.chrome.com/docs/extensions/get-started
• https://github.com/Jonghakseo/chrome-extension-boilerplate-react-vite
• https://github.com/nanobrowser/nanobrowser
• https://github.com/lxieyang/chrome-extension-boilerplate-react

---

End reindexed PRD