cadence.

your voice, preserved.

Cadence is a multi-agent AI platform that preserves your writing voice. It analyzes how you write — your rhythm, word choices, sentence structure, the things that make your writing yours — then generates new content that sounds exactly like you while passing AI detection.

DiamondHacks 2026 @ UCSD

---

what it does

Writing Studio — Upload past writing, describe what you need, and Cadence writes essays, papers, and reports that match your voice. Watch the full pipeline live: voice fingerprinting, draft streaming, parallel AI detection across ZeroGPT and Originality.ai, and iterative revision until the draft passes.

Communication Studio — Upload emails and messages. Cadence learns your messaging style — brevity, tone, emoji habits. A Browser Use agent reads your Gmail inbox and drafts replies in your voice.

Voice Interview — No writing samples? A 5-minute ElevenLabs ConversationalAI interview builds your voice profile from how you speak, think, and explain.

Fetch.ai Pipeline — The long-form writing pipeline runs as 5 autonomous agents on Agentverse, accessible through ASI:One chat. Attach your voice profile PDF, describe what you need, and the agents coordinate: digest → write → detect → revise.

---

architecture

frontend/          React 18 + Vite
backend/           FastAPI + WebSockets
cadence-fetchai/   Fetch.ai uAgents (5 agents on Agentverse)

the pipeline

voice fingerprint → claude writer → noise injection → browser use detectors → consensus
                         ↑                                                        │
                         └──── revise flagged sentences (max 5 iterations) ←──────┘

agents

agent	framework	what it does
Voice Analyst	Claude	Builds structured voice fingerprint from writing samples
Writer	Claude Sonnet	Generates and revises drafts guided by fingerprint
ZeroGPT Detector	Browser Use	Navigates zerogpt.com, scrapes AI score + flagged sentences
Originality Detector	Browser Use	Navigates originality.ai, scrapes AI score + flagged sentences
Inbox Agent	Browser Use	Reads Gmail inbox, sends replies through live browser
Interview Analyst	Claude	Converts voice interview transcript into fingerprint
Comm Writer	Claude	Drafts email replies matching communication voice

fetch.ai agents (ASI:One)

agent	port	role
Long-Form Orchestrator	8001	User-facing on ASI:One, coordinates pipeline
Profile Digester	8002	Parses voice profiles into fingerprints
Long-Form Writer	8003	Claude-powered draft generation + revision
ZeroGPT Detector	8004	Browser Use AI detection via zerogpt.com
Originality Detector	8005	Browser Use AI detection via originality.ai

---

humanization engine

Two-stage approach to making AI-generated text pass detection:

Stage 1 — Structural post-processing (_humanize_text):

• Replace AI sentence starters (Furthermore, Moreover) with "And" / "But"
• Inject comma splices, ellipsis on reflective sentences
• Drop Oxford commas randomly, strip em-dashes
• Break tricolon patterns (AI loves groups of three)

Stage 2 — Progressive noise injection (_inject_noise, intensity 1→5):

• Cyrillic homoglyphs (Latin a → Cyrillic а) — visually identical, breaks tokenization
• Zero-width spaces between words — invisible, disrupts token boundaries
• Double words, contraction swaps, comma drops — human-like errors

---

tech stack

layer	tech
Frontend	React, Framer Motion, Three.js (particle + dither shaders), jsPDF
Backend	FastAPI, WebSockets, Anthropic SDK, Browser Use Cloud SDK
Database	Supabase (Postgres, Auth, Storage)
Voice	ElevenLabs ConversationalAI
AI Detection	Browser Use → ZeroGPT, Originality.ai (no APIs — real browser scraping)
Multi-Agent	Fetch.ai uagents, Agentverse, ASI:One Chat Protocol
PDF	PyMuPDF (extraction), jsPDF (generation)

---

running locally

frontend

cd frontend
npm install
npm run dev

backend

cd backend
pip install -r requirements.txt
uvicorn main:app --reload

fetch.ai agents

cd cadence-fetchai
pip install -r requirements.txt
python3 run_all.py

environment variables

backend/.env

ANTHROPIC_API_KEY=
BROWSER_USE_API_KEY=
ELEVENLABS_API_KEY=
ELEVENLABS_AGENT_ID=
SUPABASE_URL=
SUPABASE_SERVICE_KEY=
GMAIL_PROFILE_ID=          # from setup_gmail_profile.py

cadence-fetchai/.env

ANTHROPIC_API_KEY=
BROWSER_USE_API_KEY=
AGENTVERSE_API_KEY=

---

voice profile export

The webapp generates structured PDFs that can be used with the Fetch.ai pipeline:

• writing.cadence.pdf — writing rules, metrics, signature phrases, exemplar passages
• comms.cadence.pdf — email style, response patterns, communication rules
• interview.cadence.pdf — personality, speech patterns, vocabulary, transcript

Attach to ASI:One chat → orchestrator downloads + extracts text → passes to writer as style guide.

---

hackathon tracks

The Scholar's Spellbook (Education) — Voice discovery as education. Students learn what makes their writing unique. The live pipeline is a real-time writing lesson.

Best Use of Browser Use — Two parallel Browser Use agents scraping AI detection sites with no APIs. Per-sentence flagging from live DOM elements. Gmail automation through real browser sessions.

Best Use of Fetch.ai — 5 agents on Agentverse with ASI:One Chat Protocol. Multi-agent coordination with write-detect-revise loop. PDF attachment handling via ResourceContent.

---

built by matthew kim for diamondhacks 2026.