voice-os

Six-axis voice scoring. Dual-persona routing. QA gates that enforce stylistic fidelity before a draft ever reaches a human. Built on Claude. Calibrated on 6.9M+ words of executive communication.

This is the engine behind the Voice DNA RAG pipeline I shipped at Google xGE — a system that functions as a digital twin for VP-level communications. It cut drafting latency by 90% and holds 99% stylistic fidelity across production volume.

---

What it does

Ingests a voice corpus, builds a scored representation across six stylistic axes, and routes drafts through dual personas (generative + adversarial) before a QA gate decides whether output clears or cycles back. The result is drafts that sound like the person they're supposed to sound like — not a generic LLM.

---

Why it matters

Most "voice matching" is prompt engineering with a few examples. This is a calibrated scoring system. The six axes catch what vibes-based prompting misses: rhetorical pace, risk tolerance, sentence rhythm, escalation pattern, hedging behavior, and editorial register. The kill list — a curated set of rejected drafts — teaches the system what the voice refuses to do, not just what it does.

---

Quick Start

# Clone the repo
git clone https://github.com/mitwilli-create/voice-os.git
cd voice-os

# Install dependencies
pip install -r requirements.txt

# Set your API key
export ANTHROPIC_API_KEY=your_key_here

# Run a scoring pass against the sample corpus
python score.py --corpus data/sample_corpus.txt --draft data/sample_draft.txt

# Run the full dual-persona pipeline with QA gate
python pipeline.py \
  --corpus data/sample_corpus.txt \
  --kill-list data/kill_list.txt \
  --draft data/sample_draft.txt \
  --output output/scored_draft.json

Output includes axis scores, persona deltas, QA gate decision (pass / cycle), and a revision trace.

---

Architecture

Layer	Function
Corpus ingestion	Chunks and embeds voice corpus; builds axis score baseline
Six-axis scorer	Evaluates drafts against baseline across six stylistic dimensions
Dual-persona router	Generative persona drafts; adversarial persona stress-tests fidelity
QA gate	Blocks output below threshold; returns structured revision signal
Kill list enforcement	Flags patterns the voice explicitly rejects

---

What this demonstrates

• Production RAG design — not a demo, a system that ran at VP scale inside Google
• Evaluation rigor — quantified fidelity scoring, not vibes
• Agentic architecture — multi-step pipeline with conditional routing and gate logic
• Domain depth — a decade in newsrooms and eight years at Google built the editorial judgment that makes the scoring axes meaningful

---

Status

Pipeline architecture and scoring logic are documented here. Core corpus and VP-identity data are not included — that's proprietary. Sample data is synthetic but structurally representative.

CI/evaluation harness: in progress.

---

Built with

• Claude (Anthropic) — generation and adversarial persona
• Python
• Custom embedding + scoring layer

---

Mitchell Williams · LinkedIn · GitHub · thestorytellermitch.com