DraftTwin

Multi-brand AI customer support drafting tool for D2C brands. Each brand signs up, onboards a brain file, and gets a private dashboard for drafting + reviewing replies.

Stack

Python + Flask, Flask-Login, SQLite, vanilla HTML/JS, Groq SDK (llama-3.3-70b-versatile). Deployable on Railway.

Local setup

cd drafttwin
python -m venv .venv
.venv\Scripts\activate   # Windows
# source .venv/bin/activate  # macOS/Linux
pip install -r requirements.txt
cp .env.example .env     # then edit .env and paste GROQ_API_KEY
python app.py

Open http://localhost:5000.

Flow

1. /signup — email + password. Creates the brand account.
2. /onboarding — fill the brand form. Generates brain.md v1.0, saves to SQLite + brains/<id>-<slug>.md.
3. /dashboard — paste a customer DM → structured { classification, reply, reasoning } response. Last 20 drafts appear below with [Flag this reply] buttons.
4. /brain — edit brain.md section-by-section. Saving bumps the minor version (v1.0 → v1.1 → v1.2). Current version + last-updated shown on dashboard.
5. /flagged — review all flagged drafts with customer message, twin's original reply, and your correction. Use this as your gap list when editing the brain.

Auth & data isolation

• One account per brand (brands.email is UNIQUE).
• Every /dashboard, /brain, /flagged, and /api/* route is @login_required.
• All DB queries are scoped to current_user.id — brands cannot see each other's drafts, flags, or brain files.
• Passwords hashed with werkzeug.security.generate_password_hash (PBKDF2).

How the LLM call works

• System prompt = full brand brain.md + strict operating rules.
• Model: llama-3.3-70b-versatile on Groq (fast inference, OpenAI-compatible function calling).
• Structured output is enforced by a forced draft_reply tool call. Classification is constrained to AUTO | DRAFT+APPROVE | ESCALATE.

Brain versioning

Stored as a plain major.minor string in brands.version. The editor always bumps the minor on save. The parser in brain_editor.py splits the brain into 5 fixed sections + internal notes; the assembler regenerates the header/footer with the new version, so the on-disk file stays well-formed regardless of what the founder edits.

Deploy to Railway

Push this directory to a GitHub repo. In Railway, "New Project" → "Deploy from GitHub repo" → pick the repo. railway.toml + Procfile tell Railway to run gunicorn app:app --bind 0.0.0.0:$PORT.

Required env vars (set in Railway → Variables):

• GROQ_API_KEY — from https://console.groq.com/keys
• FLASK_SECRET_KEY — any long random string
• DRAFTTWIN_DB — optional. Defaults to drafttwin.db in the working directory. For persistence across redeploys, attach a Railway Volume at /data and set DRAFTTWIN_DB=/data/drafttwin.db.

Without a volume, the SQLite file is wiped on every redeploy — fine for testing, not for production users.

Files

File	Purpose
app.py	Flask routes (index, onboarding, dashboard, brain, flagged, API)
auth.py	Flask-Login wiring, User model, login/signup/logout blueprint
brain_generator.py	Onboarding form → initial brain.md
brain_editor.py	Parse/assemble brain.md sections; version bump
llm.py	Groq SDK call, forced tool use
db.py	SQLite schema, migrations, query helpers
templates/	Base, login, signup, onboarding, dashboard, brain, flagged
static/style.css	Styling
brains/	Generated brain.md files per brand
railway.toml, Procfile	Railway deploy

Next phases

• Brain editing: diff view between versions, rollback, import of an external brain.md
• Slack/Telegram notifications for DRAFT+APPROVE / ESCALATE
• "Retrain from flags" — auto-suggest brain edits from flagged examples